From 3fe0f4fcef0a72845a54da775f908e9a85f583d1 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 06:31:03 +0000 Subject: [PATCH 01/20] =?UTF-8?q?=E6=96=B0=E5=BB=BA=20AmoebaNet-D=5FID2073?= =?UTF-8?q?=5Ffor=5FTensorFlow?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.keep | 0 1 file changed, 0 insertions(+), 0 deletions(-) create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.keep diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.keep b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.keep new file mode 100644 index 000000000..e69de29bb -- Gitee From 89f32d76c38e1ebd763202d27cb363ac52c68a75 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 06:37:33 +0000 Subject: [PATCH 02/20] =?UTF-8?q?=E4=B8=8A=E4=BC=A0=E6=BA=90=E7=A0=81?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../README.md | 109 +++ .../amoeba_net.py | 525 ++++++++++++ .../amoeba_net_k8s.yaml | 53 ++ .../amoeba_net_model.py | 680 +++++++++++++++ .../AmoebaNet-D_ID2073_for_TensorFlow/env.sh | 23 + .../fusion_switch.cfg | 37 + .../inception_preprocessing.py | 419 ++++++++++ .../model_builder.py | 451 ++++++++++ .../model_specs.py | 115 +++ .../network_utils.py | 791 ++++++++++++++++++ .../network_utils_test.py | 88 ++ .../online_inference_testcase.sh | 39 + .../cv/AmoebaNet-D_ID2073_for_TensorFlow/run | 2 + .../tf_hub.py | 361 ++++++++ .../train_testcase.sh | 38 + .../AmoebaNet-D_ID2073_for_TensorFlow/xx.sh | 21 + 16 files changed, 3752 insertions(+) create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net_k8s.yaml create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net_model.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/env.sh create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/fusion_switch.cfg create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/inception_preprocessing.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/model_builder.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/model_specs.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/network_utils.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/network_utils_test.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/online_inference_testcase.sh create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/run create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/tf_hub.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/xx.sh diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md new file mode 100644 index 000000000..c0089b14a --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md @@ -0,0 +1,109 @@ +# AmoebaNet-D on TPU + +This code was implemented based on results in the AmoebaNet paper, which should +be cited as: Real, E., Aggarwal, A., Huang, Y. and Le, Q.V., 2018. Regularized +Evolution for Image Classifier Architecture Search. arXiv preprint +arXiv:1802.01548. + + +## Acknowledgements + +The starting point for this code was branched from the implementation for NASNet +in https://github.com/tensorflow/models/tree/master/research/slim/nets/nasnet +and from image processing code in +https://github.com/tensorflow/tpu/blob/master/models/experimental/inception/inception_preprocessing.py. + + +## Prerequisites + +### Setup a Google Cloud project + +Follow the instructions at the [Quickstart Guide](https://cloud.google.com/tpu/docs/quickstart) +to get a GCE VM with access to Cloud TPU. + +To run this model, you will need: + +* A GCE VM instance with an associated Cloud TPU resource +* A GCS bucket to store your training checkpoints +* (Optional): The ImageNet training and validation data preprocessed into + TFRecord format, and stored in GCS. + +### Installing extra packages + +The AmoebaNet trainer uses a few extra packages. We can install them now: + +``` +pip install -U pillow +pip install -U --no-deps tensorflow-serving-api +``` + +### Formatting the data + +The data is expected to be formatted in TFRecord format, as generated by [this +script](https://github.com/tensorflow/tpu/blob/master/tools/datasets/imagenet_to_gcs.py). + +If you do not have ImageNet dataset prepared, you can use a randomly generated +fake dataset to test the model. It is located at +`gs://cloud-tpu-test-datasets/fake_imagenet`. + + +## Training the model + +Train the model by executing the following command (substituting the appropriate +values): + +``` +python amoeba_net.py \ + --tpu=$TPU_NAME \ + --data_dir=$DATA_DIR \ + --model_dir=$MODEL_DIR +``` + +If you are not running this script on a GCE VM in the same project and zone as +your Cloud TPU, you will need to add the `--project` and `--zone` flags +specifying the corresponding values for the Cloud TPU you'd like to use. + +This will train an AmoebaNet-D model on ImageNet with 256 batch size on a +single Cloud TPU. With the default flags on everything, the model should train +to above 80% accuracy in under 48 hours (including evaluation time every +few epochs). + +You can launch TensorBoard (e.g. `tensorboard -logdir=$MODEL_DIR`) to view loss +curves and other metadata regarding your training run. (Note: if you launch +on your VM, be sure to configure ssh port forwarding or the GCE firewall rules +appropriately.) + +You can also train the AmoebaNet-D model to 93% top-5 accuracy in under 7.5 +hours using the following command: + +``` +python amoeba_net.py \ + --tpu=$TPU_NAME \ + --data_dir=$DATA_DIR \ + --model_dir=$MODEL_DIR \ + --num_cells=6 \ + --image_size=224 \ + --num_epochs=35 \ + --train_batch_size=1024 \ + --eval_batch_size=1024 \ + --lr=2.56 \ + --lr_decay_value=0.88 \ + --lr_warmup_epochs=0.35 \ + --mode=train \ + --iterations_per_loop=1251 +``` + +## Understanding the code + +For more detailed information, read the documentation within each file. + +## Additional notes + +### About the model and training regime + +The model is the result of evolutionary neural architecture search presented +in [Regularized Evolution for Image Classifier +Search](https://arxiv.org/abs/1802.01548). + +TODO: give some more details + diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net.py new file mode 100644 index 000000000..fe0db18e2 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net.py @@ -0,0 +1,525 @@ + + + +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +# Copyright 2022 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# pylint: disable=line-too-long +r"""TensorFlow AmoebaNet Example. + +GCP Run Example +python amoeba_net.py --data_dir=gs://cloud-tpu-datasets/imagenet-data --model_dir=gs://cloud-tpu-ckpts/models/ameoba_net_x/ \ +--drop_connect_keep_prob=1.0 --cell_name=evol_net_x --num_cells=12 --reduction_size=256 --image_size=299 --num_epochs=48 \ +--train_batch_size=256 --num_epochs_per_eval=4.0 --lr_decay_value=0.89 --lr_num_epochs_per_decay=1 --alsologtostderr \ +--tpu=huangyp-tpu-0 +""" +# pylint: enable=line-too-long + + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import faulthandler +faulthandler.enable() + + +from npu_bridge.npu_init import * +from npu_bridge.estimator import npu_ops +from npu_bridge.estimator.npu.npu_config import NPURunConfig +from npu_bridge.estimator.npu.npu_config import ProfilingConfig +import sys +import itertools +import math +from absl import app +from absl import flags +import absl.logging as _logging # pylint: disable=unused-import +import tensorflow.compat.v1 as tf +import amoeba_net_model as model_lib +from common import inference_warmup +from tensorflow.contrib import cluster_resolver as contrib_cluster_resolver +from tensorflow.contrib import tpu as contrib_tpu +from tensorflow.contrib.training.python.training import evaluation + + +import os +import argparse + + + + + +# Cloud TPU Cluster Resolvers +flags.DEFINE_string( + 'tpu', default=None, + help='The Cloud TPU to use for training. This should be either the name ' + 'used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 url.') + +flags.DEFINE_string( + 'gcp_project', default=None, + help='Project name for the Cloud TPU-enabled project. If not specified, we ' + 'will attempt to automatically detect the GCE project from metadata.') + +flags.DEFINE_string( + 'tpu_zone', default=None, + help='GCE zone where the Cloud TPU is located in. If not specified, we ' + 'will attempt to automatically detect the GCE project from metadata.') + +# General Parameters +flags.DEFINE_integer( + 'num_shards', 1, + 'Number of shards (TPU cores).') + +flags.DEFINE_integer( + 'distributed_group_size', 1, + help='Size of the distributed batch norm. group.' + 'Default is normalization over local examples only.' + 'When set to a value greater than 1, it will enable' + 'a distribtued batch norm. To enable a global batch norm.' + 'set distributed_group_size to FLAGS.num_shards') + +flags.DEFINE_bool( + 'use_tpu', False, + 'Use TPUs rather than CPU or GPU.') + +flags.DEFINE_string( + 'data_dir', '/home/test_user03/tf_records/', + 'Directory where input data is stored') + +flags.DEFINE_string( + 'model_dir', '/home/test_user03/hh', + 'Directory where model output is stored') + +flags.DEFINE_string( + 'export_dir', '/home/test_user03/result', + 'The directory where the exported SavedModel will be stored.') + +flags.DEFINE_bool( + 'export_to_tpu', False, + help='Whether to export additional metagraph with "serve, tpu" tags' + ' in addition to "serve" only metagraph.') + +flags.DEFINE_integer( + 'iterations_per_loop', 1251, + 'Number of iterations per TPU training loop.') + +flags.DEFINE_integer( + 'train_batch_size', 64, + 'Global (not per-shard) batch size for training') + +flags.DEFINE_integer( + 'eval_batch_size', 64, + 'Global (not per-shard) batch size for evaluation') + +flags.DEFINE_float( + 'num_epochs', 35., + 'Number of steps use for training.') + +flags.DEFINE_float( + 'num_epochs_per_eval', 1., + 'Number of training epochs to run between evaluations.') + +flags.DEFINE_string( + 'mode', 'train_and_eval', + 'Mode to run: train, eval, train_and_eval, or predict') + +flags.DEFINE_integer( + 'save_checkpoints_steps', 0, + 'Interval (in steps) at which the model data ' + 'should be checkpointed. Set to 0 to disable.') + +flags.DEFINE_bool( + 'enable_hostcall', True, + 'Skip the host_call which is executed every training step. This is' + ' generally used for generating training summaries (train loss,' + ' learning rate, etc...). When --enable_hostcall=True, there could' + ' be a performance drop if host_call function is slow and cannot' + ' keep up with the TPU-side computation.') + +# Model specific parameters +flags.DEFINE_bool('use_aux_head', True, 'Include aux head or not.') +flags.DEFINE_float( + 'aux_scaling', 0.4, 'Scaling factor of aux_head') +flags.DEFINE_float( + 'batch_norm_decay', 0.9, 'Batch norm decay.') +flags.DEFINE_float( + 'batch_norm_epsilon', 1e-5, 'Batch norm epsilon.') +flags.DEFINE_float( + 'dense_dropout_keep_prob', None, 'Dense dropout keep probability.') +flags.DEFINE_float( + 'drop_connect_keep_prob', 1.0, 'Drop connect keep probability.') +flags.DEFINE_string( + 'drop_connect_version', None, 'Drop connect version.') +flags.DEFINE_string( + 'cell_name', 'amoeba_net_d', 'Which network to run.') +flags.DEFINE_integer( + 'num_cells', 6, 'Total number of cells.') +flags.DEFINE_integer( + 'reduction_size', 256, 'Default cell reduction size.') +flags.DEFINE_integer( + 'stem_reduction_size', 32, 'Stem filter size.') +flags.DEFINE_float( + 'weight_decay', 4e-05, 'Weight decay for slim model.') +flags.DEFINE_integer( + 'num_label_classes', 1001, 'The number of classes that images fit into.') + +# Training hyper-parameters +flags.DEFINE_float( + 'lr', 2.56, 'Learning rate.') +flags.DEFINE_string( + 'optimizer', 'sgd', + 'Optimizer (one of sgd, rmsprop, momentum)') +flags.DEFINE_float( + 'moving_average_decay', 0.9999, + 'moving average decay rate') +flags.DEFINE_float( + 'lr_decay_value', 0.88, + 'Exponential decay rate used in learning rate adjustment') +flags.DEFINE_integer( + 'lr_num_epochs_per_decay', 1, + 'Exponential decay epochs used in learning rate adjustment') +flags.DEFINE_string( + 'lr_decay_method', 'exponential', + 'Method of decay: exponential, cosine, constant, stepwise') +flags.DEFINE_float( + 'lr_warmup_epochs', 0.35, + 'Learning rate increased from zero linearly to lr for the first ' + 'lr_warmup_epochs.') +flags.DEFINE_float('gradient_clipping_by_global_norm', 0, + 'gradient_clipping_by_global_norm') + +flags.DEFINE_integer( + 'image_size', 224, 'Size of image, assuming image height and width.') + +# flags.DEFINE_integer( +# 'num_train_images', 10000, 'The number of images in the training set.') +# flags.DEFINE_integer( +# 'num_eval_images', 1000, 'The number of images in the evaluation set.') +flags.DEFINE_integer( + 'num_train_images', 1281167, 'The number of images in the training set.') +flags.DEFINE_integer( + 'num_eval_images', 50000, 'The number of images in the evaluation set.') + +flags.DEFINE_bool( + 'use_bp16', True, 'If True, use bfloat16 for activations') + +flags.DEFINE_integer( + 'eval_timeout', 60*60*24, + 'Maximum seconds between checkpoints before evaluation terminates.') + +# Inference configuration. +flags.DEFINE_bool( + 'add_warmup_requests', False, + 'Whether to add warmup requests into the export saved model dir,' + 'especially for TPU inference.') +flags.DEFINE_string('model_name', 'amoeba_net', + 'Serving model name used for the model server.') +# flags.DEFINE_multi_integer( +# 'inference_batch_sizes', [8], +# 'Known inference batch sizes used to warm up for each core.') +flags.DEFINE_multi_integer( + 'inference_batch_sizes', [1], + 'Known inference batch sizes used to warm up for each core.') +FLAGS = flags.FLAGS + + +def build_run_config(): + """Return RunConfig for TPU estimator.""" + tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver( + FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) + + eval_steps = FLAGS.num_eval_images // FLAGS.eval_batch_size + iterations_per_loop = (eval_steps if FLAGS.mode == 'eval' + else FLAGS.iterations_per_loop) + save_checkpoints_steps = FLAGS.save_checkpoints_steps or iterations_per_loop + run_config = contrib_tpu.RunConfig( + cluster=tpu_cluster_resolver, + model_dir=FLAGS.model_dir, + save_checkpoints_steps=save_checkpoints_steps, + keep_checkpoint_max=None, + tpu_config=contrib_tpu.TPUConfig( + iterations_per_loop=iterations_per_loop, + num_shards=FLAGS.num_shards, + per_host_input_for_training=contrib_tpu.InputPipelineConfig + .PER_HOST_V2)) + return run_config + + +def build_image_serving_input_receiver_fn(shape, + dtype=tf.float32): + """Returns a input_receiver_fn for raw images during serving.""" + + def _preprocess_image(encoded_image): + """Preprocess a single raw image.""" + image = tf.image.decode_image(encoded_image, channels=shape[-1]) + image.set_shape(shape) + return tf.cast(image, dtype) + + def serving_input_receiver_fn(): + image_bytes_list = tf.placeholder( + shape=[None], + dtype=tf.string, + ) + images = tf.map_fn( + _preprocess_image, image_bytes_list, back_prop=False, dtype=dtype) + return tf.estimator.export.TensorServingInputReceiver( + features=images, receiver_tensors=image_bytes_list) + + return serving_input_receiver_fn + + +# TODO(ereal): simplify this. +def override_with_flags(hparams): + """Overrides parameters with flag values.""" + override_flag_names = [ + 'aux_scaling', + 'train_batch_size', + 'batch_norm_decay', + 'batch_norm_epsilon', + 'dense_dropout_keep_prob', + 'drop_connect_keep_prob', + 'drop_connect_version', + 'eval_batch_size', + 'gradient_clipping_by_global_norm', + 'lr', + 'lr_decay_method', + 'lr_decay_value', + 'lr_num_epochs_per_decay', + 'moving_average_decay', + 'image_size', + 'num_cells', + 'reduction_size', + 'stem_reduction_size', + 'num_epochs', + 'num_epochs_per_eval', + 'optimizer', + 'enable_hostcall', + 'use_aux_head', + 'use_bp16', + 'use_tpu', + 'lr_warmup_epochs', + 'weight_decay', + 'num_shards', + 'distributed_group_size', + 'num_train_images', + 'num_eval_images', + 'num_label_classes', + ] + for flag_name in override_flag_names: + flag_value = getattr(FLAGS, flag_name, 'INVALID') + if flag_value == 'INVALID': + tf.logging.fatal('Unknown flag %s.' % str(flag_name)) + if flag_value is not None: + _set_or_add_hparam(hparams, flag_name, flag_value) + + +def build_hparams(): + """Build tf.Hparams for training Amoeba Net.""" + hparams = model_lib.build_hparams(FLAGS.cell_name) + override_with_flags(hparams) + return hparams + + +def _terminate_eval(): + tf.logging.info('Timeout passed with no new checkpoints ... terminating eval') + return True + + +def _get_next_checkpoint(): + return evaluation.checkpoints_iterator( + FLAGS.model_dir, timeout=FLAGS.eval_timeout, timeout_fn=_terminate_eval) + + +def _set_or_add_hparam(hparams, name, value): + if getattr(hparams, name, None) is None: + hparams.add_hparam(name, value) + else: + hparams.set_hparam(name, value) + + +def _load_global_step_from_checkpoint_dir(checkpoint_dir): + try: + checkpoint_reader = tf.train.NewCheckpointReader( + tf.train.latest_checkpoint(checkpoint_dir)) + return checkpoint_reader.get_tensor(tf.GraphKeys.GLOBAL_STEP) + except: # pylint: disable=bare-except + return 0 + + +def main(_): + mode = FLAGS.mode + data_dir = FLAGS.data_dir + model_dir = FLAGS.model_dir + hparams = build_hparams() + + estimator_parmas = {} + + train_steps_per_epoch = int( + math.ceil(hparams.num_train_images / float(hparams.train_batch_size))) + eval_steps = hparams.num_eval_images // hparams.eval_batch_size + eval_batch_size = (None if mode == 'train' else + hparams.eval_batch_size) + + model = model_lib.AmoebaNetEstimatorModel(hparams, model_dir) + + if hparams.use_tpu: + run_config = build_run_config() + # Temporary treatment until flags are released. + image_classifier = contrib_tpu.TPUEstimator( + model_fn=model.model_fn, + use_tpu=True, + config=npu_run_config_init(run_config=run_config), + params=estimator_parmas, + predict_batch_size=eval_batch_size, + train_batch_size=hparams.train_batch_size, + eval_batch_size=eval_batch_size, + export_to_tpu=FLAGS.export_to_tpu) + else: + save_checkpoints_steps = (FLAGS.save_checkpoints_steps or + FLAGS.iterations_per_loop) + + # dump_config = npu_tf_config.estimator_dump_config(action='dump') + # dump_config = DumpConfig(enable_dump_debug=True, dump_path="/home/test_user03/tpu-master/models/official/amoeba_net/output/output",dump_step="0|5|10", dump_debug_mode="all") + + # dump_config = npu_tf_config.estimator_dump_config(action='dump') + # # dump_config = DumpConfig(enable_dump_debug=True, + # # dump_path="/home/test_user03/tpu-master/models/official/amoeba_net/output/output", + # # dump_step="0", dump_debug_mode="all") + npu_config = NPURunConfig( + model_dir=FLAGS.model_dir, + save_checkpoints_steps=save_checkpoints_steps, + save_summary_steps=0, + # dump_config=dump_config, + # fusion_switch_file="/home/test_user03/tpu-master/models/official/amoeba_net/fusion_switch.cfg", + session_config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False), + #precision_mode="allow_mix_precision") + precision_mode="allow_fp32_to_fp16") + #precision_mode="force_fp32") + + ''' + run_config = tf.estimator.RunConfig( + model_dir=FLAGS.model_dir, + save_checkpoints_steps=save_checkpoints_steps, save_summary_steps=0) + ''' + image_classifier = NPUEstimator( + model_fn=model.model_fn, + config=npu_config, + params=estimator_parmas) + + + # Input pipelines are slightly different (with regards to shuffling and + # preprocessing) between training and evaluation. + imagenet_train = model_lib.InputPipeline( + is_training=True, data_dir=data_dir, hparams=hparams) + imagenet_eval = model_lib.InputPipeline( + is_training=False, data_dir=data_dir, hparams=hparams) + + if hparams.moving_average_decay < 1: + eval_hooks = [model_lib.LoadEMAHook(model_dir, + hparams.moving_average_decay)] + else: + eval_hooks = [] + + if mode == 'eval': + for checkpoint in _get_next_checkpoint(): + tf.logging.info('Starting to evaluate.') + try: + eval_results = image_classifier.evaluate( + input_fn=imagenet_eval.input_fn, + steps=eval_steps, + hooks=eval_hooks, + checkpoint_path=checkpoint) + tf.logging.info('Evaluation results: %s' % eval_results) + except tf.errors.NotFoundError: + # skip checkpoint if it gets deleted prior to evaluation + tf.logging.info('Checkpoint %s no longer exists ... skipping') + elif mode == 'train_and_eval': + current_step = _load_global_step_from_checkpoint_dir(model_dir) + total_step = int(hparams.num_epochs * train_steps_per_epoch) + tf.logging.info('Starting training at step=%d.' % current_step) + train_steps_per_eval = int( + hparams.num_epochs_per_eval * train_steps_per_epoch) + # Final Evaluation if training is finished. + if current_step >= hparams.num_epochs * train_steps_per_epoch: + eval_results = image_classifier.evaluate( + input_fn=imagenet_eval.input_fn, steps=eval_steps, hooks=eval_hooks) + tf.logging.info('Evaluation results: %s' % eval_results) + while current_step < hparams.num_epochs * train_steps_per_epoch: + image_classifier.train( + input_fn=imagenet_train.input_fn, steps=train_steps_per_eval) + current_step += train_steps_per_eval + tf.logging.info('Starting evaluation at step=%d.' % current_step) + eval_results = image_classifier.evaluate( + input_fn=imagenet_eval.input_fn, steps=eval_steps, hooks=eval_hooks) + tf.logging.info('Evaluation results: %s' % eval_results) + + elif mode == 'predict': + for checkpoint in _get_next_checkpoint(): + tf.logging.info('Starting prediction ...') + time_hook = model_lib.SessionTimingHook() + eval_hooks.append(time_hook) + result_iter = image_classifier.predict( + input_fn=imagenet_eval.input_fn, + hooks=eval_hooks, + checkpoint_path=checkpoint, + yield_single_examples=False) + results = list(itertools.islice(result_iter, eval_steps)) + tf.logging.info('Inference speed = {} images per second.'.format( + time_hook.compute_speed(len(results) * eval_batch_size))) + elif mode == 'train': + current_step = _load_global_step_from_checkpoint_dir(model_dir) + total_step = int(hparams.num_epochs * train_steps_per_epoch) + if current_step < total_step: + tf.logging.info('Starting training ...') + image_classifier.train( + input_fn=imagenet_train.input_fn, + steps=total_step-current_step) + else: + tf.logging.info('Mode not found.') + + if FLAGS.export_dir is not None: + tf.logging.info('Starting exporting saved model ...') + serving_shape = [hparams.image_size, hparams.image_size, 3] + export_path = image_classifier.export_saved_model( + export_dir_base=FLAGS.export_dir, + serving_input_receiver_fn=build_image_serving_input_receiver_fn( + serving_shape), + as_text=True) + if FLAGS.add_warmup_requests: + inference_warmup.write_warmup_requests( + export_path, + FLAGS.model_name, + hparams.image_size, + batch_sizes=FLAGS.inference_batch_sizes) + + +if __name__ == '__main__': + tf.logging.set_verbosity(tf.logging.INFO) + app.run(main) + + diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net_k8s.yaml b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net_k8s.yaml new file mode 100644 index 000000000..b4a31e00d --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net_k8s.yaml @@ -0,0 +1,53 @@ +# Train AmoebaNet-D with fake ImageNet dataset using Cloud TPU and Google +# Kubernetes Engine. +# +# [Training Data] +# In this example, we use the randomly generated fake ImageNet dataset at +# gs://cloud-tpu-test-datasets/fake_imagenet as the training data. +# +# [Instructions] +# 1. Follow the instructions on https://cloud.google.com/tpu/docs/kubernetes-engine-setup +# to create a Kubernetes Engine cluster. +# Note: Use a base machine type with more memory than the default n1-standard-1. +# 2. Change the environment variable MODEL_BUCKET in the Job spec to the +# Google Cloud Storage location where you want to store the output model. +# 3. Run `kubectl create -f amoeba_net_k8s.yaml`. + +apiVersion: batch/v1 +kind: Job +metadata: + name: amoeba-net-tpu +spec: + template: + metadata: + annotations: + # The Cloud TPUs that will be created for this Job must support + # TensorFlow 1.11. This version MUST match the TensorFlow version that + # your model is built on. + tf-version.cloud-tpus.google.com: "1.11" + spec: + restartPolicy: Never + containers: + - name: amoeba-net-tpu + # The official TensorFlow 1.11 TPU model image built from https://github.com/tensorflow/tpu/blob/r1.11/tools/docker/Dockerfile. + image: gcr.io/tensorflow/tpu-models:r1.11 + command: + - python + - /tensorflow_tpu_models/models/official/amoeba_net/amoeba_net.py + - --data_dir=$(DATA_BUCKET) + - --model_dir=$(MODEL_BUCKET) + env: + # The Google Cloud Storage location where the fake ImageNet dataset is + # stored. + - name: DATA_BUCKET + value: "gs://cloud-tpu-test-datasets/fake_imagenet" + # [REQUIRED] Must specify the Google Cloud Storage location where your + # output model will be stored. + - name: MODEL_BUCKET + value: "gs:///amoeba_net" + resources: + limits: + # Request a single v2-8 Cloud TPU device to train the model. + # A single v2-8 Cloud TPU device consists of 4 chips, each of which + # has 2 cores, so there are 8 cores in total. + cloud-tpus.google.com/v2: 8 diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net_model.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net_model.py new file mode 100644 index 000000000..bfea61e30 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/amoeba_net_model.py @@ -0,0 +1,680 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +# Copyright 2022 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +"""AmoebaNet ImageNet model functions.""" + + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + + +import faulthandler +faulthandler.enable() + + +from npu_bridge.npu_init import * +from npu_bridge.estimator import npu_ops + +import copy +import os +import time + +import numpy as np +import six +import tensorflow.compat.v1 as tf + +import inception_preprocessing +import model_builder +import model_specs +from tensorflow.contrib import framework as contrib_framework +from tensorflow.contrib import tpu as contrib_tpu +from tensorflow.contrib import training as contrib_training + + +# Random cropping constants +_RESIZE_SIDE_MIN = 300 +_RESIZE_SIDE_MAX = 600 + +# Constants dictating the learning rate schedule. +RMSPROP_DECAY = 0.9 # Decay term for RMSProp. +RMSPROP_MOMENTUM = 0.9 # Momentum in RMSProp. +RMSPROP_EPSILON = 1.0 # Epsilon term for RMSProp. + + +def imagenet_hparams(): + """Returns default ImageNet training params. + + These defaults are for full training. For search training, some should be + modified to increase the speed of the search. + """ + return contrib_training.HParams( + ########################################################################## + # Input pipeline params. ################################################# + ########################################################################## + image_size=299, + num_train_images=1281167, + num_eval_images=50000, + num_label_classes=1001, + ########################################################################## + # Architectural params. ################################################## + ########################################################################## + + # The total number of regular cells (summed across all stacks). Reduction + # cells are not included. + num_cells=18, + reduction_size=256, + stem_reduction_size=32, + + # How many reduction cells to use between the stacks of regular cells. + num_reduction_layers=2, + + # Stem. + stem_type='imagenet', # 'imagenet' or others + num_stem_cells=2, # 2 if stem_type == 'imagenet' else 0 + + # Implementation details. + data_format='NCHW', # 'NHWC' or 'NCHW'. + + ########################################################################## + # Training params. ####################################################### + ########################################################################## + + # Summed across all TPU cores training a model. + train_batch_size=32, + num_epochs=100., + + # Auxiliary head. + use_aux_head=True, + aux_scaling=0.4, + + # Regularization. + l1_decay_rate=0.0, + label_smoothing=0.1, + drop_connect_keep_prob=0.7, + # `drop_connect_version` determines how the drop_connect probabilites are + # set/increased over time: + # -v1: increase dropout probability over training, + # -v2: increase dropout probability as you increase the number of cells, + # so the top cell has the highest dropout and the lowest cell has the + # lowest dropout, + # -v3: Do both v1 and v2. + drop_connect_version='v1', + drop_path_burn_in_steps=0, + # `drop_connect_condition` determines under what conditions drop_connect + # is used: + # -identity: Dropout all paths except identity connections, + # -all: Dropout all paths, + # -separable: Dropout only paths containing a separable conv operation. + dense_dropout_keep_prob=0.5, + batch_norm_epsilon=0.001, + batch_norm_decay=0.9997, + shuffle_buffer=20000, + + # Any value <= 0 means it is unused + gradient_clipping_by_global_norm=10.0, + + # Learning rate schedule. + lr=0.015, + lr_decay_method='exponential', + lr_decay_value=0.97, + lr_num_epochs_per_decay=2.4, + lr_warmup_epochs=3.0, + weight_decay=4e-05, + + # Optimizer. + optimizer='sgd', # 'sgd', 'mom', 'adam' or 'rmsprop' + rmsprop_decay=0.9, + rmsprop_momentum_rate=0.9, + rmsprop_epsilon=1.0, + momentum_rate=0.9, + use_nesterov=1, + + ########################################################################## + # Eval and reporting params. ############################################# + ########################################################################## + + # This number should be a multiple of the number of TPU shards + # used for eval (e.g., 2 for a 1x1 or 8 for a 2x2). + eval_batch_size=40, + + # How many different crops are fed into one model. Also affects training. + num_input_images=1, + moving_average_decay=0.9999, + write_summaries=0, + + ########################################################################## + # Other params. ########################################################## + ########################################################################## + num_shards=None, + distributed_group_size=1, + use_tpu=False) + + +def build_hparams(cell_name='amoeba_net_d'): + """Build tf.Hparams for training Amoeba Net. + + Args: + cell_name: Which of the cells in model_specs.py to use to build the + amoebanet neural network; the cell names defined in that + module correspond to architectures discovered by an + evolutionary search described in + https://arxiv.org/abs/1802.01548. + + Returns: + A set of tf.HParams suitable for Amoeba Net training. + """ + hparams = imagenet_hparams() + operations, hiddenstate_indices, used_hiddenstates = ( + model_specs.get_normal_cell(cell_name)) + hparams.add_hparam('normal_cell_operations', operations) + hparams.add_hparam('normal_cell_hiddenstate_indices', + hiddenstate_indices) + hparams.add_hparam('normal_cell_used_hiddenstates', + used_hiddenstates) + operations, hiddenstate_indices, used_hiddenstates = ( + model_specs.get_reduction_cell(cell_name)) + hparams.add_hparam('reduction_cell_operations', + operations) + hparams.add_hparam('reduction_cell_hiddenstate_indices', + hiddenstate_indices) + hparams.add_hparam('reduction_cell_used_hiddenstates', + used_hiddenstates) + hparams.set_hparam('data_format', 'NHWC') + return hparams + + +def formatted_hparams(hparams): + """Formatts the hparams into a readable string. + + Also looks for attributes that have not correctly been added to the hparams + and prints the keys as "bad keys". These bad keys may be left out of iterators + and cirumvent type checking. + + Args: + hparams: an HParams instance. + + Returns: + A string. + """ + # Look for bad keys (see docstring). + good_keys = set(hparams.values().keys()) + bad_keys = [] + for key in hparams.