From a1e2217e3ba75b202dd36d2b2fee7703cd0f221b Mon Sep 17 00:00:00 2001 From: lvmingfu Date: Thu, 14 Apr 2022 14:09:02 +0800 Subject: [PATCH] modify urls adapt docs repository structure --- .jenkins/check/config/filter_linklint.txt | 2 ++ benchmark/ascend/bert/README.md | 2 -- benchmark/ascend/resnet/README.md | 2 +- benchmark/ascend/resnet/README_CN.md | 2 +- official/audio/melgan/README.md | 2 +- official/audio/melgan/README_CN.md | 2 +- official/cv/FCN8s/README.md | 2 +- official/cv/c3d/README.md | 2 +- official/cv/cnnctc/README.md | 6 +++--- official/cv/cnnctc/README_CN.md | 6 +++--- official/cv/crnn/README.md | 2 +- official/cv/crnn_seq2seq_ocr/README.md | 2 +- official/cv/cspdarknet53/README.md | 4 ++-- official/cv/ctpn/README.md | 2 +- official/cv/darknet53/README.md | 2 +- official/cv/deeplabv3/README.md | 2 +- official/cv/deeplabv3/README_CN.md | 2 +- official/cv/deeplabv3plus/README_CN.md | 2 +- official/cv/deeptext/README.md | 2 +- official/cv/densenet/README.md | 2 +- official/cv/densenet/README_CN.md | 2 +- official/cv/depthnet/README.md | 2 +- official/cv/dpn/README.md | 2 +- official/cv/east/README.md | 2 +- official/cv/essay-recogination/README_CN.md | 2 +- official/cv/googlenet/README.md | 6 +++--- official/cv/googlenet/README_CN.md | 4 ++-- official/cv/inceptionv3/README.md | 2 +- official/cv/inceptionv3/README_CN.md | 2 +- official/cv/inceptionv4/README.md | 4 ++-- official/cv/maskrcnn/README.md | 2 +- official/cv/maskrcnn/README_CN.md | 2 +- official/cv/maskrcnn_mobilenetv1/README.md | 4 ++-- official/cv/mobilenetv1/README.md | 2 +- official/cv/mobilenetv2/README.md | 2 +- official/cv/mobilenetv2/README_CN.md | 2 +- official/cv/nima/README.md | 2 +- official/cv/openpose/README.md | 2 +- official/cv/patchcore/README_CN.md | 2 +- official/cv/predrnn++/README.md | 2 +- official/cv/psenet/README.md | 2 +- official/cv/psenet/README_CN.md | 2 +- official/cv/pvnet/README.md | 2 +- official/cv/resnet/README.md | 4 ++-- official/cv/resnet/README_CN.md | 2 +- official/cv/resnext/README.md | 2 +- official/cv/resnext/README_CN.md | 2 +- official/cv/retinanet/README_CN.md | 2 +- official/cv/semantic_human_matting/README.md | 2 +- official/cv/simple_pose/README.md | 2 +- official/cv/squeezenet/README.md | 4 ++-- official/cv/squeezenet/modelarts/README.md | 4 ++-- official/cv/srcnn/README_CN.md | 2 +- official/cv/ssd/README.md | 2 +- official/cv/ssd/README_CN.md | 2 +- official/cv/ssim-ae/README_CN.md | 2 +- official/cv/tinydarknet/README_CN.md | 2 +- official/cv/unet/README.md | 2 +- official/cv/unet/README_CN.md | 2 +- official/cv/unet3d/README.md | 2 +- official/cv/vgg16/README.md | 4 ++-- official/cv/vgg16/README_CN.md | 4 ++-- official/cv/vit/README.md | 4 ++-- official/cv/vit/README_CN.md | 4 ++-- official/cv/warpctc/README.md | 2 +- official/cv/warpctc/README_CN.md | 2 +- official/cv/xception/README.md | 4 ++-- official/cv/yolov3_resnet18/README.md | 4 ++-- official/cv/yolov3_resnet18/README_CN.md | 4 ++-- official/nlp/bert/README.md | 2 -- official/nlp/cpm/README.md | 2 +- official/nlp/cpm/README_CN.md | 2 +- official/nlp/duconv/README_CN.md | 2 +- official/nlp/mass/README.md | 4 ++-- official/nlp/mass/README_CN.md | 4 ++-- official/nlp/pangu_alpha/README.md | 12 ++++++------ official/nlp/transformer/README.md | 2 +- official/nlp/transformer/README_CN.md | 2 +- official/recommend/ncf/README.md | 6 +++--- research/audio/fcn-4/README.md | 2 +- research/audio/speech_transformer/README.md | 2 +- research/cv/3D_DenseNet/README.md | 2 +- research/cv/3D_DenseNet/README_CN.md | 4 +--- research/cv/APDrawingGAN/README_CN.md | 2 +- research/cv/AlignedReID++/README_CN.md | 4 ++-- research/cv/AlphaPose/README_CN.md | 2 +- research/cv/DDRNet/README_CN.md | 2 +- research/cv/EDSR/README_CN.md | 2 +- research/cv/EGnet/README_CN.md | 2 +- research/cv/GENet_Res50/README_CN.md | 2 +- research/cv/LightCNN/README.md | 6 +++--- research/cv/LightCNN/README_CN.md | 4 ++-- research/cv/ManiDP/Readme.md | 2 +- research/cv/NFNet/README_CN.md | 2 +- research/cv/RefineDet/README_CN.md | 2 +- research/cv/RefineNet/README.md | 2 +- research/cv/SE-Net/README.md | 2 +- research/cv/SE_ResNeXt50/README_CN.md | 2 +- research/cv/TNT/README_CN.md | 2 +- research/cv/cct/README_CN.md | 2 +- research/cv/convnext/README_CN.md | 2 +- research/cv/dcgan/README.md | 2 +- research/cv/deeplabv3plus/README_CN.md | 2 +- research/cv/dlinknet/README.md | 2 +- research/cv/dlinknet/README_CN.md | 2 +- research/cv/efficientnetv2/README_CN.md | 2 +- research/cv/fairmot/README.md | 2 +- research/cv/fishnet99/README_CN.md | 2 +- research/cv/glore_res/README_CN.md | 2 +- research/cv/glore_res200/README_CN.md | 2 +- research/cv/glore_res50/README.md | 2 +- research/cv/hardnet/README_CN.md | 6 +++--- research/cv/inception_resnet_v2/README.md | 4 ++-- research/cv/inception_resnet_v2/README_CN.md | 4 ++-- research/cv/mae/README_CN.md | 4 ++-- research/cv/metric_learn/README_CN.md | 2 +- research/cv/midas/README.md | 2 +- research/cv/nas-fpn/README_CN.md | 2 +- research/cv/ntsnet/README.md | 2 +- research/cv/osnet/README.md | 2 +- research/cv/ras/README.md | 2 +- research/cv/renas/Readme.md | 2 +- research/cv/res2net/README.md | 2 +- research/cv/res2net_deeplabv3/README.md | 2 +- research/cv/resnet3d/README_CN.md | 2 +- research/cv/resnet50_bam/README.md | 2 +- research/cv/resnet50_bam/README_CN.md | 2 +- research/cv/resnext152_64x4d/README.md | 2 +- research/cv/resnext152_64x4d/README_CN.md | 2 +- research/cv/retinanet_resnet101/README.md | 2 +- research/cv/retinanet_resnet101/README_CN.md | 2 +- research/cv/retinanet_resnet152/README.md | 2 +- research/cv/retinanet_resnet152/README_CN.md | 2 +- research/cv/siamRPN/README_CN.md | 2 +- research/cv/simple_baselines/README_CN.md | 2 +- research/cv/single_path_nas/README.md | 2 +- research/cv/single_path_nas/README_CN.md | 2 +- research/cv/sknet/README.md | 2 +- research/cv/squeezenet/README.md | 4 ++-- research/cv/squeezenet1_1/README.md | 2 +- research/cv/ssd_ghostnet/README.md | 2 +- research/cv/ssd_inception_v2/README.md | 2 +- research/cv/ssd_inceptionv2/README_CN.md | 2 +- research/cv/ssd_mobilenetV2/README.md | 2 +- research/cv/ssd_mobilenetV2_FPNlite/README.md | 2 +- research/cv/ssd_resnet34/README.md | 2 +- research/cv/ssd_resnet34/README_CN.md | 2 +- research/cv/ssd_resnet50/README.md | 2 +- research/cv/ssd_resnet50/README_CN.md | 2 +- research/cv/ssd_resnet_34/README.md | 2 +- research/cv/swin_transformer/README_CN.md | 2 +- research/cv/tsm/README_CN.md | 2 +- research/cv/vgg19/README.md | 2 +- research/cv/vgg19/README_CN.md | 4 ++-- research/cv/vnet/README_CN.md | 2 +- research/cv/wideresnet/README.md | 2 +- research/cv/wideresnet/README_CN.md | 2 +- research/hpc/pinns/README.md | 4 ++-- research/hpc/pinns/README_CN.md | 2 +- research/nlp/albert/README.md | 5 ++--- research/nlp/atae_lstm/README.md | 2 +- research/nlp/rotate/README_CN.md | 2 +- research/nlp/seq2seq/README_CN.md | 2 +- 163 files changed, 204 insertions(+), 209 deletions(-) create mode 100644 .jenkins/check/config/filter_linklint.txt diff --git a/.jenkins/check/config/filter_linklint.txt b/.jenkins/check/config/filter_linklint.txt new file mode 100644 index 000000000..bbd5911cd --- /dev/null +++ b/.jenkins/check/config/filter_linklint.txt @@ -0,0 +1,2 @@ +http://www.vision.caltech.edu/visipedia/CUB-200-2011.html +http://dl.yf.io/dla/models/imagenet/dla34-ba72cf86.pth \ No newline at end of file diff --git a/benchmark/ascend/bert/README.md b/benchmark/ascend/bert/README.md index 5d9cceade..8c46ef6a6 100644 --- a/benchmark/ascend/bert/README.md +++ b/benchmark/ascend/bert/README.md @@ -209,8 +209,6 @@ Please follow the instructions in the link below to create an hccl.json file in For distributed training among multiple machines, training command should be executed on each machine in a small time interval. Thus, an hccl.json is needed on each machine. [merge_hccl](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools#merge_hccl) is a tool to create hccl.json for multi-machine case. -For dataset, if you want to set the format and parameters, a schema configuration file with JSON format needs to be created, please refer to [tfrecord](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_loading.html#tfrecord) format. - ```text For pretraining, schema file contains ["input_ids", "input_mask", "segment_ids", "next_sentence_labels", "masked_lm_positions", "masked_lm_ids", "masked_lm_weights"]. diff --git a/benchmark/ascend/resnet/README.md b/benchmark/ascend/resnet/README.md index c1e12fc6e..f8d916c77 100644 --- a/benchmark/ascend/resnet/README.md +++ b/benchmark/ascend/resnet/README.md @@ -97,7 +97,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/benchmark/ascend/resnet/README_CN.md b/benchmark/ascend/resnet/README_CN.md index 9663c783e..3df9996ec 100644 --- a/benchmark/ascend/resnet/README_CN.md +++ b/benchmark/ascend/resnet/README_CN.md @@ -103,7 +103,7 @@ ResNet的总体网络架构如下： ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/audio/melgan/README.md b/official/audio/melgan/README.md index fbe180bef..edab4f65f 100644 --- a/official/audio/melgan/README.md +++ b/official/audio/melgan/README.md @@ -73,7 +73,7 @@ Dataset used: [LJ Speech]() ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/official/audio/melgan/README_CN.md b/official/audio/melgan/README_CN.md index 6e5727b3f..ee2279296 100644 --- a/official/audio/melgan/README_CN.md +++ b/official/audio/melgan/README_CN.md @@ -70,7 +70,7 @@ MelGAN模型是非自回归全卷积模型。它的参数比同类模型少得 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/FCN8s/README.md b/official/cv/FCN8s/README.md index 1ce8edd2a..8e5bfc381 100644 --- a/official/cv/FCN8s/README.md +++ b/official/cv/FCN8s/README.md @@ -471,7 +471,7 @@ python export.py ### 教程 -如果你需要在不同硬件平台（如GPU，Ascend 910 或者 Ascend 310）使用训练好的模型，你可以参考这个 [Link](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。以下是一个简单例子的步骤介绍： +如果你需要在不同硬件平台（如GPU，Ascend 910 或者 Ascend 310）使用训练好的模型，你可以参考这个 [Link](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。