__dict__: + if key not in good_keys and not key.startswith('_'): + bad_keys.append(key) + bad_keys.sort() + + # Format hparams. + readable_items = [ + '%s: %s' % (k, v) for k, v in sorted(six.iteritems(hparams.values())) + ] + readable_items.append('Bad keys: %s' % ','.join(bad_keys)) + readable_string = ('\n'.join(readable_items)) + return readable_string + + +class AmoebaNetEstimatorModel(object): + """Definition of AmoebaNet.""" + + def __init__(self, hparams, model_dir): + self.hparams = hparams + self.model_dir = model_dir + + def _calc_num_trainable_params(self): + self.num_trainable_params = np.sum([ + np.prod(var.get_shape().as_list()) for var in tf.trainable_variables() + ]) + tf.logging.info( + 'number of trainable params: {}'.format(self.num_trainable_params)) + + def _build_learning_rate_schedule(self, global_step): + """Build learning rate.""" + steps_per_epoch = ( + self.hparams.num_train_images // self.hparams.train_batch_size) + lr_warmup_epochs = 0 + if self.hparams.lr_decay_method == 'exponential': + lr_warmup_epochs = self.hparams.lr_warmup_epochs + learning_rate = model_builder.build_learning_rate( + self.hparams.lr, + self.hparams.lr_decay_method, + global_step, + total_steps=steps_per_epoch * self.hparams.num_epochs, + decay_steps=steps_per_epoch * self.hparams.lr_num_epochs_per_decay, + decay_factor=self.hparams.lr_decay_value, + add_summary=False, + warmup_steps=int(lr_warmup_epochs * steps_per_epoch)) + + learning_rate = tf.maximum( + learning_rate, 0.0001 * self.hparams.lr, name='learning_rate') + return learning_rate + + def _build_network(self, features, labels, mode): + """Build a network that returns loss and logits from features and labels.""" + is_training = (mode == tf.estimator.ModeKeys.TRAIN) + is_predict = (mode == tf.estimator.ModeKeys.PREDICT) + steps_per_epoch = float( + self.hparams.num_train_images) / self.hparams.train_batch_size + num_total_steps = int(steps_per_epoch * self.hparams.num_epochs) + self.hparams.set_hparam('drop_path_burn_in_steps', num_total_steps) + + hparams = copy.deepcopy(self.hparams) + if not is_training: + hparams.set_hparam('use_aux_head', False) + hparams.set_hparam('weight_decay', 0) + hparams.set_hparam('use_bp16', False) + + tf.logging.info( + 'Amoeba net received hparams for {}:\n{}'.format( + 'training' if is_training else 'eval', + formatted_hparams(hparams))) + + logits, end_points = model_builder.build_network( + features, hparams.num_label_classes, is_training, hparams) + + if not is_predict: + labels = tf.one_hot(labels, hparams.num_label_classes) + loss = model_builder.build_softmax_loss( + logits, + end_points, + labels, + label_smoothing=hparams.label_smoothing, + add_summary=False) + + # Calculate and print the number of trainable parameters in the model + if is_training: + flops = model_builder.compute_flops_per_example(hparams.train_batch_size) + else: + flops = model_builder.compute_flops_per_example(hparams.eval_batch_size) + tf.logging.info('number of flops: {}'.format(flops)) + self._calc_num_trainable_params() + + if is_predict: + return None, logits + + return loss, logits + + def _build_optimizer(self, learning_rate): + """Build optimizer.""" + if self.hparams.optimizer == 'sgd': + tf.logging.info('Using SGD optimizer') + optimizer = tf.train.GradientDescentOptimizer( + learning_rate=learning_rate) + opt_tmp = optimizer + loss_scale_manager = ExponentialUpdateLossScaleManager(init_loss_scale=2 ** 32, incr_every_n_steps=1000, + decr_every_n_nan_or_inf=2, decr_ratio=0.5) + optimizer = NPULossScaleOptimizer(opt_tmp, loss_scale_manager) + + # elif self.hparams.optimizer == 'adam': + # tf.logging.info('Using adam optimizer') + # optimizer = tf.train.MomentumOptimizer( + # learning_rate=learning_rate, + # momentum=self.hparams.momentum_rate) + # opt_tmp = optimizer + # loss_scale_manager = ExponentialUpdateLossScaleManager(init_loss_scale=2 ** 32, incr_every_n_steps=1000, + # decr_every_n_nan_or_inf=2, decr_ratio=0.5) + # optimizer = NPULossScaleOptimizer(opt_tmp, loss_scale_manager) + + elif self.hparams.optimizer == 'momentum': + tf.logging.info('Using Momoptimizerentum ') + optimizer = tf.train.MomentumOptimizer( + learning_rate=learning_rate, + momentum=self.hparams.momentum_rate) + opt_tmp = optimizer + loss_scale_manager = ExponentialUpdateLossScaleManager(init_loss_scale=2 ** 32, incr_every_n_steps=1000, + decr_every_n_nan_or_inf=2, decr_ratio=0.5) + optimizer = NPULossScaleOptimizer(opt_tmp, loss_scale_manager) + + elif self.hparams.optimizer == 'rmsprop': + tf.logging.info('Using RMSProp optimizer') + optimizer = tf.train.RMSPropOptimizer( + learning_rate, + RMSPROP_DECAY, + momentum=RMSPROP_MOMENTUM, + epsilon=RMSPROP_EPSILON) + opt_tmp = optimizer + loss_scale_manager = ExponentialUpdateLossScaleManager(init_loss_scale=2 ** 32, incr_every_n_steps=1000, + decr_every_n_nan_or_inf=2, decr_ratio=0.5) + optimizer = NPULossScaleOptimizer(opt_tmp, loss_scale_manager) + + else: + tf.logging.fatal('Unknown optimizer:', self.hparams.optimizer) + + if self.hparams.use_tpu: + optimizer = contrib_tpu.CrossShardOptimizer(optimizer) + return optimizer + + def _build_train_op(self, optimizer, loss, global_step): + """Build train_op from optimizer and loss.""" + grads_and_vars = optimizer.compute_gradients(loss) + if self.hparams.gradient_clipping_by_global_norm > 0.0: + g, v = zip(*grads_and_vars) + g, _ = tf.clip_by_global_norm( + g, self.hparams.gradient_clipping_by_global_norm) + grads_and_vars = zip(g, v) + + return optimizer.apply_gradients(grads_and_vars, global_step=global_step) + + def model_fn(self, features, labels, mode, params): + """Build the model based on features, labels, and mode. + + Args: + features: The features dictionary containing the data Tensor + and the number of examples. + labels: The labels Tensor resulting from calling the model. + mode: A string indicating the training mode. + params: A dictionary of hyperparameters. + + Returns: + A tf.estimator.EstimatorSpec. + """ + del params + is_training = (mode == tf.estimator.ModeKeys.TRAIN) + eval_active = (mode == tf.estimator.ModeKeys.EVAL) + is_predict = (mode == tf.estimator.ModeKeys.PREDICT) + + if is_training: + features = tf.transpose(features, [3, 0, 1, 2]) # HWCN to NHWC + loss, logits = self._build_network(features, labels, mode) + + if is_predict: + predictions = {'logits': logits} + if self.hparams.use_tpu: + return contrib_tpu.TPUEstimatorSpec(mode=mode, predictions=predictions) + else: + return tf.estimator.EstimatorSpec(mode=mode, + predictions=predictions) + host_call = None + train_op = None + + if is_training: + global_step = tf.train.get_or_create_global_step() + gs_t = tf.reshape(tf.cast(global_step, tf.int32), [1]) + + # Setup learning rate schedule + learning_rate = self._build_learning_rate_schedule(global_step) + + # Setup optimizer. + optimizer = self._build_optimizer(learning_rate) + ''' + opt_tmp = optimizer + loss_scale_manager = ExponentialUpdateLossScaleManager(init_loss_scale=2 ** 32, incr_every_n_steps=1000, + decr_every_n_nan_or_inf=2, decr_ratio=0.5) + optimizer = NPULossScaleOptimizer(opt_tmp, loss_scale_manager) + ''' + update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) + with tf.control_dependencies(update_ops): + train_op = self._build_train_op(optimizer, loss, + global_step=global_step) + if self.hparams.moving_average_decay > 0: + ema = tf.train.ExponentialMovingAverage( + decay=self.hparams.moving_average_decay, num_updates=global_step) + variables_to_average = (tf.trainable_variables() + + tf.moving_average_variables()) + with tf.control_dependencies([train_op]): + with tf.name_scope('moving_average'): + train_op = ema.apply(variables_to_average) + + lr_t = tf.reshape(learning_rate, [1]) + host_call = None + if self.hparams.enable_hostcall: + def host_call_fn(gs, lr): + # Outfeed supports int32 but global_step is expected to be int64. + gs = tf.cast(tf.reduce_mean(gs), tf.int64) + with tf.summary.create_file_writer(self.model_dir).as_default(): + with tf.summary.always_record_summaries(): + tf.summary.scalar('learning_rate', tf.reduce_mean(lr), step=gs) + return tf.summary.all_summary_ops() + + host_call = (host_call_fn, [gs_t, lr_t]) + + eval_metrics = None + eval_metric_ops = None + if eval_active: + def metric_fn(labels, logits): + """Evaluation metric fn. Performed on CPU, do not reference TPU ops.""" + # Outfeed supports int32 but global_step is expected to be int64. + predictions = tf.argmax(logits, axis=1) + categorical_labels = labels + top_1_accuracy = tf.metrics.accuracy(categorical_labels, predictions) + in_top_5 = tf.cast(tf.nn.in_top_k(logits, categorical_labels, 5), + tf.float32) + top_5_accuracy = tf.metrics.mean(in_top_5) + + return { + 'top_1_accuracy': top_1_accuracy, + 'top_5_accuracy': top_5_accuracy, + } + + eval_metrics = (metric_fn, [labels, logits]) + eval_metric_ops = metric_fn(labels, logits) + + if self.hparams.use_tpu: + return contrib_tpu.tpu.TPUEstimatorSpec( + mode=mode, + loss=loss, + train_op=train_op, + host_call=host_call, + eval_metrics=eval_metrics) + return tf.estimator.EstimatorSpec( + mode=mode, loss=loss, train_op=train_op, + eval_metric_ops=eval_metric_ops) + + +class InputPipeline(object): + """Generates ImageNet input_fn for training or evaluation. + + The training data is assumed to be in TFRecord format with keys as specified + in the dataset_parser below, sharded across 1024 files, named sequentially: + train-00000-of-01024 + train-00001-of-01024 + ... + train-01023-of-01024 + + The validation data is in the same format but sharded in 128 files. + + The format of the data required is created by the script at: + https://github.com/tensorflow/tpu/blob/master/tools/datasets/imagenet_to_gcs.py + + Args: + is_training: `bool` for whether the input is for training + """ + + def __init__(self, is_training, data_dir, hparams, eval_from_hub=False): + self.is_training = is_training + self.data_dir = data_dir + self.hparams = hparams + self.num_classes = 1001 + self.eval_from_hub = eval_from_hub + + def _dataset_parser(self, serialized_proto): + """Parse an Imagenet record from value.""" + keys_to_features = { + 'image/encoded': + tf.FixedLenFeature((), tf.string, default_value=''), + 'image/format': + tf.FixedLenFeature((), tf.string, default_value='jpeg'), + 'image/class/label': + tf.FixedLenFeature([], dtype=tf.int64, default_value=-1), + 'image/class/text': + tf.FixedLenFeature([], dtype=tf.string, default_value=''), + 'image/object/bbox/xmin': + tf.VarLenFeature(dtype=tf.float32), + 'image/object/bbox/ymin': + tf.VarLenFeature(dtype=tf.float32), + 'image/object/bbox/xmax': + tf.VarLenFeature(dtype=tf.float32), + 'image/object/bbox/ymax': + tf.VarLenFeature(dtype=tf.float32), + 'image/object/class/label': + tf.VarLenFeature(dtype=tf.int64), + } + + features = tf.parse_single_example(serialized_proto, keys_to_features) + + bbox = None + + image = features['image/encoded'] + image = tf.image.decode_jpeg(image, channels=3) + image = tf.image.convert_image_dtype(image, dtype=tf.float32) + + image = inception_preprocessing.preprocess_image( + image=image, + output_height=self.hparams.image_size, + output_width=self.hparams.image_size, + is_training=self.is_training, + # If eval_from_hub, do not scale the images during preprocessing. + scaled_images=not self.eval_from_hub, + bbox=bbox) + + label = tf.cast( + tf.reshape(features['image/class/label'], shape=[]), dtype=tf.int32) + + return image, label + + def input_fn(self, params): + """Input function which provides a single batch for train or eval. + + Args: + params: `dict` of parameters passed from the `TPUEstimator`. + + Returns: + A callable dataset object. + """ + # Retrieves the batch size for the current shard. The # of shards is + # computed according to the input pipeline deployment. See + # tf.contrib.tpu.RunConfig for details. + if 'batch_size' in params: + batch_size = params['batch_size'] + else: + batch_size = (self.hparams.train_batch_size if self.is_training + else self.hparams.eval_batch_size) + file_pattern = os.path.join( + self.data_dir, 'train/train-*' if self.is_training else 'validation/validation-*') + dataset = tf.data.Dataset.list_files(file_pattern, shuffle=self.is_training) + if self.is_training: + dataset = dataset.repeat() + def fetch_dataset(filename): + buffer_size = 8 * 1024 * 1024 # 8 MiB per file + dataset = tf.data.TFRecordDataset(filename, buffer_size=buffer_size) + return dataset + + dataset = dataset.apply( + tf.data.experimental.parallel_interleave( + fetch_dataset, cycle_length=64, sloppy=True)) + dataset = dataset.shuffle(1024) + + # Use the fused map-and-batch operation. + # + # For XLA, we must used fixed shapes. Because we repeat the source training + # dataset indefinitely, we can use `drop_remainder=True` to get fixed-size + # batches without dropping any training examples. + # + # When evaluating, `drop_remainder=True` prevents accidentally evaluating + # the same image twice by dropping the final batch if it is less than a full + # batch size. As long as this validation is done with consistent batch size, + # exactly the same images will be used. + + dataset = dataset.apply( + tf.data.experimental.map_and_batch( + self._dataset_parser, batch_size=batch_size, + num_parallel_batches=8, drop_remainder=True)) + ''' + dataset = dataset.apply( + tf.data.experimental.map_and_batch( + self._dataset_parser, batch_size=batch_size, + num_parallel_batches=1, drop_remainder=True)) + ''' + if self.is_training: + dataset = dataset.map( + lambda images, labels: (tf.transpose(images, [1, 2, 3, 0]), labels), + num_parallel_calls=8) + ''' + if self.is_training: + dataset = dataset.map( + lambda images, labels: (tf.transpose(images, [1, 2, 3, 0]), labels)) + ''' + dataset = dataset.prefetch(32) # Prefetch overlaps in-feed with training + return dataset # Must return the dataset and not tensors for high perf! + + +class LoadEMAHook(tf.train.SessionRunHook): + """Hook to load EMA into their corresponding variables.""" + + def __init__(self, model_dir, moving_average_decay): + super(LoadEMAHook, self).__init__() + self._model_dir = model_dir + self.moving_average_decay = moving_average_decay + + def begin(self): + ema = tf.train.ExponentialMovingAverage(self.moving_average_decay) + variables_to_restore = ema.variables_to_restore() + self._load_ema = contrib_framework.assign_from_checkpoint_fn( + tf.train.latest_checkpoint(self._model_dir), variables_to_restore) + + def after_create_session(self, sess, coord): + tf.logging.info('Reloading EMA...') + self._load_ema(sess) + + +class SessionTimingHook(tf.train.SessionRunHook): + """Hook that computes speed based on session run time.""" + + def __init__(self): + # Lists of walltime. + self._before_runs = [] + self._after_runs = [] + + def before_run(self, run_context): + self._before_runs.append(time.time()) + + def after_run(self, run_context, results): + self._after_runs.append(time.time()) + + def compute_speed(self, num_samples): + """Returns speed, in number of samples per second.""" + num_runs = len(self._before_runs) + if num_runs == 0: + raise ValueError('Session run time never recorded') + if len(self._after_runs) != num_runs: + raise ValueError( + 'Number of before_run events (%d) does not match ' + 'number of after_run events (%d)' % + (len(self._before_runs), len(self._after_runs))) + total_eval_time = sum(self._after_runs[i] - self._before_runs[i] + for i in range(num_runs)) + if num_runs <= 1: + tf.logging.warn( + 'Speed will be inaccurate with only one session run') + else: + # Exclude the first run, which tends to take much longer than other runs. + total_eval_time -= (self._after_runs[0] - self._before_runs[0]) + # We assume num_samples are evenly distributed across runs. + num_samples *= (float(num_runs - 1) / num_runs) + return num_samples / max(total_eval_time, 1e-6) + diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/env.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/env.sh new file mode 100644 index 000000000..7802ada07 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/env.sh @@ -0,0 +1,23 @@ +export RANK_INDEX=0 +export RANK_SIZE=1 +export RANK_ID=0 +export ASCEND_DEVICE_ID=2 +export DEVICE_INDEX=0 +export install_path=/usr/local/Ascend +export JOB_ID=100001 +#nnae +export LD_LIBRARY_PATH=${install_path}/nnae/latest/fwkacllib/lib64:$LD_LIBRARY_PATH +export PATH=${install_path}/nnae/latest/fwkacllib/ccec_compiler/bin:${install_path}/nnae/latest/fwkacllib/bin:$PATH +export PYTHONPATH=${install_path}/nnae/latest/fwkacllib/python/site-packages:$PYTHONPATH +export ASCEND_OPP_PATH=${install_path}/nnae/latest/opp +export ASCEND_AICPU_PATH=${install_path}/nnae/latest +#tfplugin +export PYTHONPATH=${install_path}/tfplugin/latest/tfplugin/python/site-packages:$PYTHONPATH +#Ascend-dmi +export LD_LIBRARY_PATH=/usr/local/dcmi:${install_path}/toolbox/latest/Ascend-DMI/lib64:${LD_LIBRARY_PATH} +export PATH=${install_path}/toolbox/latest/Ascend-DMI/bin:${PATH} +export LD_LIBRARY_PATH=/usr/local/gcc7.3.0/lib64:${LD_LIBRARY_PATH} +export PATH=/usr/local/gcc7.3.0/bin:${PATH} +export LD_LIBRARY_PATH=/usr/local/python3.7.5/lib:$LD_LIBRARY_PATH +export PATH=/usr/local/python3.7.5/bin:$PATH +export LD_LIBRARY_PATH=/usr/include/hdf5/lib:$LD_LIBRARY_PATH diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/fusion_switch.cfg b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/fusion_switch.cfg new file mode 100644 index 000000000..71a3d36cf --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/fusion_switch.cfg @@ -0,0 +1,37 @@ +{ + "Switch":{ + "GraphFusion":{ + "AReduceMeanFusionPass":"off", + "AReduceSumFusionPass":"off", + "ConvToFullyConnectionFusionPass":"off", + "ConcatCToNOptimizeFusionPass":"off", + "Conv2DbpFilterMulFusionPass":"off", + "Conv2DbpInputDilationFusionPass":"off", + "ConvBatchnormFusionPass":"off", + "ConvConcatFusionPass":"off", + "ConvWeightCompressFusionPass":"off", + "FusedBatchNormBertFusionPass":"off", + "SoftmaxFusionPass":"off", + "V100RequantFusionPass":"off", + "V100NotRequantFusionPass":"off", + "SplitConvConcatFusionPass":"off", + "StridedSliceGradFusionPass":"off", + "MatMulBiasAddFusionPass":"off", + "MulAddFusionPass":"off", + "MulAddNL2LossFusionPass":"off", + "MulGradFusionPass":"off", + "MulAddNPass":"off", + "PoolingFusionPass":"off", + "Pow2SquareFusionPass":"off", + "Resnet50DbnDwFusionPass":"off", + "ZConcatv2dFusionPass":"off", + "ZConcatExt2FusionPass":"off", + "TransdataCastFusionPass":"off", + "TfMergeSubFusionPass":"off" + }, + "UBFusion":{ + "TbePool2dQuantFusionPass":"off", + "FusionVirtualOpSetSwitch":"off" + } + } +} diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/inception_preprocessing.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/inception_preprocessing.py new file mode 100644 index 000000000..7cac012db --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/inception_preprocessing.py @@ -0,0 +1,419 @@ + + + +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +# Copyright 2022 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +"""Provides utilities to preprocess images for the Inception networks.""" + + + + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + + +import faulthandler +faulthandler.enable() + + +from npu_bridge.npu_init import * + +from absl import flags +import tensorflow.compat.v1 as tf + +from tensorflow.python.ops import control_flow_ops +from tensorflow.python.ops import random_ops + + +flags.DEFINE_float( + 'cb_distortion_range', 0.1, 'Cb distortion range +/-') + +flags.DEFINE_float( + 'cr_distortion_range', 0.1, 'Cr distortion range +/-') + +flags.DEFINE_boolean( + 'use_fast_color_distort', True, + 'apply fast color/chroma distortion if True, else apply' + 'brightness/saturation/hue/contrast distortion') + +FLAGS = flags.FLAGS + + +def apply_with_random_selector(x, func, num_cases): + """Computes func(x, sel), with sel sampled from [0...num_cases-1]. + + Args: + x: input Tensor. + func: Python function to apply. + num_cases: Python int32, number of cases to sample sel from. + + Returns: + The result of func(x, sel), where func receives the value of the + selector as a python integer, but sel is sampled dynamically. + """ + sel = tf.random_uniform([], maxval=num_cases, dtype=tf.int32) + # Pass the real x only to one of the func calls. + return control_flow_ops.merge([ + func(control_flow_ops.switch(x, tf.equal(sel, case))[1], case) + for case in range(num_cases)])[0] + + +def distort_color(image, color_ordering=0, fast_mode=True, scope=None): + """Distort the color of a Tensor image. + + Each color distortion is non-commutative and thus ordering of the color ops + matters. Ideally we would randomly permute the ordering of the color ops. + Rather then adding that level of complication, we select a distinct ordering + of color ops for each preprocessing thread. + + Args: + image: 3-D Tensor containing single image in [0, 1]. + color_ordering: Python int, a type of distortion (valid values: 0-3). + fast_mode: Avoids slower ops (random_hue and random_contrast) + scope: Optional scope for name_scope. + Returns: + 3-D Tensor color-distorted image on range [0, 1] + Raises: + ValueError: if color_ordering not in [0, 3] + """ + with tf.name_scope(scope, 'distort_color', [image]): + if fast_mode: + if color_ordering == 0: + image = tf.image.random_brightness(image, max_delta=32. / 255.) + image = tf.image.random_saturation(image, lower=0.5, upper=1.5) + else: + image = tf.image.random_saturation(image, lower=0.5, upper=1.5) + image = tf.image.random_brightness(image, max_delta=32. / 255.) + else: + if color_ordering == 0: + image = tf.image.random_brightness(image, max_delta=32. / 255.) + image = tf.image.random_saturation(image, lower=0.5, upper=1.5) + image = tf.image.random_hue(image, max_delta=0.2) + image = tf.image.random_contrast(image, lower=0.5, upper=1.5) + elif color_ordering == 1: + image = tf.image.random_saturation(image, lower=0.5, upper=1.5) + image = tf.image.random_brightness(image, max_delta=32. / 255.) + image = tf.image.random_contrast(image, lower=0.5, upper=1.5) + image = tf.image.random_hue(image, max_delta=0.2) + elif color_ordering == 2: + image = tf.image.random_contrast(image, lower=0.5, upper=1.5) + image = tf.image.random_hue(image, max_delta=0.2) + image = tf.image.random_brightness(image, max_delta=32. / 255.) + image = tf.image.random_saturation(image, lower=0.5, upper=1.5) + elif color_ordering == 3: + image = tf.image.random_hue(image, max_delta=0.2) + image = tf.image.random_saturation(image, lower=0.5, upper=1.5) + image = tf.image.random_contrast(image, lower=0.5, upper=1.5) + image = tf.image.random_brightness(image, max_delta=32. / 255.) + else: + raise ValueError('color_ordering must be in [0, 3]') + + # The random_* ops do not necessarily clamp. + return tf.minimum(tf.maximum(image, 0.0), 1.0) + + +def distort_color_fast(image, scope=None): + """Distort the color of a Tensor image. + + Distort brightness and chroma values of input image + + Args: + image: 3-D Tensor containing single image in [0, 1]. + scope: Optional scope for name_scope. + Returns: + 3-D Tensor color-distorted image on range [0, 1] + """ + with tf.name_scope(scope, 'distort_color', [image]): + br_delta = random_ops.random_uniform([], -32./255., 32./255., seed=None) + cb_factor = random_ops.random_uniform( + [], -FLAGS.cb_distortion_range, FLAGS.cb_distortion_range, seed=None) + cr_factor = random_ops.random_uniform( + [], -FLAGS.cr_distortion_range, FLAGS.cr_distortion_range, seed=None) + + channels = tf.split(axis=2, num_or_size_splits=3, value=image) + red_offset = 1.402 * cr_factor + br_delta + green_offset = -0.344136 * cb_factor - 0.714136 * cr_factor + br_delta + blue_offset = 1.772 * cb_factor + br_delta + channels[0] += red_offset + channels[1] += green_offset + channels[2] += blue_offset + image = tf.concat(axis=2, values=channels) + image = tf.minimum(tf.maximum(image, 0.), 1.) + return image + + +def distorted_bounding_box_crop(image, + bbox, + min_object_covered=0.1, + aspect_ratio_range=(3./4., 4./3.), + area_range=(0.05, 1.0), + max_attempts=100, + scope=None): + """Generates cropped_image using a one of the bboxes randomly distorted. + + See `tf.image.sample_distorted_bounding_box` for more documentation. + + Args: + image: 3-D Tensor of image (it will be converted to floats in [0, 1]). + bbox: 3-D float Tensor of bounding boxes arranged [1, num_boxes, coords] + where each coordinate is [0, 1) and the coordinates are arranged + as [ymin, xmin, ymax, xmax]. If num_boxes is 0 then it would use the whole + image. + min_object_covered: An optional `float`. Defaults to `0.1`. The cropped + area of the image must contain at least this fraction of any bounding box + supplied. + aspect_ratio_range: An optional list of `floats`. The cropped area of the + image must have an aspect ratio = width / height within this range. + area_range: An optional list of `floats`. The cropped area of the image + must contain a fraction of the supplied image within in this range. + max_attempts: An optional `int`. Number of attempts at generating a cropped + region of the image of the specified constraints. After `max_attempts` + failures, return the entire image. + scope: Optional scope for name_scope. + Returns: + A tuple, a 3-D Tensor cropped_image and the distorted bbox + """ + with tf.name_scope(scope, 'distorted_bounding_box_crop', [image, bbox]): + # Each bounding box has shape [1, num_boxes, box coords] and + # the coordinates are ordered [ymin, xmin, ymax, xmax]. + + # A large fraction of image datasets contain a human-annotated bounding + # box delineating the region of the image containing the object of interest. + # We choose to create a new bounding box for the object which is a randomly + # distorted version of the human-annotated bounding box that obeys an + # allowed range of aspect ratios, sizes and overlap with the human-annotated + # bounding box. If no box is supplied, then we assume the bounding box is + # the entire image. + sample_distorted_bounding_box = tf.image.sample_distorted_bounding_box( + tf.shape(image), + bounding_boxes=bbox, + min_object_covered=min_object_covered, + aspect_ratio_range=aspect_ratio_range, + area_range=area_range, + max_attempts=max_attempts, + use_image_if_no_bounding_boxes=True) + bbox_begin, bbox_size, distort_bbox = sample_distorted_bounding_box + + # Crop the image to the specified bounding box. + cropped_image = tf.slice(image, bbox_begin, bbox_size) + return cropped_image, distort_bbox + + +def preprocess_for_train(image, height, width, bbox, + min_object_covered=0.1, + fast_mode=True, + scope=None, + add_image_summaries=True): + """Distort one image for training a network. + + Distorting images provides a useful technique for augmenting the data + set during training in order to make the network invariant to aspects + of the image that do not effect the label. + + Additionally it would create image_summaries to display the different + transformations applied to the image. + + Args: + image: 3-D Tensor of image. If dtype is tf.float32 then the range should be + [0, 1], otherwise it would converted to tf.float32 assuming that the range + is [0, MAX], where MAX is largest positive representable number for + int(8/16/32) data type (see `tf.image.convert_image_dtype` for details). + height: integer + width: integer + bbox: 3-D float Tensor of bounding boxes arranged [1, num_boxes, coords] + where each coordinate is [0, 1) and the coordinates are arranged + as [ymin, xmin, ymax, xmax]. + min_object_covered: An optional `float`. Defaults to `0.1`. The cropped + area of the image must contain at least this fraction of any bounding box + supplied. + fast_mode: Optional boolean, if True avoids slower transformations (i.e. + bi-cubic resizing, random_hue or random_contrast). + scope: Optional scope for name_scope. + add_image_summaries: Enable image summaries. + Returns: + 3-D float Tensor of distorted image used for training with range [-1, 1]. + """ + with tf.name_scope(scope, 'distort_image', [image, height, width, bbox]): + if bbox is None: + bbox = tf.constant([0.0, 0.0, 1.0, 1.0], + dtype=tf.float32, + shape=[1, 1, 4]) + if image.dtype != tf.float32: + image = tf.image.convert_image_dtype(image, dtype=tf.float32) + if add_image_summaries: + # Each bounding box has shape [1, num_boxes, box coords] and + # the coordinates are ordered [ymin, xmin, ymax, xmax]. + image_with_box = tf.image.draw_bounding_boxes(tf.expand_dims(image, 0), + bbox) + tf.summary.image('image_with_bounding_boxes', image_with_box) + + distorted_image, distorted_bbox = distorted_bounding_box_crop( + image, + bbox, + min_object_covered=min_object_covered, + area_range=(min_object_covered, 1.0)) + # Restore the shape since the dynamic slice based upon the bbox_size loses + # the third dimension. + distorted_image.set_shape([None, None, 3]) + if add_image_summaries: + image_with_distorted_box = tf.image.draw_bounding_boxes( + tf.expand_dims(image, 0), distorted_bbox) + tf.summary.image('images_with_distorted_bounding_box', + image_with_distorted_box) + + # This resizing operation may distort the images because the aspect + # ratio is not respected. We select a resize method in a round robin + # fashion based on the thread number. + # Note that ResizeMethod contains 4 enumerated resizing methods. + + # We select only 1 case for fast_mode bilinear. + num_resize_cases = 1 if fast_mode else 4 + distorted_image = apply_with_random_selector( + distorted_image, + lambda x, method: tf.image.resize_images(x, [height, width], method), + num_cases=num_resize_cases) + + if add_image_summaries: + tf.summary.image('cropped_resized_image', + tf.expand_dims(distorted_image, 0)) + + # Randomly flip the image horizontally. + distorted_image = tf.image.random_flip_left_right(distorted_image) + + # Randomly distort the colors. There are 1 or 4 ways to do it. + if FLAGS.use_fast_color_distort: + distorted_image = distort_color_fast(distorted_image) + else: + num_distort_cases = 1 if fast_mode else 4 + distorted_image = apply_with_random_selector( + distorted_image, + lambda x, ordering: distort_color(x, ordering, fast_mode), + num_cases=num_distort_cases) + + if add_image_summaries: + tf.summary.image('final_distorted_image', + tf.expand_dims(distorted_image, 0)) + return distorted_image + + +def preprocess_for_eval(image, height, width, + central_fraction=0.875, scope=None): + """Prepare one image for evaluation. + + If height and width are specified it would output an image with that size by + applying resize_bilinear. + + If central_fraction is specified it would crop the central fraction of the + input image. + + Args: + image: 3-D Tensor of image. If dtype is tf.float32 then the range should be + [0, 1], otherwise it would converted to tf.float32 assuming that the range + is [0, MAX], where MAX is largest positive representable number for + int(8/16/32) data type (see `tf.image.convert_image_dtype` for details). + height: integer + width: integer + central_fraction: Optional Float, fraction of the image to crop. + scope: Optional scope for name_scope. + Returns: + 3-D float Tensor of prepared image. + """ + with tf.name_scope(scope, 'eval_image', [image, height, width]): + if image.dtype != tf.float32: + image = tf.image.convert_image_dtype(image, dtype=tf.float32) + # Crop the central region of the image with an area containing 87.5% of + # the original image. + if central_fraction: + image = tf.image.central_crop(image, central_fraction=central_fraction) + + if height and width: + # Resize the image to the specified height and width. + image = tf.expand_dims(image, 0) + image = tf.image.resize_bilinear(image, [height, width], + align_corners=False) + image = tf.squeeze(image, [0]) + image.set_shape([height, width, 3]) + return image + + +def preprocess_image(image, + output_height, + output_width, + is_training=False, + scaled_images=True, + bbox=None, + min_object_covered=0.1, + fast_mode=True, + add_image_summaries=False): + """Pre-process one image for training or evaluation. + + Args: + image: 3-D Tensor [height, width, channels] with the image. If dtype is + tf.float32 then the range should be [0, 1], otherwise it would converted + to tf.float32 assuming that the range is [0, MAX], where MAX is largest + positive representable number for int(8/16/32) data type (see + `tf.image.convert_image_dtype` for details). + output_height: integer, image expected height. + output_width: integer, image expected width. + is_training: Boolean. If true it would transform an image for train, + otherwise it would transform it for evaluation. + scaled_images: Whether to scale pixel values to the range [-1, 1]. + If set to false, pixel values are in the range [0, 1]. + bbox: 3-D float Tensor of bounding boxes arranged [1, num_boxes, coords] + where each coordinate is [0, 1) and the coordinates are arranged as + [ymin, xmin, ymax, xmax]. + min_object_covered: An optional `float`. Defaults to `0.1`. The cropped + area of the image must contain at least this fraction of any bounding box + supplied. + fast_mode: Optional boolean, if True avoids slower transformations. + add_image_summaries: Enable image summaries. + + Returns: + 3-D float Tensor containing an appropriately scaled image + + Raises: + ValueError: if user does not provide bounding box + """ + if is_training: + image = preprocess_for_train( + image, + output_height, + output_width, + bbox, + min_object_covered, + fast_mode, + add_image_summaries=add_image_summaries) + else: + image = preprocess_for_eval(image, output_height, output_width) + if scaled_images: + image = tf.subtract(image, 0.5) + image = tf.multiply(image, 2.0) + return image + diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/model_builder.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/model_builder.py new file mode 100644 index 000000000..532d0d288 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/model_builder.py @@ -0,0 +1,451 @@ + + +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +# Copyright 2022 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +"""Constructs a generic image model based on the hparams the user passes in. +""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import faulthandler +faulthandler.enable() + + +from npu_bridge.npu_init import * + +import functools + +import numpy as np +import tensorflow.compat.v1 as tf + +import network_utils +from tensorflow.contrib import layers as contrib_layers +from tensorflow.contrib import slim +from tensorflow.contrib.framework.python.ops import arg_scope + + +def _build_loss(loss_fn, loss_name, logits, end_points, labels, + add_summary=False): + """Compute total loss based on the specified loss function.""" + # Collect all losses explicitly to sum up the total_loss. + losses = [] + + # Whethere to add aux loss is controled in network_fn. Once an aux head is + # built, an aux loss would be added here automatically. + aux_head_endpoint = None + if 'AuxLogits' in end_points: + # For Inception/Genet aux head. + aux_head_endpoint = end_points['AuxLogits'] + elif 'aux_logits' in end_points: + # For AmoebaNet aux head. + aux_head_endpoint = end_points['aux_logits'], + if aux_head_endpoint: + aux_loss = loss_fn( + labels, + tf.squeeze(aux_head_endpoint, axis=[0]), + weights=0.4, + scope='aux_loss') + tf.logging.info('Adding to aux loss.') + if add_summary: + tf.summary.scalar('losses/aux_loss', aux_loss) + + losses.append(aux_loss) + + # Add the empirical loss. + primary_loss = loss_fn(labels, logits, weights=1.0, scope=loss_name) + losses.append(primary_loss) + + # Add regularization losses. + reg_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES) + if reg_losses: + fp32_reg_losses = [] + for reg_loss in reg_losses: + fp32_reg_losses.append(tf.cast(reg_loss, tf.float32)) + reg_loss = tf.add_n(fp32_reg_losses, name='regularization_loss') + losses.append(reg_loss) + + total_loss = tf.add_n(losses, name='total_loss') + if add_summary: + tf.summary.scalar('losses/' + loss_name, primary_loss) + tf.summary.scalar('losses/regularization_loss', reg_loss) + tf.summary.scalar('losses/total_loss', total_loss) + + return total_loss + + +def build_softmax_loss(logits, + end_points, + labels, + label_smoothing=0.1, + add_summary=True): + loss_fn = functools.partial( + tf.losses.softmax_cross_entropy, label_smoothing=label_smoothing) + return _build_loss( + loss_fn=loss_fn, + loss_name='softmax_loss', + logits=logits, + end_points=end_points, + labels=labels, + add_summary=add_summary) + + +def compute_flops_per_example(batch_size): + # TODO(ereal): remove this function and other unnecessary reporting. + options = tf.profiler.ProfileOptionBuilder.float_operation() + options['output'] = 'none' + flops = ( + tf.profiler.profile( + tf.get_default_graph(), + options=options + ).total_float_ops / batch_size) + return flops + + +def build_learning_rate(initial_lr, + lr_decay_type, + global_step, + decay_factor=None, + decay_steps=None, + stepwise_epoch=None, + total_steps=None, + add_summary=True, + warmup_steps=0): + """Build learning rate.""" + if lr_decay_type == 'exponential': + assert decay_steps is not None + assert decay_factor is not None + lr = tf.train.exponential_decay( + initial_lr, global_step, decay_steps, decay_factor, staircase=True) + elif lr_decay_type == 'cosine': + assert total_steps is not None + lr = 0.5 * initial_lr * ( + 1 + tf.cos(np.pi * tf.cast(global_step, tf.float32) / total_steps)) + elif lr_decay_type == 'constant': + lr = initial_lr + elif lr_decay_type == 'stepwise': + assert stepwise_epoch is not None + boundaries = [ + 10 * stepwise_epoch, + 20 * stepwise_epoch, + ] + values = [initial_lr, initial_lr * 0.1, initial_lr * 0.01] + lr = tf.train.piecewise_constant(global_step, boundaries, values) + else: + assert False, 'Unknown lr_decay_type : %s' % lr_decay_type + + # By default, warmup_steps_fraction = 0.0 which means no warmup steps. + tf.logging.info('Learning rate warmup_steps: %d' % warmup_steps) + warmup_lr = ( + initial_lr * tf.cast(global_step, tf.float32) / tf.cast( + warmup_steps, tf.float32)) + lr = tf.cond(global_step < warmup_steps, lambda: warmup_lr, lambda: lr) + + if add_summary: + tf.summary.scalar('learning_rate', lr) + + return lr + + +def _build_aux_head(net, end_points, num_classes, hparams, scope): + """Auxiliary head used for all models across all datasets.""" + aux_scaling = 1.0 + # TODO(huangyp): double check aux_scaling with vrv@. + if hasattr(hparams, 'aux_scaling'): + aux_scaling = hparams.aux_scaling + tf.logging.info('aux scaling: {}'.format(aux_scaling)) + with tf.variable_scope(scope, custom_getter=network_utils.bp16_getter): + aux_logits = tf.identity(net) + with tf.variable_scope('aux_logits'): + aux_logits = slim.avg_pool2d( + aux_logits, [5, 5], stride=3, padding='VALID') + aux_logits = slim.conv2d(aux_logits, int(128 * aux_scaling), + [1, 1], scope='proj') + aux_logits = network_utils.batch_norm(aux_logits, scope='aux_bn0') + aux_logits = tf.nn.relu(aux_logits) + # Shape of feature map before the final layer. + shape = aux_logits.shape + if hparams.data_format == 'NHWC': + shape = shape[1:3] + else: + shape = shape[2:4] + aux_logits = slim.conv2d(aux_logits, int(768 * aux_scaling), + shape, padding='VALID') + aux_logits = network_utils.batch_norm(aux_logits, scope='aux_bn1') + aux_logits = tf.nn.relu(aux_logits) + aux_logits = contrib_layers.flatten(aux_logits) + aux_logits = slim.fully_connected(aux_logits, num_classes) + end_point_name = ( + 'aux_logits' if 'aux_logits' not in end_points else 'aux_logits_2') + end_points[end_point_name] = tf.cast(aux_logits, tf.float32) + + +def _imagenet_stem(inputs, hparams, stem_cell, filter_scaling_rate): + """Stem used for models trained on ImageNet.""" + # 149 x 149 x 32 + num_stem_filters = hparams.stem_reduction_size + with tf.variable_scope('stem', custom_getter=network_utils.bp16_getter): + net = slim.conv2d( + inputs, num_stem_filters, [3, 3], stride=2, scope='conv0', + padding='VALID') + net = network_utils.batch_norm(net, scope='conv0_bn') + tf.logging.info('imagenet_stem shape: {}'.format(net.shape)) + # Run the reduction cells + cell_outputs = [None, net] + filter_scaling = 1.0 / (filter_scaling_rate**hparams.num_stem_cells) + for cell_num in range(hparams.num_stem_cells): + net = stem_cell( + net, + scope='cell_stem_{}'.format(cell_num), + filter_scaling=filter_scaling, + stride=2, + prev_layer=cell_outputs[-2], + cell_num=cell_num) + cell_outputs.append(net) + filter_scaling *= filter_scaling_rate + tf.logging.info('imagenet_stem net shape at reduction layer {}: {}'.format( + cell_num, net.shape)) + # Only include cells in the cell_outputs. + return net, cell_outputs + + +def _basic_stem(inputs, hparams): + num_stem_filters = hparams.stem_reduction_size + with tf.variable_scope('stem', custom_getter=network_utils.bp16_getter): + net = slim.conv2d( + inputs, num_stem_filters, [3, 3], stride=1, scope='conv0', + padding='VALID') + net = network_utils.batch_norm(net, scope='conv0_bn') + tf.logging.info('basic_stem shape: {}'.format(net.shape)) + return net, [None, net] + + +def network_arg_scope(weight_decay=5e-5, + batch_norm_decay=0.9997, + batch_norm_epsilon=1e-3, + is_training=True, + data_format='NHWC', + num_shards=None, + distributed_group_size=1): + """Defines the default arg scope for the AmoebaNet ImageNet model. + + Args: + weight_decay: The weight decay to use for regularizing the model. + batch_norm_decay: Decay for batch norm moving average. + batch_norm_epsilon: Small float added to variance to avoid dividing by zero + in batch norm. + is_training: whether is training or not. + Useful for fine-tuning a model with different num_classes. + data_format: Input data format. + num_shards: Number of shards in the job + distributed_group_size: Size of the group to average for batch norm. + Returns: + An `arg_scope` to use for the AmoebaNet Model. + """ + batch_norm_params = { + # Decay for the moving averages. + 'decay': batch_norm_decay, + # epsilon to prevent 0s in variance. + 'epsilon': batch_norm_epsilon, + 'scale': True, + 'moving_vars': 'moving_vars', + 'is_training': is_training, + 'data_format': data_format, + 'num_shards': num_shards, + 'distributed_group_size': distributed_group_size, + } + weights_regularizer = contrib_layers.l2_regularizer(weight_decay) + weights_initializer = contrib_layers.variance_scaling_initializer( + mode='FAN_OUT') + with arg_scope([slim.fully_connected, slim.conv2d, slim.separable_conv2d], + weights_regularizer=weights_regularizer, + weights_initializer=weights_initializer): + with arg_scope([slim.fully_connected], + activation_fn=None, scope='FC'): + with arg_scope([slim.conv2d, slim.separable_conv2d], + activation_fn=None, biases_initializer=None): + with arg_scope([network_utils.batch_norm], **batch_norm_params): + with arg_scope( + [slim.dropout, network_utils.drop_path], is_training=is_training): + with arg_scope([slim.avg_pool2d, + slim.max_pool2d, + slim.conv2d, + slim.separable_conv2d, + network_utils.factorized_reduction, + network_utils.global_avg_pool, + network_utils.get_channel_index, + network_utils.get_channel_dim], + data_format=data_format) as sc: + return sc + + +def build_network(inputs, + num_classes, + is_training=True, + hparams=None): + """Builds an image model. + + Builds a model the takes inputs and return logits and endpoints. + + Args: + inputs: a tensor of size [batch_size, height, width, channels]. + num_classes: number of classes needed to be predicted by the model. If None, + only returns the feature vector endpoints after global_pool. + is_training: whether is training or not. + Useful for fine-tuning a model with different num_classes. + hparams: hparams used to construct the imagenet model. + + Returns: + a list containing 'logits', 'aux_logits' Tensors. + + Raises: + ValueError: upon invalid hparams. + """ + total_num_cells = (hparams.num_cells + + hparams.num_reduction_layers + + hparams.num_stem_cells) + normal_cell = network_utils.BaseCell( + hparams.reduction_size, hparams.normal_cell_operations, + hparams.normal_cell_used_hiddenstates, + hparams.normal_cell_hiddenstate_indices, hparams.drop_connect_keep_prob, + total_num_cells, hparams.drop_path_burn_in_steps) + reduction_cell = network_utils.BaseCell( + hparams.reduction_size, hparams.reduction_cell_operations, + hparams.reduction_cell_used_hiddenstates, + hparams.reduction_cell_hiddenstate_indices, + hparams.drop_connect_keep_prob, total_num_cells, + hparams.drop_path_burn_in_steps) + num_shards = hparams.num_shards + distributed_group_size = hparams.distributed_group_size + assert distributed_group_size == 1 or hparams.use_tpu + sc = network_arg_scope(weight_decay=hparams.weight_decay, + batch_norm_decay=hparams.batch_norm_decay, + batch_norm_epsilon=hparams.batch_norm_epsilon, + is_training=is_training, + data_format=hparams.data_format, + num_shards=num_shards, + distributed_group_size=distributed_group_size) + with arg_scope(sc): + return _build_network_base(inputs, + normal_cell=normal_cell, + reduction_cell=reduction_cell, + num_classes=num_classes, + hparams=hparams, + is_training=is_training) + + +def _build_network_base(images, + normal_cell, + reduction_cell, + num_classes, + hparams, + is_training): + """Constructs a AmoebaNet image model.""" + if hparams.get('use_bp16', False) and hparams.get('use_tpu', False): + images = tf.cast(images, dtype=tf.bfloat16) + end_points = {} + filter_scaling_rate = 2 + # Find where to place the reduction cells or stride normal cells + reduction_indices = network_utils.calc_reduction_layers( + hparams.num_cells, hparams.num_reduction_layers) + stem_cell = reduction_cell + + if hparams.stem_type == 'imagenet': + net, cell_outputs = _imagenet_stem(images, hparams, stem_cell, + filter_scaling_rate) + else: + net, cell_outputs = _basic_stem(images, hparams) + + # Setup for building in the auxiliary head. + aux_head_cell_idxes = [] + if len(reduction_indices) >= 2: + aux_head_cell_idxes.append(reduction_indices[1] - 1) + # Run the cells + filter_scaling = 1.0 + # true_cell_num accounts for the stem cells + true_cell_num = hparams.num_stem_cells + for cell_num in range(hparams.num_cells): + tf.logging.info('Current cell num: {}'.format(true_cell_num)) + + if cell_num in reduction_indices: + filter_scaling *= filter_scaling_rate + net = reduction_cell( + net, + scope='reduction_cell_{}'.format(reduction_indices.index(cell_num)), + filter_scaling=filter_scaling, + stride=2, + prev_layer=cell_outputs[-2], + cell_num=true_cell_num) + cell_outputs.append(net) + tf.logging.info('Reduction cell shape at layer {}: {}'.format( + true_cell_num, net.shape)) + true_cell_num += 1 + net = normal_cell( + net, + scope='cell_{}'.format(cell_num), + filter_scaling=filter_scaling, + stride=1, + prev_layer=cell_outputs[-2], + cell_num=true_cell_num) + if (hparams.use_aux_head and cell_num in aux_head_cell_idxes and + num_classes and is_training): + aux_net = tf.nn.relu(net) + _build_aux_head(aux_net, end_points, num_classes, hparams, + scope='aux_{}'.format(cell_num)) + cell_outputs.append(net) + tf.logging.info('Normal net shape at layer {}: {}'.format( + true_cell_num, net.shape)) + true_cell_num += 1 + + # Final softmax layer + with tf.variable_scope('final_layer', + custom_getter=network_utils.bp16_getter): + net = tf.nn.relu(net) + net = network_utils.global_avg_pool(net) + end_points['global_pool'] = net + if not num_classes: + return net, end_points + ''' + net = tf.to_float( + net, + name='ToFloat' + ) + net = tf.cast(net, tf.float16) + net = npu_ops.dropout(net, hparams.dense_dropout_keep_prob) + ''' + net = slim.dropout(net, hparams.dense_dropout_keep_prob, scope='dropout') + logits = slim.fully_connected(net, num_classes) + logits = tf.cast(logits, tf.float32) + predictions = tf.nn.softmax(logits, name='predictions') + end_points['logits'] = logits + end_points['predictions'] = predictions + end_points['cell_outputs'] = cell_outputs + return logits, end_points + diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/model_specs.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/model_specs.py new file mode 100644 index 000000000..787e44511 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/model_specs.py @@ -0,0 +1,115 @@ + + + +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +# Copyright 2022 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +"""Configs for various AmoebaNet architectures.""" + + + + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + + +import faulthandler +faulthandler.enable() + + +from npu_bridge.npu_init import * +from npu_bridge.estimator import npu_ops + +def get_normal_cell(cell_name): + """Return normal cell spec.""" + operations = [] + hiddenstate_indices = [] + used_hiddenstates = [] + if cell_name == 'evol_net_g' or cell_name == 'amoeba_net_a': + operations = ['avg_pool_3x3', 'max_pool_3x3', 'separable_3x3_2', 'none', + 'none', 'avg_pool_3x3', 'separable_3x3_2', 'separable_5x5_2', + 'avg_pool_3x3', 'separable_3x3_2'] + hiddenstate_indices = [0, 0, 1, 1, 0, 1, 0, 2, 5, 0] + used_hiddenstates = [1, 0, 1, 0, 0, 1, 0] + elif cell_name == 'evol_net_h' or cell_name == 'amoeba_net_b': + operations = ['1x1', 'max_pool_3x3', 'none', 'separable_3x3_2', '1x1', + 'separable_3x3_2', '1x1', 'none', 'avg_pool_3x3', '1x1'] + hiddenstate_indices = [1, 1, 1, 0, 1, 0, 2, 2, 1, 5] + used_hiddenstates = [0, 1, 1, 0, 0, 1, 0] + elif cell_name == 'evol_net_a' or cell_name == 'amoeba_net_c': + operations = ['avg_pool_3x3', 'separable_3x3_2', 'none', 'separable_3x3_2', + 'avg_pool_3x3', 'separable_3x3_2', 'none', 'separable_3x3_2', + 'avg_pool_3x3', 'separable_3x3_2'] + hiddenstate_indices = [0, 0, 0, 0, 2, 1, 0, 1, 3, 0] + used_hiddenstates = [1, 0, 0, 1, 0, 0, 0] + elif cell_name == 'evol_net_x' or cell_name == 'amoeba_net_d': + operations = ['1x1', 'max_pool_3x3', 'none', '1x7_7x1', '1x1', '1x7_7x1', + 'max_pool_3x3', 'none', 'avg_pool_3x3', '1x1'] + hiddenstate_indices = [1, 1, 1, 0, 0, 0, 2, 2, 1, 5] + used_hiddenstates = [0, 1, 1, 0, 0, 1, 0] + else: + raise ValueError('Unsupported cell name: %s.' % cell_name) + return operations, hiddenstate_indices, used_hiddenstates + + +def get_reduction_cell(cell_name): + """Return reduction cell spec.""" + operations = [] + hiddenstate_indices = [] + used_hiddenstates = [] + if cell_name == 'evol_net_g' or cell_name == 'amoeba_net_a': + operations = ['separable_3x3_2', 'avg_pool_3x3', 'max_pool_3x3', + 'separable_7x7_2', 'max_pool_3x3', 'max_pool_3x3', + 'separable_3x3_2', '1x7_7x1', 'avg_pool_3x3', + 'separable_7x7_2'] + hiddenstate_indices = [1, 0, 0, 2, 1, 0, 4, 0, 1, 0] + used_hiddenstates = [1, 1, 0, 0, 0, 0, 0] + elif cell_name == 'evol_net_h' or cell_name == 'amoeba_net_b': + operations = ['max_pool_2x2', 'max_pool_3x3', 'none', '3x3', + 'dil_2_separable_5x5_2', 'max_pool_3x3', 'none', + 'separable_3x3_2', 'avg_pool_3x3', '1x1'] + hiddenstate_indices = [0, 0, 2, 1, 2, 2, 3, 1, 4, 3] + used_hiddenstates = [1, 1, 1, 1, 1, 0, 0] + elif cell_name == 'evol_net_a' or cell_name == 'amoeba_net_c': + operations = ['max_pool_3x3', 'max_pool_3x3', 'separable_7x7_2', + 'separable_3x3_2', 'separable_7x7_2', 'max_pool_3x3', + 'separable_5x5_2', 'separable_5x5_2', 'max_pool_3x3', + 'separable_3x3_2'] + hiddenstate_indices = [0, 0, 2, 0, 0, 1, 4, 4, 1, 1] + used_hiddenstates = [0, 1, 0, 0, 0, 0, 0] + elif cell_name == 'evol_net_x' or cell_name == 'amoeba_net_d': + operations = ['max_pool_2x2', 'max_pool_3x3', 'none', '3x3', '1x7_7x1', + 'max_pool_3x3', 'none', 'max_pool_2x2', 'avg_pool_3x3', + '1x1'] + hiddenstate_indices = [0, 0, 2, 1, 2, 2, 3, 1, 2, 3] + used_hiddenstates = [1, 1, 1, 1, 0, 0, 0] + else: + raise ValueError('Unsupported cell name: %s.' % cell_name) + return operations, hiddenstate_indices, used_hiddenstates + diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/network_utils.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/network_utils.py new file mode 100644 index 000000000..ae5e85fac --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/network_utils.py @@ -0,0 +1,791 @@ + + + + +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +# Copyright 2022 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +"""A custom module for some common operations used by AmoebaNet. + +Functions exposed in this file: +- bp16_getter +- calc_reduction_layers +- get_channel_index +- get_channel_dim +- global_avg_pool +- factorized_reduction +- drop_path + +Classes exposed in this file: +- BaseCell +""" + + + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import faulthandler +faulthandler.enable() + + +from npu_bridge.npu_init import * +from npu_bridge.estimator import npu_ops + +import tensorflow.compat.v1 as tf + +from tensorflow.contrib import framework as contrib_framework +from tensorflow.contrib import slim +from tensorflow.contrib.framework.python.ops import add_arg_scope +from tensorflow.contrib.tpu.python.ops import tpu_ops +from tensorflow.python.training import moving_averages # pylint: disable=g-direct-tensorflow-import + +DATA_FORMAT_NCHW = 'NCHW' +DATA_FORMAT_NHWC = 'NHWC' +INVALID = 'null' + + +def cross_replica_average(inputs, num_shards, distributed_group_size): + """Calculates the average value of inputs tensor across TPU replicas.""" + group_assignment = None + if num_shards is not None and distributed_group_size != num_shards: + group_size = distributed_group_size + group_assignment = [] + for g in range(num_shards // group_size): + replica_ids = [g * group_size + i for i in range(group_size)] + group_assignment.append(replica_ids) + + return tpu_ops.cross_replica_sum(inputs, group_assignment) / tf.cast( + distributed_group_size, inputs.dtype) + + +def bp16_getter(getter, *args, **kwargs): + """Returns a custom getter that this class's methods must be called.""" + cast_to_bfloat16 = False + requested_dtype = kwargs['dtype'] + if requested_dtype == tf.bfloat16: + # Keep a master copy of weights in fp32 and cast to bp16 when the weights + # are used. + kwargs['dtype'] = tf.float32 + cast_to_bfloat16 = True + var = getter(*args, **kwargs) + # This if statement is needed to guard the cast, because batch norm + # assigns directly to the return value of this custom getter. The cast + # makes the return value not a variable so it cannot be assigned. Batch + # norm variables are always in fp32 so this if statement is never + # triggered for them. + if cast_to_bfloat16: + var = tf.cast(var, tf.bfloat16) + return var + + +def calc_reduction_layers(num_cells, num_reduction_layers): + """Figure out what layers should have reductions.""" + reduction_layers = [] + for pool_num in range(1, num_reduction_layers + 1): + layer_num = (float(pool_num) / (num_reduction_layers + 1)) * num_cells + layer_num = int(layer_num) + reduction_layers.append(layer_num) + return reduction_layers + + +@add_arg_scope +def get_channel_index(data_format=INVALID): + assert data_format != INVALID + axis = 3 if data_format == 'NHWC' else 1 + return axis + + +@add_arg_scope +def get_channel_dim(shape, data_format=INVALID): + assert data_format != INVALID + assert len(shape) == 4 + if data_format == 'NHWC': + return int(shape[3]) + elif data_format == 'NCHW': + return int(shape[1]) + else: + raise ValueError('Not a valid data_format', data_format) + + +@add_arg_scope +def global_avg_pool(x, data_format=INVALID): + """Average pool away the height and width spatial dimensions of x.""" + assert data_format != INVALID + assert data_format in ['NHWC', 'NCHW'] + assert x.shape.ndims == 4 + if data_format == 'NHWC': + return tf.reduce_mean(x, [1, 2]) + else: + return tf.reduce_mean(x, [2, 3]) + + +@add_arg_scope +def factorized_reduction(net, output_filters, stride, data_format=INVALID): + """Reduces the shape of net without information loss due to striding.""" + assert output_filters % 2 == 0, ( + 'Need even number of filters when using this factorized reduction.') + assert data_format != INVALID + if stride == 1: + net = slim.conv2d(net, output_filters, 1, scope='path_conv') + net = batch_norm(net, scope='path_bn') + return net + if data_format == 'NHWC': + stride_spec = [1, stride, stride, 1] + else: + stride_spec = [1, 1, stride, stride] + + # Skip path 1 + path1 = tf.nn.avg_pool( + net, [1, 1, 1, 1], stride_spec, 'VALID', data_format=data_format) + path1 = slim.conv2d(path1, int(output_filters / 2), 1, scope='path1_conv') + + # Skip path 2 + # First pad with 0's on the right and bottom, then shift the filter to + # include those 0's that were added. + if data_format == 'NHWC': + pad_arr = [[0, 0], [0, 1], [0, 1], [0, 0]] + path2 = tf.pad(net, pad_arr)[:, 1:, 1:, :] + concat_axis = 3 + else: + pad_arr = [[0, 0], [0, 0], [0, 1], [0, 1]] + path2 = tf.pad(net, pad_arr)[:, :, 1:, 1:] + concat_axis = 1 + + path2 = tf.nn.avg_pool( + path2, [1, 1, 1, 1], stride_spec, 'VALID', data_format=data_format) + path2 = slim.conv2d(path2, int(output_filters / 2), 1, scope='path2_conv') + + # Concat and apply BN + final_path = tf.concat(values=[path1, path2], axis=concat_axis) + final_path = batch_norm(final_path, scope='final_path_bn') + return final_path + + +@add_arg_scope +def drop_path(net, keep_prob, is_training=True): + """Drops out a whole example hiddenstate with the specified probability.""" + if is_training: + batch_size = tf.shape(net)[0] + noise_shape = [batch_size, 1, 1, 1] + keep_prob = tf.cast(keep_prob, dtype=net.dtype) + random_tensor = keep_prob + random_tensor += tf.random_uniform(noise_shape, dtype=net.dtype) + binary_tensor = tf.floor(random_tensor) + net = tf.div(net, keep_prob) * binary_tensor + return net + + +def _operation_to_filter_shape(operation): + splitted_operation = operation.split('x') + filter_shape = int(splitted_operation[0][-1]) + assert filter_shape == int( + splitted_operation[1][0]), 'Rectangular filters not supported.' + return filter_shape + + +def _operation_to_num_layers(operation): + splitted_operation = operation.split('_') + if 'x' in splitted_operation[-1]: + return 1 + return int(splitted_operation[-1]) + + +def _operation_to_info(operation): + """Takes in operation name and returns meta information. + + An example would be 'separable_3x3_4' -> (3, 4). + + Args: + operation: String that corresponds to convolution operation. + + Returns: + Tuple of (filter shape, num layers). + """ + num_layers = _operation_to_num_layers(operation) + filter_shape = _operation_to_filter_shape(operation) + return num_layers, filter_shape + + +def _stacked_separable_conv(net, stride, operation, filter_size): + """Takes in an operations and parses it to the correct sep operation.""" + num_layers, kernel_size = _operation_to_info(operation) + for layer_num in range(num_layers - 1): + net = tf.nn.relu(net) + net = slim.separable_conv2d( + net, + filter_size, + kernel_size, + depth_multiplier=1, + scope='separable_{0}x{0}_{1}'.format(kernel_size, layer_num + 1), + stride=stride) + net = batch_norm( + net, scope='bn_sep_{0}x{0}_{1}'.format(kernel_size, layer_num + 1)) + stride = 1 + net = tf.nn.relu(net) + net = slim.separable_conv2d( + net, + filter_size, + kernel_size, + depth_multiplier=1, + scope='separable_{0}x{0}_{1}'.format(kernel_size, num_layers), + stride=stride) + net = batch_norm( + net, scope='bn_sep_{0}x{0}_{1}'.format(kernel_size, num_layers)) + return net + + +def _operation_to_pooling_type(operation): + """Takes in the operation string and returns the pooling type.""" + splitted_operation = operation.split('_') + return splitted_operation[0] + + +def _operation_to_pooling_shape(operation): + """Takes in the operation string and returns the pooling kernel shape.""" + splitted_operation = operation.split('_') + shape = splitted_operation[-1] + assert 'x' in shape + filter_height, filter_width = shape.split('x') + assert filter_height == filter_width + return int(filter_height) + + +def _operation_to_pooling_info(operation): + """Parses the pooling operation string to return its type and shape.""" + pooling_type = _operation_to_pooling_type(operation) + pooling_shape = _operation_to_pooling_shape(operation) + return pooling_type, pooling_shape + + +def _pooling(net, stride, operation): + """Parses operation and performs the correct pooling operation on net.""" + padding = 'SAME' + pooling_type, pooling_shape = _operation_to_pooling_info(operation) + if pooling_type == 'avg': + net = slim.avg_pool2d(net, pooling_shape, stride=stride, padding=padding) + elif pooling_type == 'max': + net = slim.max_pool2d(net, pooling_shape, stride=stride, padding=padding) + elif pooling_type == 'min': + net = slim.max_pool2d(-1 * net, pooling_shape, stride=stride, + padding=padding) + net = -1 * net + else: + raise NotImplementedError('Unimplemented pooling type: ', pooling_type) + return net + + +class BaseCell(object): + """Base Cell class that is used as a 'layer' in image architectures. + + Args: + num_conv_filters: The number of filters for each convolution operation. + operations: List of operations that are performed in the AmoebaNet Cell in + order. + used_hiddenstates: Binary array that signals if the hiddenstate was used + within the cell. This is used to determine what outputs of the cell + should be concatenated together. + hiddenstate_indices: Determines what hiddenstates should be combined + together with the specified operations to create the AmoebaNet cell. + """ + + def __init__(self, num_conv_filters, operations, used_hiddenstates, + hiddenstate_indices, drop_path_keep_prob, total_num_cells, + drop_path_burn_in_steps): + self._num_conv_filters = num_conv_filters + self._operations = operations + self._used_hiddenstates = used_hiddenstates + self._hiddenstate_indices = hiddenstate_indices + self._drop_path_keep_prob = drop_path_keep_prob + self._total_num_cells = total_num_cells + self._drop_path_burn_in_steps = drop_path_burn_in_steps + + def _reduce_prev_layer(self, prev_layer, curr_layer): + """Matches dimension of prev_layer to the curr_layer.""" + # Set the prev layer to the current layer if it is none + if prev_layer is None: + return curr_layer + curr_num_filters = self._filter_size + prev_num_filters = get_channel_dim(prev_layer.shape) + curr_filter_shape = int(curr_layer.shape[2]) + prev_filter_shape = int(prev_layer.shape[2]) + if curr_filter_shape != prev_filter_shape: + prev_layer = tf.nn.relu(prev_layer) + prev_layer = factorized_reduction( + prev_layer, curr_num_filters, stride=2) + elif curr_num_filters != prev_num_filters: + prev_layer = tf.nn.relu(prev_layer) + prev_layer = slim.conv2d( + prev_layer, curr_num_filters, 1, scope='prev_1x1') + prev_layer = batch_norm(prev_layer, scope='prev_bn') + return prev_layer + + def _cell_base(self, net, prev_layer): + """Runs the beginning of the conv cell before the predicted ops are run.""" + num_filters = self._filter_size + + # Check to be sure prev layer stuff is setup correctly + prev_layer = self._reduce_prev_layer(prev_layer, net) + + net = tf.nn.relu(net) + net = slim.conv2d(net, num_filters, 1, scope='1x1') + net = batch_norm(net, scope='beginning_bn') + split_axis = get_channel_index() + net = tf.split(axis=split_axis, num_or_size_splits=1, value=net) + for split in net: + assert int(split.shape[split_axis] == int(self._num_conv_filters * + self._filter_scaling)) + net.append(prev_layer) + return net + + def __call__(self, net, scope=None, filter_scaling=1, stride=1, + prev_layer=None, cell_num=-1): + """Runs the conv cell.""" + self._cell_num = cell_num + self._filter_scaling = filter_scaling + self._filter_size = int(self._num_conv_filters * filter_scaling) + + i = 0 + with tf.variable_scope(scope, custom_getter=bp16_getter): + net = self._cell_base(net, prev_layer) + for iteration in range(5): + with tf.variable_scope('comb_iter_{}'.format(iteration)): + left_hiddenstate_idx, right_hiddenstate_idx = ( + self._hiddenstate_indices[i], + self._hiddenstate_indices[i + 1]) + original_input_left = left_hiddenstate_idx < 2 + original_input_right = right_hiddenstate_idx < 2 + h1 = net[left_hiddenstate_idx] + h2 = net[right_hiddenstate_idx] + + operation_left = self._operations[i] + operation_right = self._operations[i+1] + i += 2 + # Apply conv operations + with tf.variable_scope('left'): + h1 = self._apply_operation(h1, operation_left, + stride, original_input_left) + with tf.variable_scope('right'): + h2 = self._apply_operation(h2, operation_right, + stride, original_input_right) + + # Combine hidden states using 'add'. + with tf.variable_scope('combine'): + h = h1 + h2 + + # Add hiddenstate to the list of hiddenstates we can choose from + net.append(h) + + with tf.variable_scope('cell_output'): + net = self._combine_unused_states(net) + + return net + + def _apply_conv_operation(self, net, operation, stride, filter_size): + """Takes in a hiddenstate and applies an operation to it. + + Args: + net: This is a hiddenstate from the network that will have an operation + applied to it. + operation: This is a string that specifies what operations should be + applied to net. + stride: Stride for the operations being passed in. + filter_size: Number of filters output from this operation. + + Returns: + The hiddenstate net after it had the operation passed in applied to it. + """ + + if operation == '1x1': + net = slim.conv2d(net, filter_size, 1, scope='1x1') + elif operation == '3x3': + net = slim.conv2d(net, filter_size, 3, scope='3x3', stride=stride) + elif operation == '1x7_7x1': + net = slim.conv2d(net, filter_size, [1, 7], scope='1x7', + stride=[1, stride]) + net = batch_norm(net, scope='bn_1x7_7x1') + net = tf.nn.relu(net) + net = slim.conv2d(net, filter_size, [7, 1], scope='7x1', + stride=[stride, 1]) + elif operation == '1x3_3x1': + net = slim.conv2d(net, filter_size, [1, 3], scope='1x3', + stride=[1, stride]) + net = batch_norm(net, scope='bn_1x3_3x1') + net = tf.nn.relu(net) + net = slim.conv2d(net, filter_size, [3, 1], scope='3x1', + stride=[stride, 1]) + elif operation in ['dilated_3x3_rate_2', 'dilated_3x3_rate_4', + 'dilated_3x3_rate_6']: + dilation_rate = int(operation.split('_')[-1]) + net = slim.conv2d( + net, + filter_size, + 3, + rate=dilation_rate, + stride=stride, + scope='dilated_3x3') + else: + raise NotImplementedError('Unimplemented conv operation: ', operation) + return net + + def _apply_operation(self, net, operation, + stride, is_from_original_input): + """Applies the predicted conv operation to net.""" + # Dont stride if this is not one of the original hiddenstates + if stride > 1 and not is_from_original_input: + stride = 1 + input_filters = get_channel_dim(net.shape) + filter_size = self._filter_size + if 'separable' in operation: + net = _stacked_separable_conv(net, stride, operation, filter_size) + elif operation in ['dilated_3x3_rate_2', 'dilated_3x3_rate_4', + 'dilated_3x3_rate_6', '3x3', '1x7_7x1', '1x3_3x1']: + if operation == '1x3_3x1': + reduced_filter_size = int(3 * filter_size / 8) + else: + reduced_filter_size = int(filter_size / 4) + if reduced_filter_size < 1: + # If the intermediate number of channels would be too small, then don't + # use a bottleneck layer. + net = tf.nn.relu(net) + net = self._apply_conv_operation(net, operation, stride, filter_size) + net = batch_norm(net, scope='bn') + else: + # Use a bottleneck layer. + net = tf.nn.relu(net) + net = slim.conv2d(net, reduced_filter_size, 1, scope='pre_1x1') + net = batch_norm(net, scope='pre_bn') + net = tf.nn.relu(net) + net = self._apply_conv_operation( + net, operation, stride, reduced_filter_size) + net = batch_norm(net, scope='bn') + net = tf.nn.relu(net) + net = slim.conv2d(net, filter_size, 1, scope='post_1x1') + net = batch_norm(net, scope='post_bn') + elif operation in ['none', '1x1']: + # Check if a stride is needed, then use a strided 1x1 here + if stride > 1 or operation == '1x1' or (input_filters != filter_size): + net = tf.nn.relu(net) + net = slim.conv2d(net, filter_size, 1, stride=stride, scope='1x1') + net = batch_norm(net, scope='bn_1') + elif 'pool' in operation: + net = _pooling(net, stride, operation) + if input_filters != filter_size: + net = slim.conv2d(net, filter_size, 1, stride=1, scope='1x1') + net = batch_norm(net, scope='bn_1') + else: + raise ValueError('Unimplemented operation', operation) + + if operation != 'none': + net = self._apply_drop_path(net) + + tf.logging.info('Net shape after {}: {}'.format(operation, net.shape)) + return net + + def _combine_unused_states(self, net): + """Concatenate the unused hidden states of the cell.""" + used_hiddenstates = self._used_hiddenstates + + final_height = int(net[-1].shape[2]) + final_num_filters = get_channel_dim(net[-1].shape) + assert len(used_hiddenstates) == len(net) + for idx, used_h in enumerate(used_hiddenstates): + curr_height = int(net[idx].shape[2]) + curr_num_filters = get_channel_dim(net[idx].shape) + + # Determine if a reduction should be applied to make the number of + # filters match. + should_reduce = final_num_filters != curr_num_filters + should_reduce = (final_height != curr_height) or should_reduce + should_reduce = should_reduce and not used_h + if should_reduce: + stride = 2 if final_height != curr_height else 1 + with tf.variable_scope('reduction_{}'.format(idx)): + net[idx] = factorized_reduction( + net[idx], final_num_filters, stride) + + states_to_combine = ( + [h for h, is_used in zip(net, used_hiddenstates) if not is_used]) + + # Return the concat of all the states + concat_axis = get_channel_index() + net = tf.concat(values=states_to_combine, axis=concat_axis) + return net + + @add_arg_scope # No public API. For internal use only. + def _apply_drop_path(self, net, current_step=None, + drop_connect_version='v1'): + """Apply drop_path regularization. + + Args: + net: the Tensor that gets drop_path regularization applied. + current_step: a float32 Tensor with the current global_step value, + to be divided by hparams.drop_path_burn_in_steps. Usually None, which + defaults to tf.train.get_or_create_global_step() properly casted. + drop_connect_version: one of 'v1', 'v2', 'v3', controlling whether + the dropout rate is scaled by current_step (v1, the default), + layer (v2), or both (v3). + + Returns: + The dropped-out value of `net`. + """ + drop_path_keep_prob = self._drop_path_keep_prob + if drop_path_keep_prob < 1.0: + assert drop_connect_version in ['v1', 'v2', 'v3'] + if drop_connect_version in ['v2', 'v3']: + # Scale keep prob by layer number + assert self._cell_num != -1 + # The added 2 is for the reduction cells + num_cells = self._total_num_cells + layer_ratio = (self._cell_num + 1)/float(num_cells) + drop_path_keep_prob = 1 - layer_ratio * (1 - drop_path_keep_prob) + if drop_connect_version in ['v1', 'v3']: + if self._drop_path_burn_in_steps: + # Decrease the keep probability over time + current_step = ( + current_step or + tf.cast(tf.train.get_or_create_global_step(), tf.float32)) + current_ratio = current_step / self._drop_path_burn_in_steps + current_ratio = tf.minimum(1.0, current_ratio) + drop_path_keep_prob = (1 - current_ratio * (1 - drop_path_keep_prob)) + net = drop_path(net, drop_path_keep_prob) + return net + + +# TODO(huangyp): find out the difference and use the layers batch_norm. +@add_arg_scope +def batch_norm(inputs, + decay=0.999, + center=True, + scale=False, + epsilon=0.001, + moving_vars='moving_vars', + activation_fn=None, + is_training=True, + data_format='NHWC', + reuse=None, + num_shards=None, + distributed_group_size=1, + scope=None): + """Adds a Batch Normalization layer from http://arxiv.org/abs/1502.03167. + + "Batch Normalization: Accelerating Deep Network Training by Reducing + Internal Covariate Shift" + + Sergey Ioffe, Christian Szegedy + + Can be used as a normalizer function for conv2d and fully_connected. + + Note: When is_training is True the moving_mean and moving_variance need to be + updated, by default the update_ops are placed in `tf.GraphKeys.UPDATE_OPS` so + they need to be added as a dependency to the `train_op`, example: + + update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) + if update_ops: + updates = tf.group(*update_ops) + total_loss = control_flow_ops.with_dependencies([updates], total_loss) + + One can set updates_collections=None to force the updates in place, but that + can have speed penalty, especially in distributed settings. + + Args: + inputs: A tensor with 2 or more dimensions, where the first dimension has + `batch_size`. The normalization is over all but the last dimension if + `data_format` is `NHWC` and the second dimension if `data_format` is + `NCHW`. + decay: Decay for the moving average. Reasonable values for `decay` are close + to 1.0, typically in the multiple-nines range: 0.999, 0.99, 0.9, etc. + Lower `decay` value (recommend trying `decay`=0.9) if model experiences + reasonably good training performance but poor validation and/or test + performance. + center: If True, add offset of `beta` to normalized tensor. If False, + `beta` is ignored. + scale: If True, multiply by `gamma`. If False, `gamma` is + not used. When the next layer is linear (also e.g. `nn.relu`), this can be + disabled since the scaling can be done by the next layer. + epsilon: Small float added to variance to avoid dividing by zero. + moving_vars: Name of collection created for moving variables. + activation_fn: Activation function, default set to None to skip it and + maintain a linear activation. + is_training: Whether or not the layer is in training mode. In training mode + it would accumulate the statistics of the moments into `moving_mean` and + `moving_variance` using an exponential moving average with the given + `decay`. When it is not in training mode then it would use the values of + the `moving_mean` and the `moving_variance`. + data_format: input data format. NHWC or NCHW + reuse: Whether or not the layer and its variables should be reused. To be + able to reuse the layer scope must be given. + num_shards: Number of shards that participate in the global + reduction. Default is set to None, that will skip the cross replica sum in + and normalize across local examples only. + distributed_group_size: Number of replicas to normalize across in the + distributed batch normalization. + scope: Optional scope for `variable_scope`. + + Returns: + A `Tensor` representing the output of the operation. + + Raises: + ValueError: If the rank of `inputs` is undefined. + ValueError: If the rank of `inputs` is neither 2 or 4. + ValueError: If rank or `C` dimension of `inputs` is undefined. + """ + trainable = True + + with tf.variable_scope(scope, 'BatchNorm', [inputs], reuse=reuse): + inputs = tf.convert_to_tensor(inputs) + original_shape = inputs.get_shape() + original_rank = original_shape.ndims + if original_rank is None: + raise ValueError('Inputs %s has undefined rank' % inputs.name) + elif original_rank not in [2, 4]: + raise ValueError('Inputs %s has unsupported rank.' + ' Expected 2 or 4 but got %d' % (inputs.name, + original_rank)) + if original_rank == 2: + channels = inputs.get_shape()[-1].value + if channels is None: + raise ValueError('`C` dimension must be known but is None') + new_shape = [-1, 1, 1, channels] + + if data_format == 'NCHW': + new_shape = [-1, channels, 1, 1] + inputs = tf.reshape(inputs, new_shape) + inputs_shape = inputs.get_shape() + if data_format == 'NHWC': + params_shape = inputs_shape[-1:] + else: + params_shape = inputs_shape[1:2] + if not params_shape.is_fully_defined(): + raise ValueError('Inputs %s has undefined `C` dimension %s.' % + (inputs.name, params_shape)) + + # Allocate parameters for the beta and gamma of the normalization. + trainable_beta = trainable and center + collections = [tf.GraphKeys.MODEL_VARIABLES, tf.GraphKeys.GLOBAL_VARIABLES] + beta = contrib_framework.variable( + 'beta', + params_shape, + collections=collections, + initializer=tf.zeros_initializer(), + trainable=trainable_beta) + trainable_gamma = trainable and scale + gamma = contrib_framework.variable( + 'gamma', + params_shape, + collections=collections, + initializer=tf.ones_initializer(), + trainable=trainable_gamma) + + # Create moving_mean and moving_variance variables and add them to the + # appropiate collections. + moving_collections = [moving_vars, + tf.GraphKeys.MOVING_AVERAGE_VARIABLES, + tf.GraphKeys.MODEL_VARIABLES, + tf.GraphKeys.GLOBAL_VARIABLES] + # Disable partition setting for moving_mean and moving_variance + # as assign_moving_average op below doesn't support partitioned variable. + scope = tf.get_variable_scope() + partitioner = scope.partitioner + scope.set_partitioner(None) + moving_mean = contrib_framework.variable( + 'moving_mean', + params_shape, + initializer=tf.zeros_initializer(), + trainable=False, + collections=moving_collections) + moving_variance = contrib_framework.variable( + 'moving_variance', + params_shape, + initializer=tf.ones_initializer(), + trainable=False, + collections=moving_collections) + # Restore scope's partitioner setting. + scope.set_partitioner(partitioner) + + # Add cross replica sum to do subset mean and variance calculation + # First compute mean and variance + if is_training: + if distributed_group_size > 1: + # Execute a distributed batch normalization + if data_format == 'NCHW': + axis = 1 + else: + axis = 3 + input_shape = inputs.get_shape() + inputs_dtype = inputs.dtype + inputs = tf.cast(inputs, tf.float32) + ndims = len(input_shape) + reduction_axes = [i for i in range(ndims) if i != axis] + counts, mean_ss, variance_ss, _ = tf.nn.sufficient_statistics( + inputs, reduction_axes, keep_dims=False) + mean_ss = cross_replica_average(mean_ss, num_shards, + distributed_group_size) + variance_ss = cross_replica_average(variance_ss, num_shards, + distributed_group_size) + mean, variance = tf.nn.normalize_moments( + counts, mean_ss, variance_ss, shift=None) + outputs = tf.nn.batch_normalization(inputs, mean, variance, beta, gamma, + epsilon) + outputs = tf.cast(outputs, inputs_dtype) + else: + with tf.compat.forward_compatibility_horizon(2019, 5, 1): + outputs, mean, variance = tf.nn.fused_batch_norm(inputs, gamma, beta, epsilon=epsilon, data_format=data_format) + else: + with tf.compat.forward_compatibility_horizon(2019, 5, 1): + outputs, mean, variance = tf.nn.fused_batch_norm( + inputs, + gamma, + beta, + mean=moving_mean, + variance=moving_variance, + epsilon=epsilon, + is_training=False, + data_format=data_format) + + if is_training: + update_moving_mean = moving_averages.assign_moving_average( + moving_mean, + tf.cast(mean, moving_mean.dtype), + decay, + zero_debias=False) + update_moving_variance = moving_averages.assign_moving_average( + moving_variance, + tf.cast(variance, moving_variance.dtype), + decay, + zero_debias=False) + tf.add_to_collection('update_ops', update_moving_mean) + tf.add_to_collection('update_ops', update_moving_variance) + outputs.set_shape(inputs_shape) + if original_shape.ndims == 2: + outputs = tf.reshape(outputs, original_shape) + if activation_fn is not None: + outputs = activation_fn(outputs) + return outputs + diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/network_utils_test.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/network_utils_test.py new file mode 100644 index 000000000..62c24ad58 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/network_utils_test.py @@ -0,0 +1,88 @@ + + +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +# Copyright 2022 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +"""Tests for slim.nets.nasnet.nasnet_utils.""" + + + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + + +import faulthandler +faulthandler.enable() + + +from npu_bridge.npu_init import * + +import tensorflow.compat.v1 as tf + +import network_utils + + +class NetworkUtilsTest(tf.test.TestCase): + + def testCalcReductionLayers(self): + num_cells = 18 + num_reduction_layers = 2 + reduction_layers = network_utils.calc_reduction_layers( + num_cells, num_reduction_layers) + self.assertEqual(len(reduction_layers), 2) + self.assertEqual(reduction_layers[0], 6) + self.assertEqual(reduction_layers[1], 12) + + def testGetChannelIndex(self): + data_formats = ['NHWC', 'NCHW'] + for data_format in data_formats: + index = network_utils.get_channel_index(data_format) + correct_index = 3 if data_format == 'NHWC' else 1 + self.assertEqual(index, correct_index) + + def testGetChannelDim(self): + data_formats = ['NHWC', 'NCHW'] + shape = [10, 20, 30, 40] + for data_format in data_formats: + dim = network_utils.get_channel_dim(shape, data_format) + correct_dim = shape[3] if data_format == 'NHWC' else shape[1] + self.assertEqual(dim, correct_dim) + + def testGlobalAvgPool(self): + data_formats = ['NHWC', 'NCHW'] + inputs = tf.placeholder(tf.float32, (5, 10, 20, 10)) + for data_format in data_formats: + output = network_utils.global_avg_pool( + inputs, data_format) + self.assertEqual(output.shape, [5, 10]) + + +if __name__ == '__main__': + tf.test.main() + diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/online_inference_testcase.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/online_inference_testcase.sh new file mode 100644 index 000000000..f241eb3e5 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/online_inference_testcase.sh @@ -0,0 +1,39 @@ +#! /bin/bash + +DATA_DIR=/home/test_user03/tf_records/ +MODEL_DIR=/home/test_user03/hh + +#߸˶ԤõݼԤѵģ͡ATC-OMģ͵ȣִ֧OBS +#obsutil cp obs://obsxxx/xxx/xxx.ckpt ./model/ -f -r + +#testcase壬߸ݲͬģд +python3 amoeba_net.py \ + --data_dir=${DATA_DIR} \ + --model_dir=${MODEL_DIR} \ + --num_cells=6 \ + --image_size=224 \ + --num_epochs=1 \ + --train_batch_size=64 \ + --eval_batch_size=64 \ + --lr=2.56 \ + --lr_decay_value=0.88 \ + --lr_warmup_epochs=0.35 \ + --mode=predict \ + --iterations_per_loop=1251 \ + --num_train_images=10000 \ + --num_eval_images=1000 \ + --eval_timeout=10 \ + > predict.log 2>&1 +#ֻ1000ͼƬӡϢpredict.log + +#жϣܼckpt/־ؼ֡ȼlossֵ/accucyؼܼ֡ʱ/ThroughOutputȹؼ +key1="Restoring parameters from" #ܼ +key2="Inference speed =" #ܼ + + + +if [ `grep -c "$key1" "predict.log"` -ne '0' ] && [ `grep -c "$key2" "predict.log"` -ne '0' ];then #ԸҪ߼ + echo "Run testcase success!" +else + echo "Run testcase failed!" +fi \ No newline at end of file diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/run b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/run new file mode 100644 index 000000000..fbbd642aa --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/run @@ -0,0 +1,2 @@ +bash train_testcase.sh +bash online_inference_testcase.sh \ No newline at end of file diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/tf_hub.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/tf_hub.py new file mode 100644 index 000000000..7b7fa094d --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/tf_hub.py @@ -0,0 +1,361 @@ + + + +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +# Copyright 2022 Huawei Technologies Co., Ltd +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +r"""Provide functionalities to export and eval tf_hub module. + +Example use of export_to_hub: + + tf_hub.py --tf_hub_mode='export_to_hub' --cell_name=amoeba_net_a \ + --reduction_size=448 --num_cells=18 --image_size=331 \ + --drop_connect_keep_prob=0.7 \ + --export_path=/tmp/module_export \ + --model_dir=/ADD_PATH_WITH_1001_CLASSES_HERE \ + --alsologtostderr + +Example use of eval_from_hub + tf_hub.py --tf_hub_mode='eval_from_hub' --cell_name=amoeba_net_a \ + --reduction_size=448 --num_cells=18 --image_size=331 \ + --export_path=/tmp/module_export \ + --data_dir=/ADD_DATA_PATH_HERE \ + --model_dir=/ADD_PATH_WITH_1001_CLASSES_HERE \ + --eval_batch_size=40 --alsologtostderr +""" + + + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import faulthandler +faulthandler.enable() + + +from npu_bridge.npu_init import * +from npu_bridge.estimator import npu_ops + +import re + +# Standard Imports +from absl import app +from absl import flags +import tensorflow.compat.v1 as tf +import tensorflow_hub as hub + +import amoeba_net +import amoeba_net_model as model_lib +import model_builder +from tensorflow.contrib import slim + +flags.DEFINE_string('tf_hub_mode', 'export_to_hub', + 'export_to_hub|eval_from_hub') + +flags.DEFINE_string('export_path', None, + 'Directory where export output is stored') + +flags.DEFINE_bool( + 'export_feature_vector', False, + 'If true, network builder only returns feature vector after global_pool ' + 'without the fully connected layer.') + +flags.DEFINE_bool( + 'dryrun_with_untrained_weights', None, + 'FOR TESTING USE ONLY. If set, export_to_hub is done without restoring ' + 'the model\'s trained weights. This helps test the Python code quickly but ' + 'makes the resulting module useless.') + + +FLAGS = flags.FLAGS + + +def _make_module_fn(hparams, num_classes): + """Returns a module_fn for use with hub.create_module_spec().""" + + def _module_fn(is_training): + """A module_fn for use with hub.create_module_spec(). + + Args: + is_training: a boolean, passed to the config.network_fn. + This is meant to control whether batch norm, dropout etc. are built + in training or inference mode for this graph version. + + Raises: + ValueError: if network_fn outputs are not as expected. + """ + # Set up the module input, and attach an ImageModuleInfo about it. + with tf.name_scope('hub_input'): + default_size = (hparams.image_size,) * 2 + image_module_info = hub.ImageModuleInfo() + size_info = image_module_info.default_image_size + size_info.height, size_info.width = default_size + # TODO(b/72731449): Support variable input size. + shape = (None,) + default_size + (3,) + images = tf.placeholder(dtype=tf.float32, shape=shape, name='images') + hub.attach_image_module_info(image_module_info) + # The input is expected to have RGB color values in the range [0,1] + # and gets converted for AmoebaNet to the Inception-style range [-1,+1]. + scaled_images = tf.multiply(images, 2.0) + scaled_images = tf.subtract(scaled_images, 1.0) + + # Build the net. + logits, end_points = model_builder.build_network(scaled_images, num_classes, + is_training, hparams) + + with tf.name_scope('hub_output'): + # Extract the feature_vectors output. + try: + feature_vectors = end_points['global_pool'] + except KeyError: + tf.logging.error('Valid keys of end_points are:', ', '.join(end_points)) + raise + with tf.name_scope('feature_vector'): + if feature_vectors.shape.ndims != 2: + raise ValueError( + 'Wrong rank (expected 2 after squeeze) ' + 'in feature_vectors:', feature_vectors) + # Extract the logits output (if applicable). + if num_classes: + with tf.name_scope('classification'): + if logits.shape.ndims != 2: + raise ValueError('Wrong rank (expected 2) in logits:', logits) + + # Add named signatures. + hub.add_signature('image_feature_vector', dict(images=images), + dict(end_points, default=feature_vectors)) + if num_classes: + hub.add_signature('image_classification', dict(images=images), + dict(end_points, default=logits)) + # Add the default signature. + if num_classes: + hub.add_signature('default', dict(images=images), dict(default=logits)) + else: + hub.add_signature('default', dict(images=images), + dict(default=feature_vectors)) + return _module_fn + + +def export_to_hub(checkpoint_path, export_path, num_classes, hparams): + """Exports the network as a TF-Hub Module. + + If a positive integer num_classes is given, a module for image classification + is exported. If num_classes is 0 or None, a module for feature vector + extraction is exported. In both cases, the default signature returns + a default output that matches the Python slim API net, _ = network_fn(...). + + Args: + checkpoint_path: a string with the file name of the checkpoint from which + the trained weights are copied into the Module. + FOR TESTING USE ONLY, this can be set to empty or None, to skip + restoring weights, which ignores the checkpoint and copies the random + initializer values of the weights instead. + export_path: a string with the directory to pass to hub.Module.export(). + num_classes: an integer with the number of classes for which the given + checkpoint has been trained. If 0 or None, the classification layer + is omitted. + hparams: hyper parameters. + """ + module_fn = _make_module_fn(hparams, num_classes) + tags_and_args = [ + # The default graph is built with batch_norm, dropout etc. in inference + # mode. This graph version is good for inference, not training. + ([], { + 'is_training': False + }), + # A separate 'train' graph builds batch_norm, dropout etc. in training + # mode. + (['train'], { + 'is_training': True + }), + ] + drop_collections = [ + 'moving_vars', tf.GraphKeys.GLOBAL_STEP, + tf.GraphKeys.MOVING_AVERAGE_VARIABLES + ] + spec = hub.create_module_spec(module_fn, tags_and_args, drop_collections) + + with tf.Graph().as_default(): + module = hub.Module(spec) + init_fn = _get_init_fn( + module, + checkpoint_path, + hparams.moving_average_decay > 0, + moving_averages_blacklist_regex='global_step') + config = tf.ConfigProto() + custom_op = config.graph_options.rewrite_options.custom_optimizers.add() + custom_op.name = "NpuOptimizer" + + config.graph_options.rewrite_options.remapping = RewriterConfig.OFF # 必须显式关闭 + config.graph_options.rewrite_options.memory_optimization = RewriterConfig.OFF # 必须显式关闭 + with tf.Session(config=config) as session: + init_fn(session) + module.export(export_path, session=session) + + tf.logging.info('Export to {} succeeded.'.format(export_path)) + + +def _get_init_fn(module, + checkpoint_path, + export_moving_averages=False, + moving_averages_blacklist_regex=None): + """Returns init_fn for the session that calls hub.Module.export().""" + if not checkpoint_path: + tf.logging.warn('DRYRUN: using random weight initializers, no checkpoint') + return lambda session: session.run(tf.global_variables_initializer()) + + # Build `variables_to_restore` as a map from names in the checkpoint to the + # variable in the instantiated module. + if export_moving_averages: + variables_to_restore = {} + num_averaged = num_blacklisted = 0 + for variable_name, variable in module.variable_map.items(): + if (moving_averages_blacklist_regex and + re.match(moving_averages_blacklist_regex, variable_name)): + num_blacklisted += 1 + else: + variable_name += '/ExponentialMovingAverage' + num_averaged += 1 + variables_to_restore[variable_name] = variable + tf.logging.info('Export of moving averages is applied to %d variables ' + 'with %d exempted by matching the blacklist_regex' % + (num_averaged, num_blacklisted)) + else: + variables_to_restore = module.variable_map + tf.logging.info('Export of moving averages is disabled') + + unchecked_init_fn = slim.assign_from_checkpoint_fn(checkpoint_path, + variables_to_restore) + def init_fn(session): + unchecked_init_fn(session) + _check_shapes_of_restored_variables(session, variables_to_restore) + + return init_fn + + +def _check_shapes_of_restored_variables(session, variables_to_restore): + """Raises TypeError if restored variables have unexpected shapes.""" + num_errors = 0 + for variable_name, variable in variables_to_restore.items(): + graph_shape = variable.value().shape + # Values are big, but tf.shape(..) whould echo graph_shape if fully defined. + checkpoint_shape = session.run(variable.value()).shape + if not graph_shape.is_compatible_with(checkpoint_shape): + tf.logging.error('Shape mismatch for variable %s: ' + 'graph expects %s but checkpoint has %s' % + (variable_name, graph_shape, checkpoint_shape)) + num_errors += 1 + if num_errors: + raise TypeError( + 'Shape mismatch for %d variables, see error log for list.' % num_errors) + + +def _make_model_fn(hub_module_spec): + """Returns a model_fn for estimator using hub_module.""" + + def _model_fn(features, labels, mode, params): + """model_fn for estimator.""" + del params + features = tf.transpose(features, [3, 0, 1, 2]) # HWCN to NHWC + hub_module = hub.Module(spec=hub_module_spec, trainable=False) + logits = hub_module(features) + labels_onehot = tf.one_hot(labels, logits.shape[1]) + loss = tf.losses.softmax_cross_entropy(labels_onehot, logits) + + eval_metric_ops = None + + def metric_fn(labels, logits): + """Evaluation metric fn. Performed on CPU, do not reference TPU ops.""" + predictions = tf.argmax(logits, axis=1) + top_1_accuracy = tf.metrics.accuracy(labels, predictions) + in_top_5 = tf.cast(tf.nn.in_top_k(logits, labels, 5), tf.float32) + top_5_accuracy = tf.metrics.mean(in_top_5) + + return { + 'top_1_accuracy': top_1_accuracy, + 'top_5_accuracy': top_5_accuracy, + } + + eval_metric_ops = metric_fn(labels, logits) + return tf.estimator.EstimatorSpec( + mode=mode, loss=loss, train_op=None, eval_metric_ops=eval_metric_ops) + + return _model_fn + + +def eval_from_hub(model_dir, input_fn, eval_steps): + """Eval using hub module.""" + hub_module_spec = hub.load_module_spec(model_dir) + npu_config = NPURunConfig(model_dir=model_dir, save_summary_steps=0) + #run_config = tf.estimator.RunConfig(model_dir=model_dir, save_summary_steps=0) + image_classifier = NPUEstimator( + model_fn=_make_model_fn(hub_module_spec), config=npu_config, params={}) + eval_results = image_classifier.evaluate(input_fn=input_fn, steps=eval_steps) + tf.logging.info('Evaluation results: %s' % eval_results) + + +def main(_): + mode = FLAGS.tf_hub_mode + data_dir = amoeba_net.FLAGS.data_dir + model_dir = amoeba_net.FLAGS.model_dir + hparams = amoeba_net.build_hparams() + hparams.add_hparam('drop_path_burn_in_steps', 0) + hparams.set_hparam('use_tpu', False) + # So far, there is no standardized way of exposing aux heads for + # fine-tuning Hub image modules. Disable aux heads to avoid putting unused + # variables and ops into the module. + hparams.set_hparam('use_aux_head', False) + eval_steps = FLAGS.num_eval_images // FLAGS.eval_batch_size + export_path = FLAGS.export_path or (model_dir + '/export') + + input_pipeline = model_lib.InputPipeline( + is_training=False, data_dir=data_dir, hparams=hparams, eval_from_hub=True) + + if mode == 'eval_from_hub': + eval_from_hub(export_path, input_pipeline.input_fn, eval_steps=eval_steps) + elif mode == 'export_to_hub': + num_classes = (None if FLAGS.export_feature_vector else + input_pipeline.num_classes) + + if FLAGS.dryrun_with_untrained_weights: + checkpoint_path = None + else: + checkpoint_path = tf.train.latest_checkpoint(model_dir) + if not checkpoint_path: + raise IOError('No checkpoint found.') + export_to_hub( + checkpoint_path, export_path, num_classes, hparams) + else: + raise ValueError('Unsupported tf_hub_mode = {}'.format(mode)) + + +if __name__ == '__main__': + tf.logging.set_verbosity(tf.logging.INFO) + app.run(main) + diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh new file mode 100644 index 000000000..622b3a99d --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh @@ -0,0 +1,38 @@ +#! /bin/bash + +DATA_DIR=/home/test_user03/tf_records/ +MODEL_DIR=/home/test_user03/hh + +#߸˶ԤõݼԤѵģ͡ATC-OMģ͵ȣִ֧OBS +#obsutil cp obs://obsxxx/xxx/xxx.ckpt ./model/ -f -r + +#testcase壬߸ݲͬģд +python3 amoeba_net.py \ + --data_dir=${DATA_DIR} \ + --model_dir=${MODEL_DIR} \ + --num_cells=6 \ + --image_size=224 \ + --num_epochs=1 \ + --train_batch_size=64 \ + --eval_batch_size=64 \ + --lr=2.56 \ + --lr_decay_value=0.88 \ + --lr_warmup_epochs=0.35 \ + --mode=train \ + --iterations_per_loop=1251 \ + --num_train_images=10000 \ + --num_eval_images=1000 \ + > train.log 2>&1 +#ѵֻѵ10000ͼƬӡϢtrain.log + +#жϣܼckpt/־ؼ֡ȼlossֵ/accucyؼܼ֡ʱ/ThroughOutputȹؼ +key1="Saving checkpoints" #ܼ +key2="global_step/sec:" #ܼ +key3="loss = " #ȼ + + +if [ `grep -c "$key1" "train.log"` -ne '0' ] && [ `grep -c "$key2" "train.log"` -ne '0' ] && [ `grep -c "$key3" "train.log"` -ne '0' ];then #ԸҪ߼ + echo "Run testcase success!" +else + echo "Run testcase failed!" +fi \ No newline at end of file diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/xx.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/xx.sh new file mode 100644 index 000000000..71381ed8f --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/xx.sh @@ -0,0 +1,21 @@ +#! /bin/bash + +DATA_DIR=/home/test_user03/tf_records/ +MODEL_DIR=/home/test_user03/xx + + +nohup python3 amoeba_net.py \ + --data_dir=${DATA_DIR} \ + --model_dir=${MODEL_DIR} \ + --num_cells=6 \ + --image_size=224 \ + --num_epochs=35 \ + --train_batch_size=64 \ + --eval_batch_size=64 \ + --lr=2.56 \ + --lr_decay_value=0.88 \ + --lr_warmup_epochs=0.35 \ + --mode=train \ + --iterations_per_loop=1251 \ + > xx.log 2>&1 & + -- Gitee From 4d367690fdbe3e07c8b862026e288c55b9e824da Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 06:38:22 +0000 Subject: [PATCH 03/20] =?UTF-8?q?=E4=B8=8A=E4=BC=A0=E6=BA=90=E7=A0=81?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../.idea/amoeba_net.iml | 12 ++ .../inspectionProfiles/profiles_settings.xml | 6 + .../.idea/misc.xml | 4 + .../.idea/modules.xml | 8 ++ .../.idea/workspace.xml | 63 +++++++++ .../amoeba_net_model.cpython-37.pyc | Bin 0 -> 16830 bytes .../inception_preprocessing.cpython-37.pyc | Bin 0 -> 11975 bytes .../__pycache__/model_builder.cpython-37.pyc | Bin 0 -> 10334 bytes .../__pycache__/model_specs.cpython-37.pyc | Bin 0 -> 2103 bytes .../__pycache__/network_utils.cpython-37.pyc | Bin 0 -> 21032 bytes .../common/__init__.py | 14 ++ .../__pycache__/__init__.cpython-37.pyc | Bin 0 -> 111 bytes .../inference_warmup.cpython-37.pyc | Bin 0 -> 2803 bytes .../common/imagenet.py | 74 +++++++++++ .../common/inference_warmup.py | 97 ++++++++++++++ .../common/tpu_profiler_hook.py | 120 ++++++++++++++++++ 16 files changed, 398 insertions(+) create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/amoeba_net.iml create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/inspectionProfiles/profiles_settings.xml create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/misc.xml create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/modules.xml create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/workspace.xml create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/amoeba_net_model.cpython-37.pyc create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/inception_preprocessing.cpython-37.pyc create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/model_builder.cpython-37.pyc create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/model_specs.cpython-37.pyc create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/network_utils.cpython-37.pyc create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/__init__.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/__pycache__/__init__.cpython-37.pyc create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/__pycache__/inference_warmup.cpython-37.pyc create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/imagenet.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/inference_warmup.py create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/tpu_profiler_hook.py diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/amoeba_net.iml b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/amoeba_net.iml new file mode 100644 index 000000000..15fc560d0 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/amoeba_net.iml @@ -0,0 +1,12 @@ + + + + + + + + + + \ No newline at end of file diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/inspectionProfiles/profiles_settings.xml b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/inspectionProfiles/profiles_settings.xml new file mode 100644 index 000000000..105ce2da2 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/inspectionProfiles/profiles_settings.xml @@ -0,0 +1,6 @@ + + + + \ No newline at end of file diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/misc.xml b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/misc.xml new file mode 100644 index 000000000..44fd3cc6e --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/misc.xml @@ -0,0 +1,4 @@ + + + + \ No newline at end of file diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/modules.xml b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/modules.xml new file mode 100644 index 000000000..933dff8b6 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/modules.xml @@ -0,0 +1,8 @@ + + + + + + + + \ No newline at end of file diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/workspace.xml b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/workspace.xml new file mode 100644 index 000000000..cca82cf07 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/.idea/workspace.xml @@ -0,0 +1,63 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + 1633974512150 + + + + \ No newline at end of file diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/amoeba_net_model.cpython-37.pyc b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/amoeba_net_model.cpython-37.pyc new file mode 100644 index 0000000000000000000000000000000000000000..7b4b828e812518b4c163eb609e5908216b5929dd GIT binary patch literal 16830 zcmb7rd#oJSdEd+Y_}3`K{C{@>{J<%I{QdN`9w1GZy;k%+}`8mh-iFd9tszPk#5;_T$%S zALuO94oaO8@`q{_f>DpE82|!GZUU;6V7~om}lH+%E(Naeq+mpT_;6;33>U6wb94ThE}*$5D1T zco=05kIIe)`K!j_BZH5f>GZ-?eC^DZxQtb9Oyk0*-wzQ`FPI$AYuHz;b z_c!q4#L91Wo83n9D7Lft8Otyl|L))N^FK~C_-4cSU#IFiqthk)oW&J^FpSt}nOJ2j z$OU$g4+=psaDo!9GIA3^C8!3IT+LvHE6G*l`f<&;PFxqP@gw-UxE>82!}T%Ld>p?| z1W(>QidA}Q@U<&pyrH_08-#2AMtjFvWrKJ7KBg8emCLSsZ9R;_tZL-?DsUBalTJLSN19vqaW3Kqe`C&QUjB9D?&S>Q-YnMX9%$9L=Q4*CHWH`F?e2zgVy~Qi^{sTJ z-#&e<@%R2NKfj$VKDI-b{PL;BntcA?RASvYuF#*6Hv7h@#-D!eyT5s9_427k?R&qR z{Ix&(&Z(xmGJ5{_>BcV|`LS<*`j@_XDwzmx^?Tj08#n!Sqx@Sx*Z<4q@0@BpC!gHu zv_t1~iG3=`MUCKoinzzu-+P*4IP&w~dUfe9zH@5j!s5}SEc5cB<{(U_ zyBi%(RvZ+I*+$6}OTrs|J1dxBLEB#q+g`oxM-l3jSX2+&?I@X4VXz@wpgWw4;;`fG zF09S7{|UMB)S_C1h9$B&a~J-$^xuCMEtlvzFO_v-6W zQrUM^t%rV)OfVPK{WiFzR&5{mg0Sx2_LLuo$&{W%)amu&bv#e@2ddZi>b-6^ ztjFHgz#^=HDfRksvjgH# zwW{hwSdqSdQrnl^cEE$(n0tzKtxaXc*&M68p6eazhFC1syOB%@ZH+gxCM4K|1dKC^ zoA2~)aN0hW9yFQ0ed?xaLa;^~osO@XphBU7Xx&#qgsp5wv1+bD^afr-^)NZ1?;@## zQq)SOvC*ue$({ROeD$)3q^}w3N5CbLeE<0JI%c;VW4s<3t7o5E9w~rh=azj@1|Fn> zCrV(ce><7Z1ckS*W!#JCEvg|g3bt*T@-0f5vvT;&S-jIZK6P*vaPg~a=)A6z_O<)BWwz_0k)4n ztNO+3R2C2X=tabqy4`R?`EC5f)WDw<&A}JA9DwMyW*0gIZ;RLc_`~`GVQjQRx874K zj3CKHHBpU$e(TVZ?5P>mdmwLgwtCz3yC2q7c!O{P;>TCFGhG;jQ5|$Fv7zV zPA)I|>Q?i{QcpFOflQWs(sm`I!NI(>D!%GCjk0TQsLu1R_Hnyy7_l>w1J6k2)F%_%QcF5CPa1aA$9uoB1KZcF}(hf&#D)YCn#I`_iQY%Yl=9Y^om!MN-;# z8(LPhqn_LD^{z)kr)a7b0Xy(xcO7cP?e=2LkM+2H8@Ms-y4X#y5PBt6OkV?f_Q9?K z)AjqWhqr-GBlodYKX7^T*iwpKvjVr{vq!fbu7Q6xgonWBV1u5D(hloQwb8i&o#_(p zxb^k0etk6jk*G@taS~&2GZt);X}b0Bfef3jf$JuwH`-jlS&41R~ zG&ijPl748dS^JH@WxZ=-$<0CGiRh&JL?qg8F_&0N&nD*e#Jn-qi9&hbFTd1=j;;4P z{g;1_T@xc0jKLEcrBUyu<6nBI-2(!PUPdNsF{r$ft{>m^ee-BymKP@gT!F3&0$8HN z=JFjgk{>=S#Xb2;4CfLS#adD093CcD>EwP)tZ%zab*Iu~Fbi7`2C+TgocbJBF| z!`8f6Hr3lGuk&H{;ZmK(6}^fiG-@VHaw{-vIT~d!S#9zKe#JP$Z-E9Hev2^CCh+Uf zNW*W5h8lj$!8EpRCOM#GGL+`|2$+o&cZrtiVBuW228)q`m?C{tZK+P`;D^5))>&N9 zG!oE-;enBnVtAyCl|@TdGAU`X7&H~=!;ZrCnv|=*ie`J)3fspQd)|Lz$3!6pM9^S7 zgxw=u%BqjZmmr1!Mh12QnHZ9Z<}2iiwUyg8C{4CGSUa{_a8(4kts*29q=-lq4&7cE zb;mhU)Co5nK}NV?KhO}m1yX^)W2f7u<_FCMc4Bq=i4Eu*Bqd6Epv*qxxF30>XspWx z^pxao_-fDm)kQq`LtGhqTo|qW-id=n>VnV(15v1X8W=s0e~tjluZBt#G`#^L1aVBd8%?YPpa4#ADOhk0nI zpm>*HnZnW;%xe_ehBtu@8J7molGyob24lRin6FL5y%=yv+lCX`{*)Vd$;234^Xp(# zn5uvi+M2mttE3j0RBRkOd-&V`<%L71l4|xMv9#a{vI!^_CpG`=#~andOrBwK3`qjp z3jzSPB)~Q-xH%0Mc6Fl`S)06|jxzZ;+vT8ORlr=a11iH+z7}!+H$fu2;h}Ux&Y-9sUtoXyB650@MM`5VBYhpk?`;wrq3W9DL$K zY3Kn{8}(lo$%B_vaTRe187)D$40H5{wX9a|x`l1c?rhxUV{Ut)RlGwoE3}-4jLkw^ z8j_p}tul#ZlSEb6g5Psk6#J;MRb>evj5D;O_l6FhL8y!E!#qf7l!vI2Hb!f5Ybq#i zO%HALlY!HJy)bRgCMj9_!kZq0#co!0!&9y;4LGb5B4#2xtzMJ2iK%akv9u`pMU{sZECA~YyJItN$p1`{+T3`aE9iq2)nE#um zdNQaCOKa9uV>qFI7psE{n&kb=D$=M*rS86VF9JRSW`YoD0<*2DUdLSveQoB*1GV4*0fb!~L^9ImahW;RuAmmP#zp$xFXq(t{ zB0vm3%8srh8O)94<60Z8R0_qF#fLOWmtf3m^_VzRkbi}sAb;)3nTsn)@$!}P=PsVT zmgLdSkN0v`k8coOIgMOm*LyJdodCWf=F}P=(<7(WnNahqE|OY`(^E@QEubtZ!l~!2 z_u|^DCf%sF7K7P5uAxk&c98Ij?qivQX0uyFo0hb61h*WgfdxL6^RDRZaO zD}t}0SJ1s-O#%p6fJ4*fq*Vb=SNMf4u))<8aP%y$N!*FA1HVoV_oz8*I;iDX86Dh* zX_j`sa+jWb3(yQW{J0AsJcJ7*E|NX9L;S7{+3!FX z*muBM-G7|+QbI3(JME<`%*kF#1WTenChns4NA&giw66;K`u()8>Zq@Yp&3`=YCIWC zVhkVE*Q~Uk3Z`_QB3d|-X)v!H&uq=owqh^PR?~xN&AKCp-jbTdN)6`Th^V`+e(Ib% zdIWV(Dc;T@;m0bAGTBf6uk>vz4L z0xVoh%Fuv{x~$ntoXkr&SkU9S)a#6fBkn~P94@NmuVWBNF?BNz4(RGvE?vD0-{g)d zSV~t+{g8v%H@ermy_F;+$siW%MYrVa&lvUXW3;@Tt`QdJcXTjN|pJaq{T- zv0wg;K3vq{YgmV?07T&>zY9;LO7^W>eq&TPZaIUw$&^M#>RHZdI-SI*IH|Ja%IlZC zbLY>Vd0SWD?WNZ*onN{3#-(H`t8o7E)r+saz5)(~ue_~ZV1E-MA7_6fnLDdEJUU26 z{ijM@FDnt8uV;Mc-Q} zrD7-ZXC-9)QN7%(MTI|j`Xki*=#?YQPNYdKFN}|a6#6g$@Ie0XSW?471la0y43_?? z0Uer$E{6+ZNF9%-+m@6S@jO2)fa&26g?IoW2&}tgpWNUn%|1dFIG)qGZ8VTn>qXI0 z${#!dRPSOSi^XIn6XcmQKhkvfG#)3pL9?$F(;Qv=d)@m}uyxz;_@`tT0ow}_7n*$^ z9mq?PNmhtFs8|mOKT6DodX*2)GuZ>irqYeern1I;n?ywetIvJ0A+XEV;Hi(;g{*y@ z=$xN}xOfaru@U__di~-UQW0RmN0h=X0<6O5E|0t7Ft5%5(oiO+$CC{>oj1*|7_jJw z+H>1EY?8BC8Wz=8fXPcubF=&vqvf

Q-EC5xbXz0=G-vQW@U|#mxy^6@fKBi>KAC z$!%k^iuX<7njRw8n%M?aYXEj_PU4-jLrCS#DdgscQ^-xD-aKmUYwgF~4DJpLXSU4` z_L~OAurQn+I>Tzz5%C?tjK=UJTNSv)^G6O_kCk?VYq;`E=i63IuCX@Aib4P z@4sWEU0G{qSOGBHA57haw+?eNTU>FL9tfuI(n6Z?e-uQJ@(uiIVvV?~IFiLlY$ax+sN_R4^tOwqlbkDfjdjwY{h6ojx zRu!J%jh?1{X=b^f**P6}YS!T%RPr!{S4$t3yY4PWJr;?lC?iO_G;%Cl#o31*6VJ3Y z!d7$4QH!fPNT5-_jGts81%+@BMF>OCV#`YMVxcAzJUHAC4Tlrn-KTLQ;ANp7vOZrB zwUB@y=~a<4CJ+Z~olO!6WG+Yv7SdL6+xXxlNgQPk@{j9$s}$&V;?{xm&82n9+i=@x zwgTCwIq^;9Rz!o*Af`dOIHAR?Zsz za-yijj|K03go8$jT}Tf2?VJAX2o8Rn;bcZSj~oj;TJ~}J3R1{ycM06E5~~r3VZF<5 zB4D|)-osfUsmELE{sJX0m(75$a2}F(!5qBsVNzG942{Mfp`Bfgc7paTFY$3*Mux1Z zg{r_)15np!!FdK|h28_uHMVfI;5>tv&A}?=bJ(-6fLYHO<{@;b3k!k$Gt`Co!I|^) z(#p|}ixX$nth*SOdpT5e7isQ3d-;u{ZqO5Fy9#lZ9l~+#Uc3AT02mBF5(;+=?uFEd z^`t14+;b;j5P<|1i|R|LBjm6CA#$}@aA(-)VGi)qXqH$1oDZ@tKB4{vTkKJyCH->j z$FFb`=`=NRv}Gn~y_gl8=MJ9x;kRs$>bKR+2`Lp^P0NNYzacCi=#=LZ~r)-({>XCss0WhPh-z^ zNtL9!>sEc0wGPrV3Sggd)J`{15u%eWQ@uOn!RJn>h%fp3Y&pr6^tdCc-?)74%(e60>zA)4^MYo3ZSWerb2RVKuSHM>FcT+A8*x}>b-ETu@23p;)?KN=^N#pKtS{7WXJ(oE=xyn$gw7vNg=U6Ls!vr)A`r1*+UjV8E*@1(`lHf;TSqJcd!gA&B#yis!w%;kVN#MVtfm>Yk@Q$4S!hp$SOnl~kk&;kTI^Gpgjc3aW#b50WNERKyYP^AOZ3uA1zk}0ukzr9t)sp z^869%OCMI^pb^mr*DhQU)x_R#?Bnu`Oim|xRLKwsiHHNhkw^tZ+5s8ybQ**Umd9yj zyLS_zbTkdotzW01=pKLe#PcroVsupE9RfrIyt@Ij6=7)Y+d4|X)q^|sEI-G3Ysa2t zs}xDexaHCDy^ofbGO(0ZLbK<_mCykvj-_Z~s64gGrk+k90>fsLMUc;E=y-hm#EW{= zqxDM{B@Gafo)l)%1TJu~np$2-;02K(gDLv4>&Mv)v)BpvGuJm(mte^u8i=40YUSoK zf{B(_5sbCm0oDlBa@^~+qvdpgqvcU35O}vCktmFJ9K+fj&bgEBJF95?j*!I7b%b$Y zYd9jI6^?W#t)Q0+BeRRz6Lh9+-^Y-lLn=GHgO)cu3m$62gYCS5HdFtJ$tz6$Gm|mr3Dc2R|0ku6ARLhhA{F-U z;*s&8ALqai(aQv`v$&!HIsguVQAjjGZ~9#^BGY?2yKk^ojt)HiT<%$4-wVRQ;SGEF zVe>iG51gMV!MiSosobY=ZD3J7Z7+f$p)T zj5IIQu!sGVDKC|SsiOHFMFAb1hCOZRFWi9goy`GC0Dw%(iGU@dfHF15642`zCbOwS{duy6b z9LZ&MX3z`V^NBXTLEbNGKJ!Vj;xl5;Kf|(D&QcKsSHS+(*OzaEe;V7F{t_z z^qv&c$?PTd>RQNf^z`YAP-GE_k)aRM=HS?F@@-0V^G-W{SL=JH(eN51bftc~4w-=W zHyHC2;UKeryUF?96LTb|{ts#<5>C(Uf&qo2e(pJZ6O$63u3$pI0Trqdk5I>syaQY@ z`e}w(H)__vrpt9`inhtG?C=&*n93O~Tt<|1>R%(l7z@Lzr# zFa6)BZw#LPsFlc?V4r5`b{fr#kv-83y|*#U-~;tewWGbF(F8kSt$H1 zYv+(;aDwB0A60&b!kc90%-B}B=p0ilt>YSf$c+5R$afk|;(rEtz3-oX?OJ`n1@aEuori{nLF?`^n6gXH%XExQ2SZaDwM;7w#cs34)S9AnPjM(*| zamQkW7vtLn4Kun;qhUn1<)O+qKmlr=%fMBEIjYT6!eY7@AT4BLCDKg_@DD9d{|^SH z>3G4|y1_lvLV%|8V1CE!#*i29k`D(;7=IR5M0%zWqw^GwQ;0MiX4z0SRtz;IPpGI6 z&_O_oNtCh>y|QLbI%L$pXTnkL@gcw&gMg!uhf?|<5TpN?B?lqtn~km~SYN_Y<@6uU zgJh;n3l@GF2CGV`+K<6w{@frBawc1y|H$_)a1r;AcYZZQU|E~%LP9b?!*n;T&$`7!Tupo&t$X1U2i%Q-p2&vXNoa}=d#%Wylu^^>8yc9nmv{bwoGgtLQ zui-$q+Ym;4PBS87r_)nR@hJchjLvb=)p6AZ{s{pIx-`jPrHhpDmYz<>ai#F} ztcEZ$LJvJe$%LmDlQ)?wYd0wgdk#jM0iZZIbl=FwOvY8=hrirCp{99ch~#ieAyR=9 zm9bGyeT*yrX;evU&InqMoR7nVjV}KpMpI`m#gxZ$_!5UEZ#4&p?;qH>;x0n}A%})@ z5p2|dg5t8e*%L;ZOHEr;0XY~=!WK>Gg9X*Bw@%o2&4Bmp?(ksB<1>;R< zvUd@~k6B7|wh#sh7u0$I?Gm%Oi+w0azF&HYe`3AWS1*5%kpe{UiTg<*3t5o=&5n

CiUuUy+T4WwSNmawOGV)BG_m3E=_*A;c+;;1tVRfV9^b%*id_pl(DQV=+m;poHql za;^Lz$^R7jcag~nCS%DaIg$2{;lWLgg^n}vR_C3ZlT$3m@Lxjd;xU15beXShFuW1~ z1T73jMpfk6q$0US_k|UCjiDDB2|o zfBeV$nB(D-1bC{ygj{XfyK?^O#oBrA>cw;Cy-OEY_LiLajQUA7`6(tp&E%6z&M|o# zNm9@zsJQzDzUq^!$&~o7Y;>vXGyvpMKWJ}Y)#xLNm#5IBggM$d#>`Z5AQ)T~IY)6M foz;0hH7mAl+6ZrC>U7GFI!C6zWZc44o}c@F$7_n{ literal 0 HcmV?d00001 diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/inception_preprocessing.cpython-37.pyc b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/inception_preprocessing.cpython-37.pyc new file mode 100644 index 0000000000000000000000000000000000000000..36531ab7aee0b5d38681c8bee572dc98fdb095a7 GIT binary patch literal 11975 zcmeHNU2GiJb)KL7=kiA+CI35PH?g%2MM`Gi1d3u>wk*d%Lpw5)3TrE?*}b#Nk!EMs zcV;P4SOP|6z{-PQ7bsA)C<^i|28seL`p`ZU>1$t#KBdLHG(dr-DvH8D(IO4{(C?gk zXa7XnMhXNe+97s$=gz%%&iy&(`_8%d<=NSaf$MLc_*2)tXc+%M57|{f;wAjzf5lB0 zbwilK3e2umw@gagNZWOr(>bJb^&F@3LB5{%@&|=OgP#jQaf`|qgHpF#FY7Xu?u>=^ z-D-W7b4tNncfLN)=`zv_^##-L$_FPzMa+omA-=;;vv``*Pm5w+EQk~8>BdrhNi2$` z_l^3p_>x!_C*L>gC&e@3lz0fKQ{q|in0Oqihs0^|gm@CEheb^|;whva5zmQV6JJE? zQIUVksD1hND{snZ#}{7g^b)3bdqh)c|B}-y~K|~ zC-jnAk-QnN-p9ZhH8Y)en{gEM6R+WSdy!0#RTBP=AEPwVb3N&YNu%8lTU37&S>|)N z<=cVViIG@nMPVYNpwSMZTaBnkc@^n~BI?ppy3}emgdZnp8uc4^)A4@#d%sK9pKsiL z=#{23dh768=gVtnP)$Bi4W;GLrBC%^uhDknq|u6kXjpUWcFXPcf<1>`Is9s^wJoEr zJKl`ncD7{STXDP~_8hLzlzwMB2{Dpuv77XzOR}$R_q{dlEwwh1x@FT);-dd~{Xc#I ziNv^Jc8r0A|CTWC**`E7`yls$@qzh))wYEt?DtDaeqh}(FBx~tLsR4)F)+b-_o`_a zuSDG*rUjFTu|K`L;>2E1TXAmr$u=dO*zNWLPdIJ#$9eng>gsCP?>1U)?8Rr!U0bbG z9Ot6!#OLwfadyu;eu(CrEia5CxyqT;@OkIWy<{5;Bb#AbNIV9r%&^U2K`e*$=T{uO zXnN9#+75q`sws(BI1-9QxZ+{>LM_D>mZJ3HevnYTseWyH(&Ic7X?gw*t!b3saf7}` z#qq#Wqwih|%gJ$L$8~y2dI;rpJh|dD(MK8`KOPQ{*bCjR-*SUskNQ-z(n8nWMZ0NU zBzrxtg~8^_!ktXc5`OV>xMA*$YftZ+cQKO(=3OhX4(z+;z&x}LEnz%j49q*$Z&~l= zFlXj{YNeKwMf^((F{Y`tomxpd&3o_m-2h*;R?nt68myX~=BV?j4M|gT_haMd-(K5} zy51V5J857o%d_iiNw0rqeXSb_EX%d1-S%6)8?3qA$ZNU{sPUSgDRpC{_;o)yg7t30s{T0*3Vq@vkqu3iWt3{m_d5?NKrT`I^E)TKFHp&$CN zD%~{C>p*ok-}Pk2OY^M0HCsM}3i2e?Di7C>d>GGoN7?ce-MG{vRO->=<@ze#$0gj1 z>Y|l1=S=A!-O5o8F8W`_?-l&wzlOXtQ3G3;1EXUe+8>!8{)q_%ls5*zhntqXM5=hh z9vH$FxqH?=si<`$hv&Q~h~hna-w2<6FmJlnM`n0gxAP^$&^kFjfvatq z!a6oKuez;mW#-4sGns-g3eP|ry8Q%}dU`FEt3F(=K0fydd z@FNfUIQC3*8AEqQ%L)YLL2-bAY`=83dD47qT2^8X=47=%&lM3O%jgr^~uI&t)i@XWi(Q-@^C?B+3{|3%zYO6Ad8{8=TN+iB78(M6y!1ddPez`$W< z!13XaO(OwNbPQr?A}8|qa{Iu4%1LrJ@-uwb`Svis!1)E8pMxPR60DHv2ejlxd5~KM z96e<07er-XFB@WpavoNBFo~4M?@NQcn9UQ%D0~&f@`8cqIesn~|8U zc^-?-*Z=;Fd%qL@$42M3m#-Rc6@L7q&QpK%`+u>X{^LgH*>j~oe7yc&8*CL=scV*8 zN4sf3jpKuQTbp6EmCuq0#D?T!R6c+C*Dt>ImZWtoXE=j}NDHu+fD*Y%FV=9Ym09=X zIi%~QCLAQPJ@qntv$jiQ8cC^>X{ojCh9OvIy&^r)0QHQ$q+abv&kM7snInQWFrAOeOHag$s>)$t!r|jUz znWg+>cz)^?=~S8$J{3Nk3V-(|!U+G6)DRuvT_FnE(gj!oi7m6NC>k@Yk44 zqAT6y0vUHedzoBFQ0H997%H5%&#{Y0;(GB4J_4+a2^#6T#l9YM7JlJ}7(wD`%s6VA zo{?y%#V@I4Faq7WZUzqE+k1eS6=#?4dq5gWySvvw+S=rzs_#?NXNjbR4t1P49K0;1 zZMU|gKslC9#w1BZG}wrmH$cGv;TRuJfUbn#C^XRE9M^duuCF?mytWI6mIE-?&#sDJX6VF|ZgX3P2$SiAUy^9`&Wfhfe9v+Yrw zgkus|GAupMzm^Lv|BvKRW2L2<9I=T}qG8R3M2|1|xDe?Gsfct0!kydL2Doqxsd7O*7-SF;k>waU#LA6L=IC+wd%$>*Jza*WGNp ze$YpdWCdXm3Z3wpfrk57|H?LS!Cwd!FFHvdF(YE^<8*k!a4FXamPHW*NKVR+Z>En8 zvnH%jg9c-(R!HYETcDIKU3^f7(!~eYYFeQcsMbkZ9UqePNeb(6ryG8|5k?c=_2SK% zicKh3ByyQP@u`Lr1asT%Ay5&+)$(dp2E{Z-9h4VQUA{~=0uXtLZUmQ-0w`&P$~2mo z=rApFnlQg!&7h95>a!CNFa?Wp1MU0)elZ0?j4v$D6hLFG;a|zD*obaWsTRjOkeI`d zz}?0#em{$DFr?i@IBH;bavbAmBF2Gm;?0tLnE(!il_+bjLj=@DWCO8{j?K{x_~17< z_ogkc<8A&1A_H8iM6_{Gly@ojpa}RYfjpKE@fDCqib|9Z@*u6HLqymU5K9J5e5GxP znR~f+ECS6z`CtZ6YKdycJ~Sa!QU2Wn?N!j;jBamsFhlL#HRMl3KD0z(znWABWe$)c z8X)H1FAb_<0pz(nC?Y!Yu&#H4;5q!_T?5ZA z>Go&2eWPt69#5^$jg1n|mWa<>FE!-$JA~_0gN%xB6BjxNBIbxb2Q1i@aDxB} ziVP5Zcio!`)%3?<|8`vIyCI+jVO_=v3T;e45eFhf#7?Z&Zd25OJ{h=8FHn>M0Sg6t zU@b!cZ9!`ZwTVzP;(oV_4}k&TJ5Y*V;O;5GeY7ai6MPjShcN{YnXLfQ6^K!lq2c`8 z9EK5|(5hD%4jkYIq^2beKW&&71UU9B9xQ2;NXw-Xq*X+wdls^o|*{& z0g1{@qepaU+Vbs-SFa&(!^{S0$_`vaMFG9ND8~2!NNIn@i;1o^5oZBCp$Iw#Bx@;5 zPJeCf+zV^#&)0ayRb1sXypJ(zsQH?=SQ!O!5;{Va#W^C%3--G?$Il{>V`33DPl)A5k2M_WF3QXc-AWhk0NP0A+XKZDNne z22_}v*zut46U8)UdCD}Ui6H=OfnYbv!w#IFeJQ21X`U5z6aq_27Xr826z+>pqEr8d zO2+MHM$z5|2D8C~IlhadDqt|afa3L&88=ZP=oy~q4viyshdk*7;i3UHzV7I5m>eF- zL@s5KPc(lPrAMO5@8EF_Qpzhx)K9kIz6E1`EDv$*#3-DBa04-8KvY`PCR|M)^SPxX ztX;Y!q(TD&-&@pgMq1AM9*r$x0w_n1zlWx!^GnV)nZRC-ElNIUR*jlgTebs{%L z(q5CK&T|x^Hm71HZ&1dYbbFm{yk@x1#17M?BebPHH?in;>+_?<#Bb_zqYP$_506zC zdj(FpKu)Y)HoVbVKS4RVklOC$(Z@n#sYc@R*4pw>%j9n%dAx<2@${o0p>qhv%pnYe zz>K|M7R)&-qooUY&LL+3r3$4glhvwOvWU9Q*;N~))tcknN|tNNYiReU)VN{%%SMZo zgf99&47d^x2FEf7ITfB6Sm2BHhwp(gg1)3ffz}*_I=Kw^D+PEu7V%%&*+U81XUbt`ki($}kK(;3YOXp8kCvdiW%oa! z<8K18De^GM@;r*ujY2L6HE2heoG0ZQQ>BsfHzZ9_AKSXv)q!Wqff%-YOu)#WaGHBS zH{J{-x}m4Ni6b7N=VH?yFY&Dw2!w=E4dVo$8b^GJ6rRb7E(xUoP*O zp2RV3aG7FoE7Q}Z7Yw*EO{^jVpBK|74q=ezKMm1jB(P#NLQ~`Q3N3O)GBuzg5WX-i zHpj3p?z{ZKsDh&GY<2>(hPDqX@J8bur}~dVRa(3E-P&imP~_(v=RD~)M$z>n zNX1bulrq+hDNhQVolQ6R23%$6jC!v$QdgODv>vFzuT=#iV_ z;>U_z?L!WaTNp0^Qi!m!p-F?tYXw+roJ!+SZrag6IwlvlEyPR=GYl4x4^oq;N?~P$(wi&?vJv@ z7WS$)(I46cl^5vt4Z6LE+a^w5Q1;g57JN9`9Yyf`^46xxyj*)MEjAk1Ol&mLN&}~T zvGd?H8j`$RNzrFX&S6^RtuA1S2t1h!0@Yeh z_{=Fd$E{U5=Hw=;JLhPRnv$utb50HW$+2SUROc#A1n~hw-ttbXmtMX6`sS;2&=~dR z^plRHDCf-~QJoRCV=hre`#xm*w(PZX=H^o3M8k!dh9fY#Hs?Nb-Z;M{RXybx*gd ztJ+!Bqfxuscvs%Fy%AbDcpNYWdRQQ0Iq*1e@CHZD95@g`MX4iV`@ihx%4Ji*r~Ws?=D&SOQU0D1(~pkGm+*u?LBN!T!c?Yt zYF}$;D#bI1XBrt1&mx{}WJNsZh#Nv zO4_p8sG>yomio($n#eN{Uump}c)?rkA7~sv{9#sPrFYfF8avLaYzeV-w#;hpDvg8e z5w^lsk#mS0U~9-Z%+}dK{2pP4*kSx0Wp!q;`%vQ^_9(lbJ%HFT_8@x*W%sh&6{Y?& zqnFMFei-p#D+(>g>bSnkyDh8RcRH@s518v&O(%4j75G-P34 z(%DIlnX$2plMv@8dBms;eZ(3o#;Ryn5GA5jq)c?QGI~gj&Di8W!u*OmR&1VEzIJe2 zii?vx>X>l}wTinM_4ei;ysjXHGFD>cTN-N7NF^#0`FU0m`Q(%A)9u;VY!uE8 zcP<5;ZWJb_Gu*K~5$9NY+U=~f(+xWx{W^XA`b=WT8ev%168*Z{?QEf0PHf_6ZP-nFE-9zz4-E`ed*;ZSDwF;WLr)c zC3@Qnoam`1lbpkt?bjq0WV?!8zsmY&0nlh!^E11m>V{#?ZHz|+6x)RG|rdM@&^ik2tWU^@%UEI zcOQ@3FtSP3pLpu=v;DwrI-73v*!jnWcCeenuE$)yF}RTwWkT)~ItZUdpk(n;4b2o! zR@LyW;4##anpG>BnORcJOa<|*YN`iOW@;5RuhtN&X#6;O_*5RQPtIIz<)}w|NE;Y< zgcf)h5lDa{}1OawLhdzwDj>ng&{6PfCfoX3oPP^;+ z5f278#G9aRi|WmLtW%P@m4ni{i^b>kQb|p$u08FLrq@7zSVN%XuzX@UX-VnJA46`7 z#F9Q4JXt(qiJnA+mWb5bt}<4c3e}~ZmU&QzliV&A3X7!m^09J68Edg7mJ8Ar_->L9 zh7q&}7EccbB#dEQ7q6Tc10J-yp39TfS7hvEQQwEj#RvZ~*ZZNe-ohP%e^KF0oRFT2j5)qSVG{w!7o> z2cFwQqmZnIqLpw4^Eg`u9{xq-v|=4ULL=Ctf2>3 zl~q`kBg!r``Kp5UZL~Mg{yQgLu-pn|M-ij1Onjc*EEsx`zwVWGvbYtLJr>uv<(XmU5 z)3L5QygwY+)NhD#)jkPQS^VXMY2&{uPfh zdPs`%bR}%D5~sZY<|TY~V_N@zH@3DoD1Q#a(0H7rJ}K`Rsa{PgXJKlguLA#h&I2wb zI2UoJ>nF<#8zPr0$zUH6%^xJ%pv#X?w@0bl+Mw&UTv~rvLAfQI)`!zZz}5vm;|D0c zhM-<-m|ee1+iZKhQQor`NwMPvO_b2WCndQ_ZKo9lJSobQ*pfzNy5qJx2wGc7Q7Q^i zDAZ?CTHHctjbP|AKnEAbbKCKTuGAQHji0BPouxo1j+MLY$u>$uQkjZLW~wC25ykTA z!$Nr=4n<;WHMOQKWtOy6l--XU7_3*(YG$TLrO*+j-3T{vjsR41kwS{(rp8y}Opnq~ zBoi2@2@KY*DnK7;;cTS$45$_DL-lqE5RXI}!aB|)mS?DW`zOdXdqt=kd^;L}SXKax z%_A2OGAaYOLP^|I;X)KqPT-PC^8_t0FjGK z|F(qFC8U;OQxHd({$wch%!2IFmHlo(I;=!-?uLfXD$b8oIjk4vP8DA!uY_@p^TYmnossa83V-B)j?I z>*wxh$h$*&rAxL4mA<-A>13lswIv|5pWLh;O!O@$bfSp!*HBM_j=AXN@#1|aN-Enp z`A*YwrSp`OeK)!u@N4!k>Jo}<4o*IuzE5Ic(on+D4X&d{0Q~I^P-oyh$w&?R+xX1@ zgHG+Rj}f2}{3Od=ZwO$&598IfA{IvFig03_b4#=?ovp*=2K; zQZFPK%u(Q?r05U(cFPlfCEiy$oMQ3hOVbYUJ}pya29a%o&U2kCMZ22fQCdSf9N$M9 zDb`vRc$6^aDljME(3-kLcvFfs#JGgCNogvFnDZgpwrHQ|Lo=X9i1l9~601FBoF%)< z)Cd@qX}7f7&p|w4?a5aJKFB}-vn)rx=B+%0KGV}-)k$Cps{#v53e1FGn1fd+lmHHW zQ674-^|17LR*EwiskA}>P2YJVDpTsF3eWPTu>t#>V`cOJ50!jd67fSyY_KYg2amEs z>IJqQ0{%s`Psd!M7GhNCi@x(ZJl{CqCiu5Js$YrR{#>X8lywP$|pGx=UkZd?I$1(sdS9i;js`05s^jcTjF$) z9B$6!byRA(o-i1F`>pYv|N88+&+0kRWl{kSg@{gtSqzq#1MY6iqln?CDUR3ZG zpd8C`B8$5WYzSw@`$Pe7-3kdmaOc3LjfR?rCC?9G~}F`$#s&r zJ>(-~@tA^rE=C0;&BBuZIjV#w5GWav4{|XKJjXN&mxP;Ph>%b;d^$W!8vi!xw@8)I zhoodC9kc*KItcIHT{^VSNk>LVM<$hy%z|`edUNSeq;$ZT!k~6Eabk;`oQ(ILV@Nuf zU#H-%hmh~c26QS%t023&t$l~DUgBeF{(sx*dX~RKt!Xd$O$umx_HP4x=h?_VMtMk% zxFU4aY!6ia9prub*tw(s_;*j6XFhHG`zOhF|MQ>DeEQQL{L5c{x%tW&{&kcKo718r zNFQ=+$oUbf$^-%-cNoMY$TB!C5OkLL*nl!nCaORef&wE&bxR!=qD++Shjq%w}i3E(=kz_f~SgS3D0uJfHEzicAm(!E9nt{s?=xr;C~cjI@DX)*~XbY@3x`L zB%_&nqZW8apaA%+0!k3?IU_!-i>bzw=wib(E$b}rgs1RpS(5OQ@cEQ=6%A#MCBQ5Z zUjQhg(-}oYE)k6&bnAxbMY_CXYNtJ&C-+6P6J-{ap~W^ecDCVKcihi4g)7tb0{@hC zrB7f;IMo)w2^2TxAs~X^B0dm31YHotplAg>*SulD*$%{1EHJy^3E_f^G36lL&=w{D zqo1;_Z@HL0x4NlapeX@=`9UNrTh^@I z?S2nM_@PPGYdoNe6;E)Dv+<5jkKWW!FLuF+)sqJFO^T5fO_oBJP#^(X2}z}`CEJqd z4D1@LT%$k{d*DP{oWMe25}_xNY-3gWlz1~Xx8F%Mbv~z%w;>p*)|LZ&>S0pM-Dy+> z*PCWB+)5PTX$n?5^^uqM3o8~zFjhi*)7=!yA{K?;rnmYKL1THLUJ7!5g$jR_X0L^< z{U%}=C`F+7*+HZ0DoL^))U+uJNWVn4f^-!In~XI1uqBvt4S+X~lKB^2tjc%M&kWp} zNX*Rl50QXBNDo~$WO~{<0bTg3i0AO+@sP)8;3*Il)-!i?g1568MdSgv_R4!|P`iro zHLHNGxGcM-g%R?W!YEn>K~+P11rPbi2LNp|y)^(=1Hf8`U0RO|H`Uu~uowpcuVKYt z)iSVag5*s>a~2vt=eW@Df`5V==r3Mt3B^4vNi`A4YUErK*I9(W(1ZaFlAPnRf-XQr z&Xi^k_bGn~+D=)H1#c&W4kQ&sO64nZrzw5{ZYtf0oPgS6(bfqoJ#OLG=87!!eNK@o zOf^}WW%;g~>hGp&0ok%HIZ_rhHqBj{5bI*w+61+BB30mIm(ow*dpFV-k_NBsI(R$R z+J-}h3D4fla@s)BRw^JQMRIxzis=7A^QXSwsoAZg@a|!-patZerFiWrfKA_@Eb<=Z z>{uRDJ~aZvC``|BxtqsNSjXoM_HpZ~*fzPXmmTn1sdj&j2qeLnow9~tEn!i>DZyl> z6Qs2funVO4Z$e`DZ&C2u6p%jLC;0UoTwJnUo&@!kxko`DpNG^$pharDSwEB%NZU(4 zVU?ou2GYgmZ&MTUKN`k!P|D}XS6Jwl4qmdhFy@?&50iTq^J#(Ytu8))fB3vCSgiKsX@0=QX@ z0r>Igpy$9U;D#X!A~Fv`uo=;9Ll*Z6`pxWDwXr!a`jyC}yqlU#mL6=MUWWA0jx z|8iWy&4I!6LxAW9#a#n4#%1(a9#=k8cg(0lqX`bEZ=<|;vWEN?;xwe@i_z{RlEHPjN5z!6Z>*lj0Lv`PTT5f7~>Ds@E<5e zo^N!dcNFn#?;iXYWa$>&0v(wgLrw;}xI8CNag(FP%jo%D%BK-}R=iCA8^I~4#mngX z5c;m+cOAb(E+57(@`ZnWWc0wfsqTd_TtK!$38xx3F-Y!+WxT-u2qzK zRACF9IcIiO&keb70Z1*7@gv@4w7gH^-x~89)TbnUB(WiArMHkPD2~yCm!}+!CAtzc z5>Y#ggw9A8+6N=#$p#SDEAZ_I#*>sYZ0!OmwAivygd0QG8T5aZlxk>9Z!*YM@{cf; z&7_zbXmMSt@+!^GqTpQw_4|?xu%#pw=x!x3n!-gp`LxvJ2Y~-upI7z2OC9_k)t9`G zSXSa+1UZmoJl7X2lC%lGLkd(w-JDeFt_w{+#CzmQ&flF;Sad`m_fuiKl z@kgnC4hxBYF1$qXRf%bFM6YkqT{}GQ?UQsxK+#0oJ}FN!qq?;{b;V@>&!a&2gYur? z_~)x^wDHd}mpmN&nIzf@{0+M27yjUJYWx@lbZal}*yItN9eior$rg%*T-au@;o_># zI1PY2^xS7_s20A6APb@&Uyz$d#?bJ^_NLxa)_h7)&5}{Y-}ch4`H*p+agR})rs>yw I2G7s_H?4jsKmY&$ literal 0 HcmV?d00001 diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/model_specs.cpython-37.pyc b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/__pycache__/model_specs.cpython-37.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e393e37178dc5b14ef5fd9b128d81e555b8dffb6 GIT binary patch literal 2103 zcmbtVO^@R=7><*Eq-my|X*=ES>@EZWhaza1j+T)STCuFSE)rjdUQAlMZ8T{T*~zF> zoeS^-_9wuRJL14E;5Yb+gw(&liT8CnZ3eC8LM%VFUnfr<$M55Oc68(zaQ*${H~)(h z!}tpax0eZ>Poc{10GKg07-42emgd+b*tei>jV;|TLBBLE>3%sZkITILbLAI9k1JvI zYrMZ2+RNIw1~Hq}elo@mb66ePIxBx=^xgIQFXCtx%%wMrg?H_XAkL)sd>QkJ|B|Pk zFQ$ti<!~ z0ts7}^apS?f%8vjkFU&EM7UZZvZB&myVr_(kq;S83R*QIL~Ya>F<@&UoAjX9*1T+| zwgYWP>;A9WoclhtZfda8jnOrJBtMW1*_az^WNF$!oxpIA-9D3x#=z30*7-lb47s_4wtHY=%dEyIy5^|iRBUOsxh^s>Q>FOyR z3CxDA22B2K8KCgS#ho`!ly$y3FHEC^gP`fwy8HC%&NNfi^VKE2T%E7azKLX(VDfU7 zFE#R){A2H3Ie<)czV*Y5e=bBU`XyDvPeE~BIdQ^;k6A5Mdl4|kBboYWgop)GF4ajU zIirOc`Kuw`1X2GyA1>l0AEsQU6mAk9UJTPD``}`@j2RE*FrLkVY2b&$!+P>5iklz@ z$@i)`hb_prhoYvpfmeP6V3;L><(SyI^hg)_4r!1c>5$ew+JTV^1)`=)#9Q#>|1@J+ zCwUc-nxWDABl%J)j&Xli6+1n zPhA239YeGL#s|}mRwLm;NLk?5@hJ;JI-&;?=WcD|gI$CU!YRTL!UKdJ!2RA5XK+iu zBp%`Ij}e|Ayn%o|$-#r3YGJzXKY-47{ik>alZTA2ki6tMZ^OvhW&L3W^WFg$RBki> z9bo9Ys!Az~r<5v(!tV-ya+C@@g#?}>0>6}UXMPr@3qN8Z7plS|JkeAe|78;qusI)K zH;95%bqYbSd9d`;ScnQzq16PwwXjINTn}Q({sazzEacB{MTzRWs8w z{o<{f-I-yv<#1_Bazu#^0>N?oW7ctyC`b&%FoM7~ko-v;$@iUGRoy*1R306^Z*-fSXh!^qS1PZU< z=XH%#%1PBzj^UVXqhr=hgXJ{J>3Ukq8I&{ijFhwOY(48{4|5Mva-D1E@34KoUFa0+ z#i&iGQ#NtGQ>jl#O`$#6nW|4qxrp*ieMZWq_H1XaJ}2dJd%km`enQHX_Cn`m{Upj0 z^+jj0e#)7upLV95nGaL-GtR6thx1uy-Z_EuBkr7gdc$xQoRc3K_l^3w=z7sPCD)He z*QcE`a&1M|XPrmn`Z0ITS#}8RrS-Nt~Z>E;tu)e$uHr zPvLyQdD^*z^F`+wr-pOYx$IoQ`Ki}Z&XRNW!<2K?efoa7ehJV&>%8E+DA1pAvbR&! zml`D?eZ^|-dVa5Cb$ZTj+qE`&%JO<0*J}1UonF`K^Drrld_wC%>-O}pRgwNbOtX!sg=Hzv~Upz2UaorB6tdRU@2gtb4uouJ78dPQRyoR27`o-IfO= zpgh@Et*&oxL<3(%mGKk~`EGyL?)5zsiyNxZaqsoi4lXiYyVb#YqTx7pLv7k#6FB7q ze)7Onr+2s2-Ne9v>`m7TGpjeg@Gg#Dc=tw_>F&1M%`|(5zijm5wb#YLF$U?taEzUz zx)Ydzv6JTc+s0;UkluUFPX)%ejMJ(6#?92%p8R&|o2hRa8>VAA=?@Ej>RX`Ho47Z~ z1nEN~$UHEN)O%08oAS-WbdUP- z*4;*X*R^^Zpn(3ak9%&{12Hw4s^@vuoiBaaQf|N9YBs#uY7>+ve0)SN5*oYJ0W*{*Oy9 zwOgK#1<-%_2K&WM=28b$842i2MiEa|?WJ}PRPDWtLi|)UA7)jfyXl6xM!)ZN9Yq!x z&Wr=_b~~ZzZ-g13#1FHMzYn^|hB=MDu+-h{*xpt{IbOKn0JLhYgDM?+IOcF#mqFpJ z&2Go-`e8ZlUd!JfUcxPpEmLO3m^L!_pEVYZNpsRrkD)f1 zvK%frWpCl<-NZpC5Hxiv1v<+3scmD&RG-1w3{d-0%1MJlF_nS&U1N|9(&`mI1^UXe zVu%OJi_6=2>1yS3q~nj0vb!Xi~%G*g?mks692f2iA15VS5d%n zHTPcfflMyuld1bA7VJ}lv~L9I`%&E)tc`gf**AlXELtv8y=dp(O$8>hiVGs%{Cf`MR_-iN0D~oSV8c0ywj`TKpLFKe}RZw!gWG@L?C~HI17a! z<)?PiUd1;)Fan?tw72M(U2wPo&_>EyRa7%RN#mfN#sRD&|N81XD{s7hGn`z3a8Q(B zx0LEB+$uW_zhOfRb{akiWUr-{(a40)JS42P0M+BVNSPI^9rZj)$yBqqe7E%YZ4?4y z*kep3_%D0yos|(l|G@~Z_*gMc`V{WnMtdiNx|>p$)pdTx$pyxJ^8>>nSI)0~LTZwM zs;ev!i!$1SldC=7YLGp*9BX87)vUV4M$h69rrT~;J;PGz9rE!Q8(A%h9O{1nQjP|z2Bpp=}L=yP%(Wd=YW=q!)2IlxK+t)-m;(7fhkAocOoW@eD- zR?uFqdF5xfbK7}5g*yIkg6BI$JXiWCJIKl#vVMU*Iz!6NIAwVg_`X2?@6bKWO==*S zkt1pVJLhK)i|8M84X5H%ejaT`j#9Icft1NR6N5tc3kjqn@Q#J_5WEOdW>APml-J`q zGNQbUC_gr$Nf}Wd&y|8~kl!?ef-?mS7J;q!67!*N>GY8k-79!B9k}M{wNKiRI zfDmF}(ygv$HXuH$O;yQ(8s+ya3Va7zVNHDwbFIjy4^~!PsMzjZw`+l^uDgoBplL(7 z_Q35?z3br-EiH$-@sQ!ygeTOC)Q#TKowG^g-xcVfp1`X4W#vSGaQ<_QD%lg9h`fjtHm2PunRK1nH zz8aR%0l>4#-0P)bZ;x`g`5^{}a_^fMf=8a2N>3U&VRExLW+2$VWoHZ^Q zIg`hXDRpVKQmc6GPNK|n5hvWx!9{?j2jvKY>;V=KNIi{Z!^OhoLDdQhK9upalLm=F zyis{VEtX(h`+N$EFa@PO`@q10N*(67;2)Sl4w{23aj06iZo)qDL`7(TY)?4kCzh4uQ)DqPY&8wQ1c@$14s+QD-lTd}7 zUdP_;wn*usW%N{-hEatlej39Li#x6hgAKsLa?7)Q)o67=$1#aezYojZUdwZ-5%=A2 zB7&)njBsKdB&YUmeR*^qi-|e-xA8Jhl-x9RGO|}H8YK$tui##yW~6bE=x58g2vUbo zE`U=58X1K2sYpBzz$T>zjES}PDaxPoO(!c%m^xW*RlNcTg04^g7yh(dKL57c^kF2x zlv2qW?m||3yM5Z7P-v@ZEKXP(zAOe=N2n?$HhDNJcd4Ha>neI(Xha=F#pz)~YPzFM z{~r42QM9B49~ls`qC%G7WK5BH5_g6;aX^<;S9Wshv;%8RFcPUR)$D3$*ct_Y7Efu4 zF;$&qsvMBDKg04@c{rkEEROe=8mQK1e-Dj4vJ>`}%TKMLob;9Rm#B%fD`5y@Ze;d& z<~lMrI6KC{f4bAyfh`NiMS?`D3j*k(l*RVw{cy7O; z8tZM>zP5MGe)fX(%*)orYnQBNtNIz39Lspt8oq7Wx(!}0Y#aJ|v!|5n^?O|yI1mG5 zmUKr@JS6=SNRhpyf!*2V>V-78sOJ>=KGw-4iiyYq!0RmwxF()tVykJvyEZg6R@K*d z^T-^`j?V$*)eq53E4<|b7@u0lb)ue-tCjH+_2f5D0PfBE##hZjW{}k$% zSU9WTQBX<_!N0J!QWVADS1=Cg2t3R~yyj>Pfq#L%KFXm*fvU&o9VbiW1xBdx-f6ln zej>EYv5-)!p-=^?<_@kW6ww$#IxJJ@jhI}zlUguId%_%LzM5FJZhvHg^cp`gB!>5| z4EFG|-(MPF4Np669}>8;+xA;n;I2~FfhhF`4hOUAT^rBYqfZ!ezgiSbi?OCz?Q3`- z%qX|Lt3#(ls@=u zGDCDhCF??zCM+aQRJ?N8fL4GkFt}K1$_7;7#!zp_g?Im`KsJ3fKgOe;=;UU} zP}{gl20@`t79>mH6%>vRVgUOzFn4m^IWcZb2$%E&E|D~SaRKZ-?*-n0eF$wHL(UxB z`2QQSo=(jUqM-{HX0N~jgbP!h(X2Q-!ivS9fig$!($_aI6>O~3(sEpl3s?Rllu@AV1a5Pz-k3A#gqwQNPlSIamu6%F(<6V zU1r;GBEFIbjt=%y2!q^_e)Y&5Al$LfQF3(rqnm&~A&s9!f$pr?H1atE)(O z2WP+Bg=4+#(p^c7a%5>PTX0GcTbh}bVq1X50)U9qALTiCG+rNq(7uYCXp%$LraZ#w z@<7kT(djD)Fmw+gAnkdk&xtMzdY!cA2@X82?0r1m#4q{d>V5@3k1C^^s>26UHyxu+ zRXvU93+>`8&aknaJkB{ph00FlL!%DM)tN*M>_ul9=b|&?%;H>f%g#xOBr$eWaSB-R zi(w&-O*CEuSLB7oNQEVCP{-wu7pTd!I*21!4b|Lg`A{?A5cG~{ zA5^6%B4RUY*}4Np`LI<9BCy;>bL-O-kugoiTejYSr!R!fq7t_bfYbdAAH3Dt5Rp96+`0$U5JeGW zTHO+&g4Y7#)S> zp%n1KSw>-;vH1(F#Em3#^5E@Lba+t^* z!h>J+Guv7C1++Dr7x@Cl^${HA-K(N}q}6xF+7z~n(kgZv8h#17n06cD&0toAgR5`S zn*+@_aiwz2^p$&8TWpECysMP1sjZ4<3rZN}g(%GRf;_Oaw}jg3sDda0T8CY+WsX-BI(UVU0xj@Hab zn^t%vnbBGMCl1aG=^}C` z51}L!GNw~}45*qapmf5@rT1rGMs|^&b6Q!{Dqv%YW?}fC0thYfD*1)`5@W)g!%jne zmXxtq3uv!B37Ue_bGsbC-+@wUyMo`uJg994^yhuWpZvgh?<h&Hrv%J$iZrHGkw1@8y*x@jCFzRz9W~%1mUc-ZZT31O{aZt|B$uv47F7;@D*CDx za6FhBQU;}xJhdvTNc5~G*#`e;LotjYe0~dHg_G0B2A^mA394k86G$!(Wt4vH9GqHt z$Nb3r*!W%zQ=zkf05SX*`pyq?Fvm+l;Q>hNdw;;TAddX0=ry6%N&aFv6(gy^YiX&D?@p5cT}yoqvS zKA}>1P(jkcx($_0+3yeZT+;a;X4>wCzlx*+hOfdrTR}5`BbyViRg~&xRbGQ4U_PZ497o+YPSDWpr3e z?`fC}^mI%bvt|J$T6_41(EwhKHOo3VB9u3qUPR5jIVR~f)F$Vif8^t2Lb6n^a;d_s zfK7LfsS;Qwund_L2V2ni*yv72VLfCGM0c|Y)4}#bAkOIiH{8wR+3zQJO{aj6TM@q6 zG|tir%inNH5#LX5GucN=h!|o)f^}s^A>wDczYNF~K>p(q$cVkbeXtMh;87hIcFaJKm+z{!8rl{BJEjfA`6ZG0K_G} zlE$cZE#^Ndx&=LD)U`+Dq{VzT7#!A0;-R%@+tM)7Cc-$Z4TY>Z$}_ZK{t(L0>!_G^G&?%W zk6j;uwC?hCeJ0B>i7$RIUcgYfmaWx?Vvd^kSP?FOrUYVy%rPgmB{t@WwoTkwjSZ}Q zI#jnBK+I^bhTS<1aq7ERD#9cqWid>{<1XPz_#nuCB1gp1-t+eJdrO~CBG#chJYO%t z#TwPt)3_Dpm*A_vwztIF*P`1|1#022opoAFd2__64^HLUSC6TC_L#cow7-$Qk9v6o z3V>z^NNd0B?!mcn*NDiB1fg(_6pyI1M->>+gvFz`2XM$v7%C$YVG;JbiCf`>bObR4 zTe@&W4>=LBeen+@9r#Yd{H6-x%%wwj8VpY7DNubKy(*crP;FypuPMHigL6ZrMN5|P zS0tQ}T#(ZxobW%A6GB_v$tZ|&Z#wdr8QXA{AOec0=pYB)19z%gro{u*2Jz1OnoE8F zo{|=xavCLi&%O>17%ZXrl#^qM!XV$RAaD=WLSI4MFm_VvSE&Xe&LI3SQh>nUkemRx zKI|c<;1n5!?FOSQ{Swq4gD}8Xi{#YOwyfKZwj6<(KMHfST?7$v5F|wmAx=bZ|22$i zQpVTUIQrP|uA^>h2g(~tV)KEc{_gQG@eEv1lo-;_nF#6|&MfhX=YG?tq6UoKXF}Bw znjR;6WgH8kPv)WD_55}KOG%I$Ss&4I{XDaG1qv7w6 zg{-2G$}kdCg5^scsYLCG`V#b{X=9Swk}wrIWcfb}eX3-hLusBRoJ%a3bkCN|1>@k< zQMQucImtYmz)MT`2}k*#D1dJu=Taio$h1hjBx!OG=3Yep6eK>~H^3jjYY>!K_Ol-t zd*7#aA|}u;B4m?A43pX!Qu7ihdErA+0}fMw0aDmWBPl2yq( z1x0mJ-%-C3r7o8I{9yrxRgrldFgQ*dw6^6PfK|FrE|-0Y=1Eo?9PiykS=`M&`499V zN$f8B60Z{q~gKSjiKLCOHZ>cGlzb60m8z3!z_N(V&$1+|?nRcK#UU)ZFi>chxLVTw z3h#)CL>dm$2d%#L5_}Pt)rpa1^dTzh2$s<@(V?^YjJ^WI*ujKETUa(Z?U1aO#kZsz ztMBmn3GHHz+()7()MxaxFnJ>yvw9P4kD}^VQTHcYY)_()GINl{lUOpdTuK?bnMzoG z)3`DXD?JZaGwop?P$IWZTI!t*yERv6tpEzwPcuOOPg zLO~4yoc69rE&L8zc+<|u0z<;gnG^&$L=1kJMFg-GnNbdd22cMwp3dTE_5&Yc+0Shk z4vP;`IG5<&qStymi|29Yunf0V9<_PE;V*$Dq70$0C^I? z#2o!Jn}Cw(j^5!QdOw{0aDB$xPL2SHwbxjw=1e2{C4TE=t6@dy=`ldYC}crK?c#ke zbaU%%7MBurA3_kJ0`5<7ptoap&K9#zI$ka5i#_I;u32C9c45ozTdNRYmk@wxHMbII zj=Rx7N&@_J$ZgoDNgh^>dA{(a>8OGo?ZIJRE>@Wj25TD7m^+V*r3jL|9OAl**prP+ zTxQlTTX-Q|ei!a8UATnfRXJY65iNbyYfFe8S&YD~O1eZu6C5lJiz|__8V`1}9_=V5 zt&4Y;E(zYEVN{1`lE@euT&-S0aLHQl0mB#XUehT#mUWDk!guyv=NZ0NJcsQ18WfH- zJxIzu{QdFg!Mc7M1+0#%_+5i0@H@Cvzsm!aer=(c(nxB-D89jZvKaN3c=$OUF7fc^ zcsRwwX&l0=%s_bRQ>Q>Z&SrVUt6ClSQW%gN73X1P6hYzqar2^nfv>CKP%jIbXcDw~ zb@-<=mCEH&2J>5JKSf@A9);8ld!V%zV5ilaMu+je?Q+g9HO1nt=skJew| zr6@?$JrsGOD)PY2j|@l@m3h`3;BcO7a}k9?C1V!Kh55oPa?GA8TrNCUIEQVChP3+8 z)8Ab9*FXFp|Ml~){^(DC@F)NB&Gj#RJ}gJuxZq;*WJpaa9sGL=zY2b7{Jihu2oq*P z+$R+nE~!m;Lh07|IQ>0%2a&XiI;KNGT~0rsXLNEq1Fz^5xMxAV;b+0qup?^$#>)pr z_cZErJ9)K%dPL{oyPeuF^*!8SHuvAva2$ABxAX9)P79pxpnZuD;1Vi3d*4Y;Ix=1`HxNY_XFT&Cxg;2XpF(Q37zt{Q{9ZS@a=fs9>KUz4B)&w zT;ML1?UTfyzZg6MmGIGTrm$P|q_c>9c&CC#sZ65QLaWo;XMk-B4z5!lrN5dPJT^E# zczo~~|7>tRCF}g?DErm)-~u5+!OUo7 zv-|RqCodj(@=S0Rn!p^MKlf4L;pc<=$JnLlpWl8wI3qoJA~>^sLQ?bQoJV0aUzB@i z<=&IQ*|B?7)PBaXf{PEZ7^8BP<>12M6Tz&^5!MSSA8)yf7KPwKFv$@^v&ifeQ47U8 zxBz@#2reWZd!(&aA8WjZWxC=Z^|E1IXFwBT5<7v~kv5=HDYd(6%lG@smoGQeUh8hH zr#3HR-{|F~pSp6jcIDdAbI(VL!}(*NEL$s0xWC2W4WHXh{qWLG}%Nf$ws+ z!APz)rYDR2RV8?PFJx>s;cHBgw7J5vfP)L5KhP z?M4$%A$emBs>y4r(ck)Mci*dh`Aauf?%cB9`qJ$+>|6#sbYOqv4HnXn;t?b361y?o zZWG&qboS<&NLajN?NXe}yR0#)yCtvx9F-x;;3kw5AOhOV#ZT!&6&?_eD2VsilNwb+I~fJs*cX3BgYdov)hBth$|1Z3 zb*qU9@Ze*J8jTFV`y3TOq5&aSObMqgYDF9iwlO8peyHC996^a=8{TN(6NqfI{e56f zywjpm;<=HHs(i5z@;#b)r1!UC5yV@BMc=(jy{yCNZ<0XZQ9X5|c<@F`A+adAhpu>n zYRg*F`$(9uy2gASkk{CnrB@m{sD&;@<8iT$$AGmqVpt;^iYutmqCPYjOc`nuh zw$a7g#W>SzN1lm6tzr*%q%1mav$4;_ADm=~5td#Q@&4%=mV3kNbs3w}3Mcl^10`$n zEP4PuG(i?_1RfK&RJ(#n-|u5jH5K|OttrYWcAr6r9rnBILteJ7)Lwk?#Y=LMBj(L- z#_IYTy?ek<{3fmKUQ`%r2T6KZsA`|Ff$09{0IJGy%a&^Ht9MlkOUjMs1KCG<9Q)R0 zugAyrZjLzpOE>@t3l1`+_a%y{MI>KF4L5H`gNzc!_=N6jlSzYW*?N6Ln9U_FZYbs( zq;cx!>qygIBT>gBKt3a8TFv4^uOds#jm?6nHD(#+VGxaY<9Kb6o8CxQM>=EVs5!8}2J znkS2GJ)UZ;DN==lG&eF&p+ihL=Y)%r_8*rrx;OqVjWVQ0wc0elZiCskwXGTPMBqte zi9qdv5_{YtN?Ve!0S}^~r=bivI1<=axjO>r9*E<%1JMv|m{ESKq6-MznL5rThRJ`ME$jk}L1`Wl5ghu7VR4*C zhsN$gXvgCKk;(w{<~>Yk1dfA>dq)vIWPfpO(r1n95eB%7mtu3CY&0e@U>9}GA`d0* z$r0rp%Eh~X!n9EJV9vI@ha@zT7AenSaowTkBkHj*P2gJYjZClPI3%A55R~a2(+dZ` zJHD+j8r9S_cpiaVxKS{rEm){5j&OagQHZ?Wg?a}^Ys}04FQ4kV(5t`>Ll^z9L1k6_>l56S-&yNG9FV{NP-d)UZ{gjaAd+81F_SS*)mm=zJ4I7`T2 zUWE_spW&VAhdlgK9K!gMI5DS@uzs>t#6%N%pZQUfkq86*y5GXrAk=XY22)2Em~Mnn zAN~K@1tg9ih8F$}pvIwvb6Th5dOS9YfRFr0k}@cPU;M#m5@3q8{sg>JvzYItF3#4vbr7<|AdWMG@&gBA4uYrQ5c z(lF6mR9+^yWgHNmgosv51r769Z(H0mJI#Qe*mKcLhQ-8+3a8$D>*lRD>~~h)etqTo z8@FzU)33ep*7cRq8orw!p+Qs-z-g^U!{MyXu=k#%F2{2|4Wb&aS$y-YcV1t8&0cxu z*4rzu-5Pn@EC(^pN5ZnOV6qz!&=fc;wv>Y>>k4&-W2KW+T%hVjmKJz;4TtcAw<}-z z0i%JJJ(kb%A)9a};j(Zjil7eXhSQzc`rLNj0s9Xh0h7ZQVjKosaNW||pY1JvcdFM7 zr#GnVA^kGh%wCor=#TQ?6{DYa3TL3O>flYhQW(09Cb0UHI?HFXBG1tQbRT2K27UOu zl~Bw-;{A!CvIAXeNJZ)&NZ z>0dp5w|sU-k*=|gJpO785@_*_EKJWXlq@*Ev9N3p3@2V&xX?K_|BAr>kPqjzn8AjG zXkm>gW#PhjT?|IpL(D3KOS}qDFOW<8dlXV{X3Rynw&Ci=k1p?wgtYn18m0kO5Y%Sy znq_<$07s;Br4gXPI%lw-0vcOd~WoA5pH@yWV*SAtxN1LUG_X;TQA33HSSa^>K~$X=T-H4 zI4P3g>K*w^#Hx1PuU2d7AE1^#`i))gHi9!AcuFZcycIb$b`CPK1Tu3#%3%)P1G0@d z%(=ae4_$1!H5Mr|@S&k$C2T?LWx_Y!RE|*QHPfR%q^6f+?e3EL&wR^o^T5zaI6Kxz zc$NZZPN{##10$E=g@>Ee61IkKfZ@{|-5MM&5-+Gfc1&x0f?~8CQwze=Bkdv>v1y@% zDNMsvwPL^e*4uBc+_9+*hvPNSll6x(9%D}#HIT0#=^*n;gc zBu!+Sr4A2$9`?2IJ*5f)z35tf%NJmV2g26-Cz2PqdologX% z8qq(95C7@HBqAgw>{`g-SB`#Fkk&+@P{2IXIkyg`kg7gp8i5r<17#@Q-Fu(|8H~?`m3y?@*2xib^^jpbL1QJFNza;eZ^!4;& mlJj$O^Ydcj<1_OzOXB183My}L*yQG?l;)(`fpmNZVg>;6P!=!% literal 0 HcmV?d00001 diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/__pycache__/inference_warmup.cpython-37.pyc b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/__pycache__/inference_warmup.cpython-37.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a90dafac5a7a750aff6052c9d4cf4728befb3262 GIT binary patch literal 2803 zcmZuz&2Jn<7Vqkh`RMtw9Vb9wmjVgIh-4xVLddLEY=EGwAOt5Os1?*|Pu0w{>917P zII-0uAzlovZ^3W+ z%e&G4`j+*7d^!EOuy_w1IfH>&fyD^36O!72O>phN+6f%9c46%Xu33AD7kF^)C5@nA z=4Q|`b365Ii&$wV=mKSfHNUlj9_z6V>wZguKJ)Hb!~W4L4@9iEpp#tCIGb?6voWWS zBazMvD)^&0mr9OKKq+fT^dQ=kc`{c#jME|)3btA-K8R(UXRz)UBFJn}6VKtnsRQnCQ+N_%OS)`~0~3T+%3`uVwSJI2_VAji#JNLPUpDp=t7$8NKqCkk243)Pqeif3RBVRLz`1+-V}VZ4?TN@! zqlmWSTsLnXDlTt-s$Cf!aNU}6br2`#!cA3-xfT0!36R?7+L?S&6dCwm)s*s6DI%fhUX+c%}n$a5A_um;SXYy)^V zRll^4$ra1UXzve}8?~f1Ypv`R`R=Dh_sI{X^MUo{_lw?>f2qw9^ntgp7yYui+yc-I z)TQONy0m|Jd8PEr4r{L*=5He`H_E~OGbKn{J_{N@r>>S8rFU#kN!i=B%HEW)&Qp*9 zKqWx=eE0(En)xqfYyYqCyjTuc?M$4t#C?e`)zo`onu?ijAX$;>Fw92)Tw9WfR>Qk=4QVrrqPM6WBM5e=0m zJgtpTGBQvNw)UWm0>je-TEP37d!&R;(KL&oE}aXCIRI9ag-Lx)62_O^82HB09k$d-UoWO>4d6yr==J4h!(fD>AsM zCLw|lFjE)7v$&l&!06;R7_el^SyXU6IHd)*Me_vPTeA-5kOXIzfgN{-F9**kp2=KH zlKkCH6yiiee1VIgvrn|2WRRsPF@kF|^WuKQ;Y3-LYZ z)j}c6ALrA+uU}!UJp*CAd5?=2kaNWE=l4vH9&QEwKh`E)3o;`7b&PfATnY!>^XaF% z=Tk}#E@Z|!7Qu$WPgsEoE1_;w?8QK9{{Fik@7JOu@A@V6d)qNf#O`Hdk!2E@_Z$c=K(W^z8=Of;~LFYR-_R1%H=p5OEW358XRr z6J~*(S6srKfo_H&%g14;{SZF3umXgkz;7*qpceS0C9s8y7h%wz`A`zC<9d6I*%5pu z0Qg1+Z|Fv~)b_y};$^&t9|1=EIb-n)Ll}!&e*-+s6aIH3lh4B7+Gu3gM%7zn%V}-H ImnZyx2UG`Gb^rhX literal 0 HcmV?d00001 diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/imagenet.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/imagenet.py new file mode 100644 index 000000000..9432e168c --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/imagenet.py @@ -0,0 +1,74 @@ +# Copyright 2019 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Provides data helpers for the ImageNet ILSVRC 2012 Dataset.""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from six.moves import urllib + + +def create_readable_names_for_imagenet_labels(): + """Create a dict mapping label id to human readable string. + + Returns: + labels_to_names: dictionary where keys are integers from to 1000 + and values are human-readable names. + + We retrieve a synset file, which contains a list of valid synset labels used + by ILSVRC competition. There is one synset one per line, eg. + # n01440764 + # n01443537 + We also retrieve a synset_to_human_file, which contains a mapping from synsets + to human-readable names for every synset in Imagenet. These are stored in a + tsv format, as follows: + # n02119247 black fox + # n02119359 silver fox + We assign each synset (in alphabetical order) an integer, starting from 1 + (since 0 is reserved for the background class). + """ + + # pylint: disable=g-line-too-long + base_url = 'https://raw.githubusercontent.com/tensorflow/models/master/research/inception/inception/data/' + # pylint: enable=g-line-too-long + synset_url = '{}/imagenet_lsvrc_2015_synsets.txt'.format(base_url) + synset_to_human_url = '{}/imagenet_metadata.txt'.format(base_url) + + filename, _ = urllib.request.urlretrieve(synset_url) + synset_list = [s.strip() for s in open(filename).readlines()] + num_synsets_in_ilsvrc = len(synset_list) + assert num_synsets_in_ilsvrc == 1000 + + filename, _ = urllib.request.urlretrieve(synset_to_human_url) + synset_to_human_list = open(filename).readlines() + num_synsets_in_all_imagenet = len(synset_to_human_list) + assert num_synsets_in_all_imagenet == 21842 + + synset_to_human = {} + for s in synset_to_human_list: + parts = s.strip().split('\t') + assert len(parts) == 2 + synset = parts[0] + human = parts[1] + synset_to_human[synset] = human + + label_index = 1 + labels_to_names = {0: 'background'} + for synset in synset_list: + name = synset_to_human[synset] + labels_to_names[label_index] = name + label_index += 1 + + return labels_to_names diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/inference_warmup.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/inference_warmup.py new file mode 100644 index 000000000..53380b4b0 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/inference_warmup.py @@ -0,0 +1,97 @@ +# Copyright 2019 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== + +"""Writer for inference warmup requests.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import io +import os +import numpy as np +import tensorflow.compat.v1 as tf + + +def _encode_image(image_array, fmt): + """encodes an (numpy) image array to string. + + Args: + image_array: (numpy) image array + fmt: image format to use + + Returns: + encoded image string + """ + from PIL import Image # pylint: disable=g-import-not-at-top + pil_image = Image.fromarray(image_array) + image_io = io.BytesIO() + pil_image.save(image_io, format=fmt) + return image_io.getvalue() + + +def write_warmup_requests(savedmodel_dir, + model_name, + image_size, + batch_sizes=None, + num_requests=8, + image_format='PNG', + input_signature='input'): + """Writes warmup requests for inference into a tfrecord file. + + Args: + savedmodel_dir: string, the file to the exported model folder. + model_name: string, a model name used inside the model server. + image_size: tuple/list or int, size of image. For list/tuple input, assuming + it contains image height and width. + batch_sizes: list, a list of batch sizes to create different input requests. + num_requests: int, number of requests per batch size. + image_format: string, the format of the image to write (PNG, JPEG) + input_signature: string, input signature defined in exported saved model. + + Raises: + ValueError: if batch_sizes is not a valid integer list. + """ + from tensorflow_serving.apis import predict_pb2 # pylint: disable=g-import-not-at-top + from tensorflow_serving.apis import prediction_log_pb2 # pylint: disable=g-import-not-at-top + if not isinstance(batch_sizes, list) or not batch_sizes: + raise ValueError('batch sizes should be a valid non-empty list.') + extra_assets_dir = os.path.join(savedmodel_dir, 'assets.extra') + tf.gfile.MkDir(extra_assets_dir) + if isinstance(image_size, int): + height = image_size + width = image_size + elif isinstance(image_size, tuple) or isinstance(image_size, list): + height = image_size[0] + width = image_size[1] + else: + raise ValueError( + 'image_size is not a supported type: %s' % type(image_size)) + + with tf.python_io.TFRecordWriter( + os.path.join(extra_assets_dir, 'tf_serving_warmup_requests')) as writer: + for batch_size in batch_sizes: + for _ in range(num_requests): + request = predict_pb2.PredictRequest() + image = np.uint8(np.random.rand(height, width, 3) * 255) + request.inputs[input_signature].CopyFrom( + tf.make_tensor_proto( + [_encode_image(image, image_format)] * batch_size, + shape=[batch_size])) + request.model_spec.name = model_name + request.model_spec.signature_name = 'serving_default' + log = prediction_log_pb2.PredictionLog( + predict_log=prediction_log_pb2.PredictLog(request=request)) + writer.write(log.SerializeToString()) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/tpu_profiler_hook.py b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/tpu_profiler_hook.py new file mode 100644 index 000000000..e8d6e262e --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/common/tpu_profiler_hook.py @@ -0,0 +1,120 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== + +"""TPU Profiler Hook.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import subprocess + +import tensorflow.compat.v1 as tf + + +class TPUProfilerHook(tf.train.SessionRunHook): + """Captures TPU profiling information every N steps or seconds. + + Collects profiles using the cloud tpu profiler. The hook only works in + google cloud with cloud_tpu_profiler installed. + See https://cloud.google.com/tpu/docs/cloud-tpu-tools for detailed usage + for the capture_tpu_profile command. These profiles can be viewed in + Tensorboard. TPU profiling should not be invoked more frequently than every + 10 seconds. + """ + + def __init__(self, + tpu, + output_dir, + save_steps=None, + save_secs=None, + tpu_profiler_command=None): + """Initializes a hook that takes periodic profiling snapshots. + + Args: + tpu: Grpc address to the tpu master. + output_dir: `string`, the directory to save the profile traces to. + save_steps: `int`, save profile traces every N steps. Exactly one of + `save_secs` and `save_steps` should be set. + save_secs: `int` or `float`, save profile traces every N seconds. + tpu_profiler_command: Custom tpu profiler command (e.g. + $install_loc/capture_tpu_profile --duration_ms=20000 + --num_tracing_attempts=10). If not specified, profiling 2 secs with + 3 attempts by default. + + Raises: + ValueError: if `tpu` is not a string. + """ + if not isinstance(tpu, str): + raise ValueError("--tpu should be provided with a string.") + + self._timer = tf.train.SecondOrStepTimer( + every_secs=save_secs, every_steps=save_steps) + self._tpu_profiler_command = None + + if tpu_profiler_command is None: + tpu_profiler_command = ["/usr/local/bin/capture_tpu_profile"] + self._tpu_profiler_command = tpu_profiler_command + if tpu.startswith("grpc://"): + tf.logging.warn( + "Profiling single TPU pointed by %s. Use tpu name to profile a pod." % + tpu) + service_addr = tpu.split("://")[1] + worker = service_addr.split(":")[0] + self._tpu_profiler_command += [ + "--service_addr=" + service_addr, "--workers_list=" + worker + ] + else: + self._tpu_profiler_command += ["--tpu=" + tpu] + self._tpu_profiler_command += ["--logdir=" + output_dir] + self._running_process = None + self._ran_first_step = False + + def begin(self): + self._global_step_tensor = tf.train.get_or_create_global_step() # pylint: disable=protected-access + + def before_run(self, run_context): + return tf.train.SessionRunArgs({"global_step": self._global_step_tensor}) + + def after_run(self, run_context, run_values): + stale_global_step = run_values.results["global_step"] + if not self._ran_first_step: + # Update the timer so that it does not activate until N steps or seconds + # have passed. + self._timer.update_last_triggered_step(stale_global_step) + self._ran_first_step = True + + global_step = stale_global_step + 1 + if (stale_global_step > 1 and + self._timer.should_trigger_for_step(stale_global_step)): + global_step = run_context.session.run(self._global_step_tensor) + self._timer.update_last_triggered_step(global_step) + self._collect_tpu_profile(global_step) + + def _collect_tpu_profile(self, step): + """Run capture_tpu_profile if not already running.""" + + if self._running_process is not None: + exit_code = self._running_process.poll() + if exit_code is not None: + tf.logging.info("Previous profile exited with status: %s", exit_code) + else: + tf.logging.info( + "Profiler is already running, skipping collection at step %d", step) + return + tf.logging.info( + "Saving profile at step %d with command %s", step, + self._tpu_profiler_command) + self._running_process = subprocess.Popen(self._tpu_profiler_command) -- Gitee From e9bdcfed1896f81bfb379a9d03e1f3f80928dc5f Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 06:40:00 +0000 Subject: [PATCH 04/20] =?UTF-8?q?=E4=B8=8A=E4=BC=A0LICENSE?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../AmoebaNet-D_ID2073_for_TensorFlow/LICENSE | 203 ++++++++++++++++++ 1 file changed, 203 insertions(+) create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/LICENSE diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/LICENSE b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/LICENSE new file mode 100644 index 000000000..15ae42140 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/LICENSE @@ -0,0 +1,203 @@ +Copyright 2017 The TensorFlow Authors. All rights reserved. + + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work by the copyright owner + or by an individual or Legal Entity authorized to submit on behalf of + the copyright owner. For the purposes of this definition, "submitted" + means any form of electronic, verbal, or written communication sent + to the Licensor or its representatives, including but not limited to + communication on electronic mailing lists, source code control systems, + and issue tracking systems that are managed by, or on behalf of, the + Licensor for the purpose of discussing and improving the Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 2. Grant of Copyright License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + copyright license to reproduce, prepare Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent License. Subject to the terms and conditions of + this License, each Contributor hereby grants to You a perpetual, + worldwide, non-exclusive, no-charge, royalty-free, irrevocable + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + where such license applies only to those patent claims licensable + by such Contributor that are necessarily infringed by their + Contribution(s) alone or by combination of their Contribution(s) + with the Work to which such Contribution(s) was submitted. If You + institute patent litigation against any entity (including a + cross-claim or counterclaim in a lawsuit) alleging that the Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (and each + Contributor provides its Contributions) on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or + implied, including, without limitation, any warranties or conditions + of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A + PARTICULAR PURPOSE. You are solely responsible for determining the + appropriateness of using or redistributing the Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. Limitation of Liability. In no event and under no legal theory, + whether in tort (including negligence), contract, or otherwise, + unless required by applicable law (such as deliberate and grossly + negligent acts) or agreed to in writing, shall any Contributor be + liable to You for damages, including any direct, indirect, special, + incidental, or consequential damages of any character arising as a + result of this License or out of the use or inability to use the + Work (including but not limited to damages for loss of goodwill, + work stoppage, computer failure or malfunction, or any and all + other commercial damages or losses), even if such Contributor + has been advised of the possibility of such damages. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works thereof, You may choose to offer, + and charge a fee for, acceptance of support, warranty, indemnity, + or other liability obligations and/or rights consistent with this + License. However, in accepting such obligations, You may act only + on Your own behalf and on Your sole responsibility, not on behalf + of any other Contributor, and only if You agree to indemnify, + defend, and hold each Contributor harmless for any liability + incurred by, or claims asserted against, such Contributor by reason + of your accepting any such warranty or additional liability. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright 2017, The TensorFlow Authors. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. -- Gitee From bb72860de316223adb786b4bbc74205f031c8b63 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:11:58 +0000 Subject: [PATCH 05/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md. --- .../README.md | 305 +++++++++++++----- 1 file changed, 231 insertions(+), 74 deletions(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md index c0089b14a..91dcb6fe2 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md @@ -1,109 +1,266 @@ -# AmoebaNet-D on TPU +- [基本信息](#基本信息.md) +- [概述](#概述.md) +- [训练环境准备](#训练环境准备.md) +- [快速上手](#快速上手.md) +- [迁移学习指导](#迁移学习指导.md) +- [高级参考](#高级参考.md) +

基本信息

-This code was implemented based on results in the AmoebaNet paper, which should -be cited as: Real, E., Aggarwal, A., Huang, Y. and Le, Q.V., 2018. Regularized -Evolution for Image Classifier Architecture Search. arXiv preprint -arXiv:1802.01548. +**发布者（Publisher）：Huawei** +**应用领域（Application Domain）： Image Classification** -## Acknowledgements +**修改时间（Modified） ：2022.5.9** -The starting point for this code was branched from the implementation for NASNet -in https://github.com/tensorflow/models/tree/master/research/slim/nets/nasnet -and from image processing code in -https://github.com/tensorflow/tpu/blob/master/models/experimental/inception/inception_preprocessing.py. +**框架（Framework）：TensorFlow 1.15.0** +**模型格式（Model Format）：ckpt** -## Prerequisites +**精度（Precision）：Mixed** -### Setup a Google Cloud project +**处理器（Processor）：昇腾910** -Follow the instructions at the [Quickstart Guide](https://cloud.google.com/tpu/docs/quickstart) -to get a GCE VM with access to Cloud TPU. +**应用级别（Categories）：Official** -To run this model, you will need: +**描述（Description）：基于TensorFlow框架的AmoebaNet-D图像分类网络训练代码** -* A GCE VM instance with an associated Cloud TPU resource -* A GCS bucket to store your training checkpoints -* (Optional): The ImageNet training and validation data preprocessed into - TFRecord format, and stored in GCS. +

概述

-### Installing extra packages +AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像分类神经网络。 -The AmoebaNet trainer uses a few extra packages. We can install them now: +- 参考论文： -``` -pip install -U pillow -pip install -U --no-deps tensorflow-serving-api -``` + [Real, E., Aggarwal, A., Huang, Y., & Le, Q. V. (2019, July). Regularized evolution for image classifier architecture search. In Proceedings of the aaai conference on artificial intelligence (Vol. 33, No. 01, pp. 4780-4789).](https://arxiv.org/pdf/1802.01548.pdf) + + +- 参考实现： + + + +- 适配昇腾 AI 处理器的实现： + + [https://gitee.com/zero167/ModelZoo-TensorFlow/tree/master/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow](https://gitee.com/zero167/ModelZoo-TensorFlow/tree/master/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow) + + + +## 混合精度训练 + +昇腾910 AI处理器提供自动混合精度功能，可以针对全网中float32数据类型的算子，按照内置的优化策略，自动将部分float32的算子降低精度到float16，从而在精度损失很小的情况下提升系统性能并减少内存使用。 + +## 开启混合精度 + +脚本已默认开启混合精度，设置precision_mode参数的脚本参考如下。 + + ``` + + npu_config = NPURunConfig( + model_dir=FLAGS.model_dir, + save_checkpoints_steps=save_checkpoints_steps, + save_summary_steps=0, + # dump_config=dump_config, + # fusion_switch_file="/home/test_user03/tpu-master/models/official/amoeba_net/fusion_switch.cfg", + session_config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False), + #precision_mode="allow_mix_precision") + precision_mode="allow_fp32_to_fp16") + #precision_mode="force_fp32") + ``` + + +

训练环境准备

+ +1. 硬件环境准备请参见各硬件产品文档"[驱动和固件安装升级指南]( https://support.huawei.com/enterprise/zh/category/ai-computing-platform-pid-1557196528909)"。需要在硬件设备上安装与CANN版本配套的固件与驱动。 +2. 宿主机上需要安装Docker并登录[Ascend Hub中心](https://ascendhub.huawei.com/#/detail?name=ascend-tensorflow-arm)获取镜像。 + + 当前模型支持的镜像列表如[表1](#zh-cn_topic_0000001074498056_table1519011227314)所示。 + + **表 1** 镜像列表 + + + + + + + + + + + + +

镜像名称

+

镜像版本

+

配套CANN版本

+
+

20.2.0

+

20.2

+
+ + +

快速上手

+ +- 数据集准备 +1. 模型训练使用ImageNet2012数据集，数据集请用户自行获取。 + +2. 数据集训练前需要做预处理操作，请用户参考[Tensorflow-Slim](https://github.com/tensorflow/models/tree/master/research/slim),将数据集封装为tfrecord格式。 + +3. 数据集处理后，放入模型目录下，在训练脚本中指定数据集路径，可正常使用。 + + +## 模型训练 -### Formatting the data +- 单击“立即下载”，并选择合适的下载方式下载源码包。 -The data is expected to be formatted in TFRecord format, as generated by [this -script](https://github.com/tensorflow/tpu/blob/master/tools/datasets/imagenet_to_gcs.py). +- 启动训练之前，首先要配置程序运行相关环境变量。 -If you do not have ImageNet dataset prepared, you can use a randomly generated -fake dataset to test the model. It is located at -`gs://cloud-tpu-test-datasets/fake_imagenet`. + 环境变量配置信息参见： + [Ascend 910训练平台环境变量设置](https://gitee.com/ascend/modelzoo/wikis/Ascend%20910%E8%AE%AD%E7%BB%83%E5%B9%B3%E5%8F%B0%E7%8E%AF%E5%A2%83%E5%8F%98%E9%87%8F%E8%AE%BE%E7%BD%AE?sort_id=3148819) -## Training the model +- 单卡训练 -Train the model by executing the following command (substituting the appropriate -values): + 1. 配置训练参数。 + + 首先在脚本xx.sh中，配置训练数据集路径和checkpoint保存路径，请用户根据实际路径配置，数据集参数如下所示： + + ``` + DATA_DIR=/home/test_user03/tf_records/ + MODEL_DIR=/home/test_user03/hh + + ``` + + 2. 启动训练。 + + （脚本为xx.sh） + + ``` + bash xx.sh + ``` + + +- 验证。 + + 1. 配置验证参数。 + + 首先在脚本xx.sh中，配置训练数据集路径和checkpoint保存路径，请用户根据实际路径配置，数据集参数如下所示： + + ``` + DATA_DIR=/home/test_user03/tf_records/ + MODEL_DIR=/home/test_user03/hh + + ``` + + 2. 启动验证。 + + （脚本为xx.sh） + + ``` + bash xx.sh + ``` ``` +- 测试用例。 + + 2. 测试用例测试指令（脚本位于xx.sh） + + ``` + bash xx.sh + ``` + + + + +

高级参考

+ +## 脚本和示例代码 + +``` +├── amoeba_net.py //训练与测试入口 +├── README.md //代码说明文档 +├── amoeba_net_model.py //模型功能 +├── model_builder.py //根据用户传入的参数构建模型 +├── model_specs.py //Amoeba_net架构配置 +├── network_utils.py //Amoeba-net使用的常见操作的自定义模块 +├── network_utils_test.py //对network_utils自定义模块的测试 +├── tf_hub.py //模型导出和评估 +├── inception_preprocessing.py //图像预处理 +├── common +│ ├──imagenet.py //为ImageNet ILSVRC 2012数据集提供数据帮助程序 +│ ├──inference_warmup.py //inference warmup实现``` + +## 脚本参数 ``` -python amoeba_net.py \ - --tpu=$TPU_NAME \ - --data_dir=$DATA_DIR \ - --model_dir=$MODEL_DIR +--use_tpu 是否使用tpu，默认：False（由于该代码从tpu版本迁移过来，在晟腾910上只能是False） +--mode 运行模式，默认train_and_eval；可选：train，eval +--data_dir 数据集目录 +--mmodel_dir 保存模型输出的目录 +--num_cells 网络结构中cell的数量，默认：6 +--image_size 图像尺寸，默认：224 +--num_epochs 训练迭代次数，默认：35 +--train_batch_size 训练的batch size，默认：64 +--eval_batch_size 验证的batch size， 默认：64 +--lr 初始学习率，默认：2.56 +--lr_decay_value 学习率指数衰减 默认：0.88 +--lr_warmup_epochs 初始学习率从0增长到指定学习率的迭代数，默认：0.35 ``` -If you are not running this script on a GCE VM in the same project and zone as -your Cloud TPU, you will need to add the `--project` and `--zone` flags -specifying the corresponding values for the Cloud TPU you'd like to use. +##下面两个模块待修改## -This will train an AmoebaNet-D model on ImageNet with 256 batch size on a -single Cloud TPU. With the default flags on everything, the model should train -to above 80% accuracy in under 48 hours (including evaluation time every -few epochs). +## 训练过程 -You can launch TensorBoard (e.g. `tensorboard -logdir=$MODEL_DIR`) to view loss -curves and other metadata regarding your training run. (Note: if you launch -on your VM, be sure to configure ssh port forwarding or the GCE firewall rules -appropriately.) +1. 通过“模型训练”中的训练指令启动单卡或者多卡训练。单卡和多卡通过运行不同脚本，支持单卡、8卡网络训练。 -You can also train the AmoebaNet-D model to 93% top-5 accuracy in under 7.5 -hours using the following command: +2. 参考脚本的模型存储路径为results/1p或者results/8p，训练脚本log中包括如下信息。 ``` -python amoeba_net.py \ - --tpu=$TPU_NAME \ - --data_dir=$DATA_DIR \ - --model_dir=$MODEL_DIR \ - --num_cells=6 \ - --image_size=224 \ - --num_epochs=35 \ - --train_batch_size=1024 \ - --eval_batch_size=1024 \ - --lr=2.56 \ - --lr_decay_value=0.88 \ - --lr_warmup_epochs=0.35 \ - --mode=train \ - --iterations_per_loop=1251 +2020-06-20 22:25:48.893067: I tf_adapter/kernels/geop_npu.cc:64] BuildOutputTensorInfo, num_outputs:1 +2020-06-20 22:25:48.893122: I tf_adapter/kernels/geop_npu.cc:93] BuildOutputTensorInfo, output index:0, total_bytes:8, shape:, tensor_ptr:140670893455168, output140653543141408 +2020-06-20 22:25:48.893165: I tf_adapter/kernels/geop_npu.cc:745] [GEOP] RunGraphAsync callback, status:0, kernel_name:GeOp133_0[ 1330us]step:150120 epoch: 30.0 FPS: 4216.5 loss: 3.373 total_loss: 4.215 lr:0.09106 +2020-06-20 22:25:48.897526: I tf_adapter/kernels/geop_npu.cc:545] [GEOP] Begin GeOp::ComputeAsync, kernel_name:GeOp133_0, num_inputs:0, num_outputs:1 +2020-06-20 22:25:48.897593: I tf_adapter/kernels/geop_npu.cc:412] [GEOP] tf session direct5649af5909132193, graph id: 51 no need to rebuild +2020-06-20 22:25:48.897604: I tf_adapter/kernels/geop_npu.cc:753] [GEOP] Call ge session RunGraphAsync, kernel_name:GeOp133_0 ,tf session: direct5649af5909132193 ,graph id: 51 +2020-06-20 22:25:48.897656: I tf_adapter/kernels/geop_npu.cc:767] [GEOP] End GeOp::ComputeAsync, kernel_name:GeOp133_0, ret_status:success ,tf session: direct5649af5909132193 ,graph id: 51 [0 ms] +2020-06-20 22:25:48.898088: I tf_adapter/kernels/geop_npu.cc:64] BuildOutputTensorInfo, num_outputs:1 +2020-06-20 22:25:48.898118: I tf_adapter/kernels/geop_npu.cc:93] BuildOutputTensorInfo, output index:0, total_bytes:8, shape:, tensor_ptr:140650333523648, output140653566153952 +2020-06-20 22:25:48.898135: I tf_adapter/kernels/geop_npu.cc:745] [GEOP] RunGraphAsync callback, status:0, kernel_name:GeOp133_0[ 529us] +2020-06-20 22:25:48.898456: I tf_adapter/kernels/geop_npu.cc:545] [GEOP] Begin GeOp::ComputeAsync, kernel_name:GeOp133_0, num_inputs:0, num_outputs:1 +2020-06-20 22:25:48.898475: I tf_adapter/kernels/geop_npu.cc:412] [GEOP] tf session direct5649af5909132193, graph id: 51 no need to rebuild +2020-06-20 22:25:48.898485: I tf_adapter/kernels/geop_npu.cc:753] [GEOP] Call ge session RunGraphAsync, kernel_name:GeOp133_0 ,tf session: direct5649af5909132193 ,graph id: 51 ``` -## Understanding the code +## 推理/验证过程 -For more detailed information, read the documentation within each file. +1. 通过“模型训练”中的测试指令启动测试。 -## Additional notes +2. 当前只能针对该工程训练出的checkpoint进行推理测试。 -### About the model and training regime +3. 推理脚本的参数eval_dir可以配置为checkpoint所在的文件夹路径，则该路径下所有.ckpt文件都会根据进行推理。 -The model is the result of evolutionary neural architecture search presented -in [Regularized Evolution for Image Classifier -Search](https://arxiv.org/abs/1802.01548). - -TODO: give some more details +4. 测试结束后会打印验证集的top1 accuracy和top5 accuracy，如下所示。 +``` +2020-06-20 19:06:09.349677: I tf_adapter/kernels/geop_npu.cc:338] [GEOP] GeOp Finalize start, tf session: direct24135e275a110a29, graph_id_: 1 +2020-06-20 19:06:09.349684: I tf_adapter/kernels/geop_npu.cc:342] tf session: direct24135e275a110a29, graph id: 1 +2020-06-20 19:06:09.397087: I tf_adapter/kernels/geop_npu.cc:347] [GEOP] GE Remove Graph success. tf session: direct24135e275a110a29 , graph id: 1 +2020-06-20 19:06:09.397105: I tf_adapter/kernels/geop_npu.cc:368] [GEOP] GeOp Finalize success, tf session: direct24135e275a110a29, graph_id_: 1 +2020-06-20 19:06:09.398108: I tf_adapter/kernels/geop_npu.cc:338] [GEOP] GeOp Finalize start, tf session: direct24135e275a110a29, graph_id_: 31 +2020-06-20 19:06:09.398122: I tf_adapter/kernels/geop_npu.cc:368] [GEOP] GeOp Finalize success, tf session: direct24135e275a110a29, graph_id_: 31 +2020-06-20 19:06:09.398247: I tf_adapter/kernels/host_queue_dataset_op.cc:71] Start destroy tdt. +2020-06-20 19:06:09.412269: I tf_adapter/kernels/host_queue_dataset_op.cc:77] Tdt client close success. +2020-06-20 19:06:09.412288: I tf_adapter/kernels/host_queue_dataset_op.cc:83] dlclose handle finish. +2020-06-20 19:06:09.412316: I tf_adapter/kernels/geop_npu.cc:338] [GEOP] GeOp Finalize start, tf session: direct24135e275a110a29, graph_id_: 51 +2020-06-20 19:06:09.412323: I tf_adapter/kernels/geop_npu.cc:342] tf session: direct24135e275a110a29, graph id: 51 +2020-06-20 19:06:09.553281: I tf_adapter/kernels/geop_npu.cc:347] [GEOP] GE Remove Graph success. tf session: direct24135e275a110a29 , graph id: 51 +2020-06-20 19:06:09.553299: I tf_adapter/kernels/geop_npu.cc:368] [GEOP] GeOp Finalize success, tf session: direct24135e275a110a29, graph_id_: 51 +2020-06-20 19:06:10.619514: I tf_adapter/kernels/host_queue_dataset_op.cc:172] HostQueueDatasetOp's iterator is released. +2020-06-20 19:06:10.620037: I tf_adapter/kernels/geop_npu.cc:338] [GEOP] GeOp Finalize start, tf session: direct24135e275a110a29, graph_id_: 41 +2020-06-20 19:06:10.620054: I tf_adapter/kernels/geop_npu.cc:342] tf session: direct24135e275a110a29, graph id: 41 +2020-06-20 19:06:10.621564: I tf_adapter/kernels/geop_npu.cc:347] [GEOP] GE Remove Graph success. tf session: direct24135e275a110a29 , graph id: 41 +2020-06-20 19:06:10.622904: I tf_adapter/util/session_manager.cc:50] find ge session connect with tf session direct24135e275a110a29 +2020-06-20 19:06:10.975070: I tf_adapter/util/session_manager.cc:55] destory ge session connect with tf session direct24135e275a110a29 success. +2020-06-20 19:06:11.380491: I tf_adapter/kernels/geop_npu.cc:388] [GEOP] Close TsdClient. +2020-06-20 19:06:11.664666: I tf_adapter/kernels/geop_npu.cc:393] [GEOP] Close TsdClient success. +2020-06-20 19:06:11.665011: I tf_adapter/kernels/geop_npu.cc:368] [GEOP] GeOp Finalize success, tf session: direct24135e275a110a29, graph_id_: 41 step epoch top1 top5 loss checkpoint_time(UTC)85068 3.0 50.988 76.99 3.09 +2020-06-20 18:06:0690072 3.0 51.569 77.51 3.03 +2020-06-20 18:11:1495076 3.0 51.689 77.33 3.00 +2020-06-20 18:16:22100080 3.0 51.426 77.04 3.08 +2020-06-20 18:25:11105084 3.0 51.581 77.50 3.03 +2020-06-20 18:34:23Finished evaluation +``` -- Gitee From d62023f76f4c4a312134afca258dc1aa09d70ef8 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:13:25 +0000 Subject: [PATCH 06/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md. --- .../contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md | 1 - 1 file changed, 1 deletion(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md index 91dcb6fe2..21c050401 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md @@ -201,7 +201,6 @@ AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像 --lr_warmup_epochs 初始学习率从0增长到指定学习率的迭代数，默认：0.35 ``` -##下面两个模块待修改## ## 训练过程 -- Gitee From b3c55278017212c00e6ea06a35e2c2989a3ef34e Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:15:08 +0000 Subject: [PATCH 07/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md. --- .../contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md index 21c050401..7f0175468 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md @@ -184,8 +184,8 @@ AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像 │ ├──imagenet.py //为ImageNet ILSVRC 2012数据集提供数据帮助程序 │ ├──inference_warmup.py //inference warmup实现``` +``` ## 脚本参数 - ``` --use_tpu 是否使用tpu，默认：False（由于该代码从tpu版本迁移过来，在晟腾910上只能是False） --mode 运行模式，默认train_and_eval；可选：train，eval -- Gitee From 97c1ee35d330a89071788ed87793180c586e9034 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:18:18 +0000 Subject: [PATCH 08/20] add TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow. --- .../contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/requirements.txt | 0 1 file changed, 0 insertions(+), 0 deletions(-) create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/requirements.txt diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/requirements.txt b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/requirements.txt new file mode 100644 index 000000000..e69de29bb -- Gitee From 6a3d6a333f66a406419480a33c3eb788629692a2 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:26:55 +0000 Subject: [PATCH 09/20] add TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/modelzoo_level.txt. --- .../cv/AmoebaNet-D_ID2073_for_TensorFlow/modelzoo_level.txt | 6 ++++++ 1 file changed, 6 insertions(+) create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/modelzoo_level.txt diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/modelzoo_level.txt b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/modelzoo_level.txt new file mode 100644 index 000000000..382bc227b --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/modelzoo_level.txt @@ -0,0 +1,6 @@ +GPUStatus:OK +NPUMigrationStatus:POK +FuncStatus:OK +PrecisionStatus:OK +AutoTune:OK +PerfStatus:PERFECT -- Gitee From 5a9a69a1e6f95e02157620d2f881ea3b8adffa18 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:33:54 +0000 Subject: [PATCH 10/20] =?UTF-8?q?=E9=87=8D=E5=91=BD=E5=90=8D=20TensorFlow/?= =?UTF-8?q?contrib/cv/AmoebaNet-D=5FID2073=5Ffor=5FTensorFlow/xx.sh=20?= =?UTF-8?q?=E4=B8=BA=20TensorFlow/contrib/cv/AmoebaNet-D=5FID2073=5Ffor=5F?= =?UTF-8?q?TensorFlow/train=5Fperformance=5F1p.sh?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../{xx.sh => train_performance_1p.sh} | 0 1 file changed, 0 insertions(+), 0 deletions(-) rename TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/{xx.sh => train_performance_1p.sh} (100%) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/xx.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_performance_1p.sh similarity index 100% rename from TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/xx.sh rename to TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_performance_1p.sh -- Gitee From f165bcc1406ab28f274e98e9898221acf920dd17 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:35:04 +0000 Subject: [PATCH 11/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_performance_1p.sh. --- .../AmoebaNet-D_ID2073_for_TensorFlow/train_performance_1p.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_performance_1p.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_performance_1p.sh index 71381ed8f..2a0c81757 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_performance_1p.sh +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_performance_1p.sh @@ -17,5 +17,5 @@ nohup python3 amoeba_net.py \ --lr_warmup_epochs=0.35 \ --mode=train \ --iterations_per_loop=1251 \ - > xx.log 2>&1 & + > train_performance_1p.log 2>&1 & -- Gitee From 533fe9daadcb9480dd716fae3caa1c90ced26c76 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:36:15 +0000 Subject: [PATCH 12/20] add TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_full_1p.sh. --- .../train_full_1p.sh | 21 +++++++++++++++++++ 1 file changed, 21 insertions(+) create mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_full_1p.sh diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_full_1p.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_full_1p.sh new file mode 100644 index 000000000..a7123af15 --- /dev/null +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_full_1p.sh @@ -0,0 +1,21 @@ +#! /bin/bash + +DATA_DIR=/home/test_user03/tf_records/ +MODEL_DIR=/home/test_user03/xx + + +nohup python3 amoeba_net.py \ + --data_dir=${DATA_DIR} \ + --model_dir=${MODEL_DIR} \ + --num_cells=6 \ + --image_size=224 \ + --num_epochs=35 \ + --train_batch_size=64 \ + --eval_batch_size=64 \ + --lr=2.56 \ + --lr_decay_value=0.88 \ + --lr_warmup_epochs=0.35 \ + --mode=train_and_eval \ + --iterations_per_loop=1251 \ + > train_full_1p.log 2>&1 & + -- Gitee From 93353beaa397cab56aae6190a46970fe75b63afe Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:37:10 +0000 Subject: [PATCH 13/20] =?UTF-8?q?=E5=88=A0=E9=99=A4=E6=96=87=E4=BB=B6=20Te?= =?UTF-8?q?nsorFlow/contrib/cv/AmoebaNet-D=5FID2073=5Ffor=5FTensorFlow/env?= =?UTF-8?q?.sh?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../AmoebaNet-D_ID2073_for_TensorFlow/env.sh | 23 ------------------- 1 file changed, 23 deletions(-) delete mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/env.sh diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/env.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/env.sh deleted file mode 100644 index 7802ada07..000000000 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/env.sh +++ /dev/null @@ -1,23 +0,0 @@ -export RANK_INDEX=0 -export RANK_SIZE=1 -export RANK_ID=0 -export ASCEND_DEVICE_ID=2 -export DEVICE_INDEX=0 -export install_path=/usr/local/Ascend -export JOB_ID=100001 -#nnae -export LD_LIBRARY_PATH=${install_path}/nnae/latest/fwkacllib/lib64:$LD_LIBRARY_PATH -export PATH=${install_path}/nnae/latest/fwkacllib/ccec_compiler/bin:${install_path}/nnae/latest/fwkacllib/bin:$PATH -export PYTHONPATH=${install_path}/nnae/latest/fwkacllib/python/site-packages:$PYTHONPATH -export ASCEND_OPP_PATH=${install_path}/nnae/latest/opp -export ASCEND_AICPU_PATH=${install_path}/nnae/latest -#tfplugin -export PYTHONPATH=${install_path}/tfplugin/latest/tfplugin/python/site-packages:$PYTHONPATH -#Ascend-dmi -export LD_LIBRARY_PATH=/usr/local/dcmi:${install_path}/toolbox/latest/Ascend-DMI/lib64:${LD_LIBRARY_PATH} -export PATH=${install_path}/toolbox/latest/Ascend-DMI/bin:${PATH} -export LD_LIBRARY_PATH=/usr/local/gcc7.3.0/lib64:${LD_LIBRARY_PATH} -export PATH=/usr/local/gcc7.3.0/bin:${PATH} -export LD_LIBRARY_PATH=/usr/local/python3.7.5/lib:$LD_LIBRARY_PATH -export PATH=/usr/local/python3.7.5/bin:$PATH -export LD_LIBRARY_PATH=/usr/include/hdf5/lib:$LD_LIBRARY_PATH -- Gitee From fe3ef2c2290356191ff300db8932f29495fa3233 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:37:33 +0000 Subject: [PATCH 14/20] =?UTF-8?q?=E5=88=A0=E9=99=A4=E6=96=87=E4=BB=B6=20Te?= =?UTF-8?q?nsorFlow/contrib/cv/AmoebaNet-D=5FID2073=5Ffor=5FTensorFlow/run?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/run | 2 -- 1 file changed, 2 deletions(-) delete mode 100644 TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/run diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/run b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/run deleted file mode 100644 index fbbd642aa..000000000 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/run +++ /dev/null @@ -1,2 +0,0 @@ -bash train_testcase.sh -bash online_inference_testcase.sh \ No newline at end of file -- Gitee From defa73b8b5a4badfb3fc4b3c98c7174b80168418 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:51:24 +0000 Subject: [PATCH 15/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md. --- .../README.md | 125 +++++++++--------- 1 file changed, 62 insertions(+), 63 deletions(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md index 7f0175468..1edbd7c6b 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md @@ -119,7 +119,7 @@ AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像 1. 配置训练参数。 - 首先在脚本xx.sh中，配置训练数据集路径和checkpoint保存路径，请用户根据实际路径配置，数据集参数如下所示： + 首先在脚本train_performance_1p.sh中，配置训练数据集路径和checkpoint保存路径，请用户根据实际路径配置，示例如下所示： ``` DATA_DIR=/home/test_user03/tf_records/ @@ -127,12 +127,10 @@ AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像 ``` - 2. 启动训练。 - - （脚本为xx.sh） + 2. 启动训练。（脚本为train_performance_1p.sh） ``` - bash xx.sh + bash train_performance_1p.sh ``` @@ -140,7 +138,7 @@ AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像 1. 配置验证参数。 - 首先在脚本xx.sh中，配置训练数据集路径和checkpoint保存路径，请用户根据实际路径配置，数据集参数如下所示： + 首先在脚本train_full_1p.sh中，配置训练数据集路径和checkpoint保存路径，请用户根据实际路径配置，示例如下所示： ``` DATA_DIR=/home/test_user03/tf_records/ @@ -148,20 +146,12 @@ AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像 ``` - 2. 启动验证。 - - （脚本为xx.sh） + 2. 启动验证。（脚本为train_full_1p.sh） ``` - bash xx.sh - ``` ``` -- 测试用例。 + bash train_full_1p.sh + ``` - 2. 测试用例测试指令（脚本位于xx.sh） - - ``` - bash xx.sh - ``` @@ -204,62 +194,71 @@ AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像 ## 训练过程 -1. 通过“模型训练”中的训练指令启动单卡或者多卡训练。单卡和多卡通过运行不同脚本，支持单卡、8卡网络训练。 +1. 通过“模型训练”中的训练指令启动单卡训练。 -2. 参考脚本的模型存储路径为results/1p或者results/8p，训练脚本log中包括如下信息。 +2. 训练脚本log中包括如下信息。 ``` -2020-06-20 22:25:48.893067: I tf_adapter/kernels/geop_npu.cc:64] BuildOutputTensorInfo, num_outputs:1 -2020-06-20 22:25:48.893122: I tf_adapter/kernels/geop_npu.cc:93] BuildOutputTensorInfo, output index:0, total_bytes:8, shape:, tensor_ptr:140670893455168, output140653543141408 -2020-06-20 22:25:48.893165: I tf_adapter/kernels/geop_npu.cc:745] [GEOP] RunGraphAsync callback, status:0, kernel_name:GeOp133_0[ 1330us]step:150120 epoch: 30.0 FPS: 4216.5 loss: 3.373 total_loss: 4.215 lr:0.09106 -2020-06-20 22:25:48.897526: I tf_adapter/kernels/geop_npu.cc:545] [GEOP] Begin GeOp::ComputeAsync, kernel_name:GeOp133_0, num_inputs:0, num_outputs:1 -2020-06-20 22:25:48.897593: I tf_adapter/kernels/geop_npu.cc:412] [GEOP] tf session direct5649af5909132193, graph id: 51 no need to rebuild -2020-06-20 22:25:48.897604: I tf_adapter/kernels/geop_npu.cc:753] [GEOP] Call ge session RunGraphAsync, kernel_name:GeOp133_0 ,tf session: direct5649af5909132193 ,graph id: 51 -2020-06-20 22:25:48.897656: I tf_adapter/kernels/geop_npu.cc:767] [GEOP] End GeOp::ComputeAsync, kernel_name:GeOp133_0, ret_status:success ,tf session: direct5649af5909132193 ,graph id: 51 [0 ms] -2020-06-20 22:25:48.898088: I tf_adapter/kernels/geop_npu.cc:64] BuildOutputTensorInfo, num_outputs:1 -2020-06-20 22:25:48.898118: I tf_adapter/kernels/geop_npu.cc:93] BuildOutputTensorInfo, output index:0, total_bytes:8, shape:, tensor_ptr:140650333523648, output140653566153952 -2020-06-20 22:25:48.898135: I tf_adapter/kernels/geop_npu.cc:745] [GEOP] RunGraphAsync callback, status:0, kernel_name:GeOp133_0[ 529us] -2020-06-20 22:25:48.898456: I tf_adapter/kernels/geop_npu.cc:545] [GEOP] Begin GeOp::ComputeAsync, kernel_name:GeOp133_0, num_inputs:0, num_outputs:1 -2020-06-20 22:25:48.898475: I tf_adapter/kernels/geop_npu.cc:412] [GEOP] tf session direct5649af5909132193, graph id: 51 no need to rebuild -2020-06-20 22:25:48.898485: I tf_adapter/kernels/geop_npu.cc:753] [GEOP] Call ge session RunGraphAsync, kernel_name:GeOp133_0 ,tf session: direct5649af5909132193 ,graph id: 51 +I0420 23:35:31.718360 281472996956240 basic_session_run_hooks.py:692] global_step/sec: 4.13978 +INFO:tensorflow:global_step...700647 +I0420 23:35:31.722282 281472996956240 npu_hook.py:132] global_step...700647 +INFO:tensorflow:global_step...700648 +I0420 23:35:31.963601 281472996956240 npu_hook.py:132] global_step...700648 +... +INFO:tensorflow:Saving checkpoints for 700662 into /home/test_user03/ckpt5/model.ckpt. +I0420 23:35:35.366074 281472996956240 basic_session_run_hooks.py:606] Saving checkpoints for 700662 into /home/test_user03/ckpt5/model.ckpt. +INFO:tensorflow:global_step...700663 +I0420 23:36:39.784266 281472996956240 npu_hook.py:132] global_step...700663 +INFO:tensorflow:global_step...700664 +I0420 23:36:40.024840 281472996956240 npu_hook.py:132] global_step...700664 +INFO:tensorflow:global_step...700665 +I0420 23:36:40.267009 281472996956240 npu_hook.py:132] global_step...700665 +INFO:tensorflow:NPUCheckpointSaverHook end... +I0420 23:36:40.267664 281472996956240 npu_hook.py:137] NPUCheckpointSaverHook end... +INFO:tensorflow:Saving checkpoints for 700665 into /home/test_user03/ckpt5/model.ckpt. +I0420 23:36:40.269501 281472996956240 basic_session_run_hooks.py:606] Saving checkpoints for 700665 into /home/test_user03/ckpt5/model.ckpt. +INFO:tensorflow:Loss for final step: 4.1664658. +I0420 23:38:08.704852 281472996956240 estimator.py:371] Loss for final step: 4.1664658. ``` ## 推理/验证过程 -1. 通过“模型训练”中的测试指令启动测试。 +1. 通过“模型训练”中的验证指令启动验证。 2. 当前只能针对该工程训练出的checkpoint进行推理测试。 -3. 推理脚本的参数eval_dir可以配置为checkpoint所在的文件夹路径，则该路径下所有.ckpt文件都会根据进行推理。 - -4. 测试结束后会打印验证集的top1 accuracy和top5 accuracy，如下所示。 +3. 测试结束后会打印测试集的top1 accuracy和top5 accuracy，如下所示。 ``` -2020-06-20 19:06:09.349677: I tf_adapter/kernels/geop_npu.cc:338] [GEOP] GeOp Finalize start, tf session: direct24135e275a110a29, graph_id_: 1 -2020-06-20 19:06:09.349684: I tf_adapter/kernels/geop_npu.cc:342] tf session: direct24135e275a110a29, graph id: 1 -2020-06-20 19:06:09.397087: I tf_adapter/kernels/geop_npu.cc:347] [GEOP] GE Remove Graph success. tf session: direct24135e275a110a29 , graph id: 1 -2020-06-20 19:06:09.397105: I tf_adapter/kernels/geop_npu.cc:368] [GEOP] GeOp Finalize success, tf session: direct24135e275a110a29, graph_id_: 1 -2020-06-20 19:06:09.398108: I tf_adapter/kernels/geop_npu.cc:338] [GEOP] GeOp Finalize start, tf session: direct24135e275a110a29, graph_id_: 31 -2020-06-20 19:06:09.398122: I tf_adapter/kernels/geop_npu.cc:368] [GEOP] GeOp Finalize success, tf session: direct24135e275a110a29, graph_id_: 31 -2020-06-20 19:06:09.398247: I tf_adapter/kernels/host_queue_dataset_op.cc:71] Start destroy tdt. -2020-06-20 19:06:09.412269: I tf_adapter/kernels/host_queue_dataset_op.cc:77] Tdt client close success. -2020-06-20 19:06:09.412288: I tf_adapter/kernels/host_queue_dataset_op.cc:83] dlclose handle finish. -2020-06-20 19:06:09.412316: I tf_adapter/kernels/geop_npu.cc:338] [GEOP] GeOp Finalize start, tf session: direct24135e275a110a29, graph_id_: 51 -2020-06-20 19:06:09.412323: I tf_adapter/kernels/geop_npu.cc:342] tf session: direct24135e275a110a29, graph id: 51 -2020-06-20 19:06:09.553281: I tf_adapter/kernels/geop_npu.cc:347] [GEOP] GE Remove Graph success. tf session: direct24135e275a110a29 , graph id: 51 -2020-06-20 19:06:09.553299: I tf_adapter/kernels/geop_npu.cc:368] [GEOP] GeOp Finalize success, tf session: direct24135e275a110a29, graph_id_: 51 -2020-06-20 19:06:10.619514: I tf_adapter/kernels/host_queue_dataset_op.cc:172] HostQueueDatasetOp's iterator is released. -2020-06-20 19:06:10.620037: I tf_adapter/kernels/geop_npu.cc:338] [GEOP] GeOp Finalize start, tf session: direct24135e275a110a29, graph_id_: 41 -2020-06-20 19:06:10.620054: I tf_adapter/kernels/geop_npu.cc:342] tf session: direct24135e275a110a29, graph id: 41 -2020-06-20 19:06:10.621564: I tf_adapter/kernels/geop_npu.cc:347] [GEOP] GE Remove Graph success. tf session: direct24135e275a110a29 , graph id: 41 -2020-06-20 19:06:10.622904: I tf_adapter/util/session_manager.cc:50] find ge session connect with tf session direct24135e275a110a29 -2020-06-20 19:06:10.975070: I tf_adapter/util/session_manager.cc:55] destory ge session connect with tf session direct24135e275a110a29 success. -2020-06-20 19:06:11.380491: I tf_adapter/kernels/geop_npu.cc:388] [GEOP] Close TsdClient. -2020-06-20 19:06:11.664666: I tf_adapter/kernels/geop_npu.cc:393] [GEOP] Close TsdClient success. -2020-06-20 19:06:11.665011: I tf_adapter/kernels/geop_npu.cc:368] [GEOP] GeOp Finalize success, tf session: direct24135e275a110a29, graph_id_: 41 step epoch top1 top5 loss checkpoint_time(UTC)85068 3.0 50.988 76.99 3.09 -2020-06-20 18:06:0690072 3.0 51.569 77.51 3.03 -2020-06-20 18:11:1495076 3.0 51.689 77.33 3.00 -2020-06-20 18:16:22100080 3.0 51.426 77.04 3.08 -2020-06-20 18:25:11105084 3.0 51.581 77.50 3.03 -2020-06-20 18:34:23Finished evaluation +ow:Evaluation [78/781] +I0420 23:40:00.258475 281472996956240 evaluation.py:167] Evaluation [78/781] +INFO:tensorflow:Evaluation [156/781] +I0420 23:40:07.798311 281472996956240 evaluation.py:167] Evaluation [156/781] +INFO:tensorflow:Evaluation [234/781] +I0420 23:40:15.336721 281472996956240 evaluation.py:167] Evaluation [234/781] +INFO:tensorflow:Evaluation [312/781] +I0420 23:40:22.876575 281472996956240 evaluation.py:167] Evaluation [312/781] +INFO:tensorflow:Evaluation [390/781] +I0420 23:40:30.432068 281472996956240 evaluation.py:167] Evaluation [390/781] +INFO:tensorflow:Evaluation [468/781] +I0420 23:40:38.020324 281472996956240 evaluation.py:167] Evaluation [468/781] +INFO:tensorflow:Evaluation [546/781] +I0420 23:40:45.564076 281472996956240 evaluation.py:167] Evaluation [546/781] +INFO:tensorflow:Evaluation [624/781] +I0420 23:40:53.106832 281472996956240 evaluation.py:167] Evaluation [624/781] +INFO:tensorflow:Evaluation [702/781] +I0420 23:41:00.634234 281472996956240 evaluation.py:167] Evaluation [702/781] +INFO:tensorflow:Evaluation [780/781] +I0420 23:41:08.236136 281472996956240 evaluation.py:167] Evaluation [780/781] +INFO:tensorflow:Evaluation [781/781] +I0420 23:41:08.331177 281472996956240 evaluation.py:167] Evaluation [781/781] +2022-04-20 23:41:08.749352: I /home/phisik3/jenkins/workspace/work_code/tmp/host-prefix/src/host-build/asl/tfadaptor/CMakeFiles/tf_adapter.dir/compiler_depend.ts:805] The model has been compiled on the Ascend AI processor, current graph id is: 71 +INFO:tensorflow:Finished evaluation at 2022-04-20-23:41:13 +I0420 23:41:13.806376 281472996956240 evaluation.py:275] Finished evaluation at 2022-04-20-23:41:13 +INFO:tensorflow:Saving dict for global step 700665: global_step = 700665, loss = 1.8883309, top_1_accuracy = 0.75600195, top_5_accuracy = 0.9269366 +I0420 23:41:13.807576 281472996956240 estimator.py:2049] Saving dict for global step 700665: global_step = 700665, loss = 1.8883309, top_1_accuracy = 0.75600195, top_5_accuracy = 0.9269366 +INFO:tensorflow:Saving 'checkpoint_path' summary for global step 700665: /home/test_user03/ckpt5/model.ckpt-700665 +I0420 23:41:13.810555 281472996956240 estimator.py:2109] Saving 'checkpoint_path' summary for global step 700665: /home/test_user03/ckpt5/model.ckpt-700665 +INFO:tensorflow:Evaluation results: {'loss': 1.8883309, 'top_1_accuracy': 0.75600195, 'top_5_accuracy': 0.9269366, 'global_step': 700665} +I0420 23:41:13.813197 281472996956240 amoeba_net.py:467] Evaluation results: {'loss': 1.8883309, 'top_1_accuracy': 0.75600195, 'top_5_accuracy': 0.9269366, 'global_step': 700665} ``` -- Gitee From 6409621f2529e060b5d2896d7268ec4cb79ec1a0 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:52:24 +0000 Subject: [PATCH 16/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md. --- .../contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md | 1 - 1 file changed, 1 deletion(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md index 1edbd7c6b..ed6b10222 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md @@ -2,7 +2,6 @@ - [概述](#概述.md) - [训练环境准备](#训练环境准备.md) - [快速上手](#快速上手.md) -- [迁移学习指导](#迁移学习指导.md) - [高级参考](#高级参考.md)

基本信息

-- Gitee From b3b7e0cdfdf9c462b19cfa3c4fe70c1be4004352 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 07:59:18 +0000 Subject: [PATCH 17/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md. --- .../cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md index ed6b10222..f4cdfb842 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/README.md @@ -169,6 +169,12 @@ AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像 ├── network_utils_test.py //对network_utils自定义模块的测试 ├── tf_hub.py //模型导出和评估 ├── inception_preprocessing.py //图像预处理 +├── train_testcase.sh //训练测试用例 +├── online_inference_testcase.sh //在线推理测试用例 +├── train_performance_1p.sh //训练入口 +├── train_full_1p.sh //训练入口，包含准确率评估 +├── modelzoo_level.txt //网络状态描述文件 +├── requirements.txt ├── common │ ├──imagenet.py //为ImageNet ILSVRC 2012数据集提供数据帮助程序 │ ├──inference_warmup.py //inference warmup实现``` @@ -177,9 +183,9 @@ AmoebaNet-D是由AmoebaNet演化神经架构搜索算法搜索出的一个图像 ## 脚本参数 ``` --use_tpu 是否使用tpu，默认：False（由于该代码从tpu版本迁移过来，在晟腾910上只能是False） ---mode 运行模式，默认train_and_eval；可选：train，eval +--mode 运行模式，可选：train_and_eval，train，eval --data_dir 数据集目录 ---mmodel_dir 保存模型输出的目录 +--mmodel_dir 保存checkpoint的目录 --num_cells 网络结构中cell的数量，默认：6 --image_size 图像尺寸，默认：224 --num_epochs 训练迭代次数，默认：35 -- Gitee From ec1e0a429a55eb96608ac7e2f99d7f26926a9765 Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 08:02:02 +0000 Subject: [PATCH 18/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh. --- .../train_testcase.sh | 17 +++++++++-------- 1 file changed, 9 insertions(+), 8 deletions(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh index 622b3a99d..dc71fd22c 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh @@ -1,12 +1,13 @@ #! /bin/bash +#使用ImageNet2012数据集，配置DATA_DIR数据集路径和MODEL_DIR保存checkpoint路径 DATA_DIR=/home/test_user03/tf_records/ MODEL_DIR=/home/test_user03/hh -#߸˶ԤõݼԤѵģ͡ATC-OMģ͵ȣִ֧OBS +#开发者个人独立预置的数据集、预训练模型、ATC-OM模型等，支持从OBS仓下载 #obsutil cp obs://obsxxx/xxx/xxx.ckpt ./model/ -f -r -#testcase壬߸ݲͬģд +#testcase主体，开发者根据不同模型写作 python3 amoeba_net.py \ --data_dir=${DATA_DIR} \ --model_dir=${MODEL_DIR} \ @@ -23,15 +24,15 @@ python3 amoeba_net.py \ --num_train_images=10000 \ --num_eval_images=1000 \ > train.log 2>&1 -#ѵֻѵ10000ͼƬӡϢtrain.log +#训练测试用例只训练10000张图片，并保存打印信息至train.log -#жϣܼckpt/־ؼ֡ȼlossֵ/accucyؼܼ֡ʱ/ThroughOutputȹؼ -key1="Saving checkpoints" #ܼ -key2="global_step/sec:" #ܼ -key3="loss = " #ȼ +#结果判断，功能检查输出ckpt/日志关键字、精度检查loss值/accucy关键字、性能检查耗时打点/ThroughOutput等关键字 +key1="Saving checkpoints" #功能检查字 +key2="global_step/sec:" #性能检查字 +key3="loss = " #精度检查字 -if [ `grep -c "$key1" "train.log"` -ne '0' ] && [ `grep -c "$key2" "train.log"` -ne '0' ] && [ `grep -c "$key3" "train.log"` -ne '0' ];then #ԸҪ߼ +if [ `grep -c "$key1" "train.log"` -ne '0' ] && [ `grep -c "$key2" "train.log"` -ne '0' ] && [ `grep -c "$key3" "train.log"` -ne '0' ];then #可以根据需要调整检查逻辑 echo "Run testcase success!" else echo "Run testcase failed!" -- Gitee From 12d627e36ea00605f255830ec5a2e3637328ddad Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 9 May 2022 08:02:35 +0000 Subject: [PATCH 19/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/online_inference_testcase.sh. --- .../online_inference_testcase.sh | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/online_inference_testcase.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/online_inference_testcase.sh index f241eb3e5..fa4b6006f 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/online_inference_testcase.sh +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/online_inference_testcase.sh @@ -1,12 +1,13 @@ #! /bin/bash +#使用ImageNet2012数据集，配置DATA_DIR数据集路径和MODEL_DIR保存checkpoint路径 DATA_DIR=/home/test_user03/tf_records/ MODEL_DIR=/home/test_user03/hh -#߸˶ԤõݼԤѵģ͡ATC-OMģ͵ȣִ֧OBS +#开发者个人独立预置的数据集、预训练模型、ATC-OM模型等，支持从OBS仓下载 #obsutil cp obs://obsxxx/xxx/xxx.ckpt ./model/ -f -r -#testcase壬߸ݲͬģд +#testcase主体，开发者根据不同模型写作 python3 amoeba_net.py \ --data_dir=${DATA_DIR} \ --model_dir=${MODEL_DIR} \ @@ -24,15 +25,15 @@ python3 amoeba_net.py \ --num_eval_images=1000 \ --eval_timeout=10 \ > predict.log 2>&1 -#ֻ1000ͼƬӡϢpredict.log +#在线推理测试用例只推理1000张图片，并保存打印信息至predict.log -#жϣܼckpt/־ؼ֡ȼlossֵ/accucyؼܼ֡ʱ/ThroughOutputȹؼ -key1="Restoring parameters from" #ܼ -key2="Inference speed =" #ܼ +#结果判断，功能检查输出ckpt/日志关键字、精度检查loss值/accucy关键字、性能检查耗时打点/ThroughOutput等关键字 +key1="Restoring parameters from" #功能检查字 +key2="Inference speed =" #性能检查字 -if [ `grep -c "$key1" "predict.log"` -ne '0' ] && [ `grep -c "$key2" "predict.log"` -ne '0' ];then #ԸҪ߼ +if [ `grep -c "$key1" "predict.log"` -ne '0' ] && [ `grep -c "$key2" "predict.log"` -ne '0' ];then #可以根据需要调整检查逻辑 echo "Run testcase success!" else echo "Run testcase failed!" -- Gitee From b0ab786b42ceb66eece940824a7f0016780793da Mon Sep 17 00:00:00 2001 From: zero167 <215051281@qq.com> Date: Mon, 16 May 2022 13:18:29 +0000 Subject: [PATCH 20/20] update TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh. --- .../cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh index dc71fd22c..1dd6d0781 100644 --- a/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh +++ b/TensorFlow/contrib/cv/AmoebaNet-D_ID2073_for_TensorFlow/train_testcase.sh @@ -1,11 +1,11 @@ #! /bin/bash #使用ImageNet2012数据集，配置DATA_DIR数据集路径和MODEL_DIR保存checkpoint路径 -DATA_DIR=/home/test_user03/tf_records/ -MODEL_DIR=/home/test_user03/hh +DATA_DIR=./tf_records/ +MODEL_DIR=./ckpt/ #开发者个人独立预置的数据集、预训练模型、ATC-OM模型等，支持从OBS仓下载 -#obsutil cp obs://obsxxx/xxx/xxx.ckpt ./model/ -f -r +obsutil cp obs://amoebanet/tf_records_testcase/ ./tf_records/ -f -r #testcase主体，开发者根据不同模型写作 python3 amoeba_net.py \ -- Gitee