以下是一个简单例子的步骤介绍： - Running on Ascend diff --git a/official/cv/c3d/README.md b/official/cv/c3d/README.md index fb6a298ec..dbee0dabb 100644 --- a/official/cv/c3d/README.md +++ b/official/cv/c3d/README.md @@ -324,7 +324,7 @@ epoch time: 150760.797 ms, per step time: 252.954 ms #### Distributed training on Ascend > Notes: -> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. +> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. > ```text diff --git a/official/cv/cnnctc/README.md b/official/cv/cnnctc/README.md index 7606795b7..f05825fed 100644 --- a/official/cv/cnnctc/README.md +++ b/official/cv/cnnctc/README.md @@ -1,4 +1,4 @@ -# Contents +# Contents - [CNNCTC Description](#CNNCTC-description) - [Model Architecture](#model-architecture) @@ -94,7 +94,7 @@ This takes around 75 minutes. ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) @@ -517,7 +517,7 @@ accuracy: 0.8533 ### Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following the steps below, this is a simple example: +If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: - Running on Ascend diff --git a/official/cv/cnnctc/README_CN.md b/official/cv/cnnctc/README_CN.md index 31b0e6257..4125a7ec3 100644 --- a/official/cv/cnnctc/README_CN.md +++ b/official/cv/cnnctc/README_CN.md @@ -95,7 +95,7 @@ python src/preprocess_dataset.py ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 @@ -250,7 +250,7 @@ bash scripts/run_distribute_train_ascend.sh [RANK_TABLE_FILE] [PRETRAINED_CKPT(o > 注意: - RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html), 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). + RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html), 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). ### 训练结果 @@ -449,7 +449,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [DVPP] [DEVICE_ID] ### 推理 -如果您需要在GPU、Ascend 910、Ascend 310等多个硬件平台上使用训练好的模型进行推理，请参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。以下为简单示例： +如果您需要在GPU、Ascend 910、Ascend 310等多个硬件平台上使用训练好的模型进行推理，请参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。以下为简单示例： - Ascend处理器环境运行 diff --git a/official/cv/crnn/README.md b/official/cv/crnn/README.md index b76d1730b..237142ab1 100644 --- a/official/cv/crnn/README.md +++ b/official/cv/crnn/README.md @@ -238,7 +238,7 @@ Parameters for both training and evaluation can be set in default_config.yaml. ## [Training Process](#contents) -- Set options in `config.py`, including learning rate and other network hyperparameters. Click [MindSpore dataset preparation tutorial](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `config.py`, including learning rate and other network hyperparameters. Click [MindSpore dataset preparation tutorial](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. ### [Training](#contents) diff --git a/official/cv/crnn_seq2seq_ocr/README.md b/official/cv/crnn_seq2seq_ocr/README.md index 1707ba078..01d056e7b 100644 --- a/official/cv/crnn_seq2seq_ocr/README.md +++ b/official/cv/crnn_seq2seq_ocr/README.md @@ -229,7 +229,7 @@ Parameters for both training and evaluation can be set in config.py. ## [Training Process](#contents) -- Set options in `default_config.yaml`, including learning rate and other network hyperparameters. Click [MindSpore dataset preparation tutorial](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `default_config.yaml`, including learning rate and other network hyperparameters. Click [MindSpore dataset preparation tutorial](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. ### [Training](#contents) diff --git a/official/cv/cspdarknet53/README.md b/official/cv/cspdarknet53/README.md index cbfd2038d..e8670cb12 100644 --- a/official/cv/cspdarknet53/README.md +++ b/official/cv/cspdarknet53/README.md @@ -49,7 +49,7 @@ Dataset used can refer to paper. ## [Mixed Precision(Ascend)](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. @@ -206,7 +206,7 @@ bash run_distribute_train.sh [RANK_TABLE_FILE] [DATA_DIR] (option)[PATH_CHECKPOI bash run_standalone_train.sh [DEVICE_ID] [DATA_DIR] (option)[PATH_CHECKPOINT] ``` -> Notes: RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html), and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV3, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. +> Notes: RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html), and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV3, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. > > This is processor cores binding operation regarding the `device_num` and total processor numbers. If you are not expect to do it, remove the operations `taskset` in `scripts/run_distribute_train.sh` diff --git a/official/cv/ctpn/README.md b/official/cv/ctpn/README.md index a89a39c8f..58be8b7ba 100644 --- a/official/cv/ctpn/README.md +++ b/official/cv/ctpn/README.md @@ -231,7 +231,7 @@ imagenet_cfg = edict({ Then you can train it with ImageNet2012. > Notes: -> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. +> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. > > This is processor cores binding operation regarding the `device_num` and total processor numbers. If you are not expect to do it, remove the operations `taskset` in `scripts/run_distribute_train.sh` > diff --git a/official/cv/darknet53/README.md b/official/cv/darknet53/README.md index 804488515..168662486 100644 --- a/official/cv/darknet53/README.md +++ b/official/cv/darknet53/README.md @@ -58,7 +58,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/official/cv/deeplabv3/README.md b/official/cv/deeplabv3/README.md index e8dc4d9f9..185eda20a 100644 --- a/official/cv/deeplabv3/README.md +++ b/official/cv/deeplabv3/README.md @@ -86,7 +86,7 @@ You can also generate the list file automatically by run script: `python get_dat ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/official/cv/deeplabv3/README_CN.md b/official/cv/deeplabv3/README_CN.md index 742956f10..8017d3f9c 100644 --- a/official/cv/deeplabv3/README_CN.md +++ b/official/cv/deeplabv3/README_CN.md @@ -93,7 +93,7 @@ Pascal VOC数据集和语义边界数据集（Semantic Boundaries Dataset，SBD ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/deeplabv3plus/README_CN.md b/official/cv/deeplabv3plus/README_CN.md index 9985407b6..29c133a23 100644 --- a/official/cv/deeplabv3plus/README_CN.md +++ b/official/cv/deeplabv3plus/README_CN.md @@ -83,7 +83,7 @@ Pascal VOC数据集和语义边界数据集（Semantic Boundaries Dataset，SBD ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/deeptext/README.md b/official/cv/deeptext/README.md index d30cccba7..dfc9863ac 100644 --- a/official/cv/deeptext/README.md +++ b/official/cv/deeptext/README.md @@ -133,7 +133,7 @@ sh run_eval_gpu.sh [IMGS_PATH] [ANNOS_PATH] [CHECKPOINT_PATH] [COCO_TEXT_PARSER_ ``` > Notes: -> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. +> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. > > This is processor cores binding operation regarding the `device_num` and total processor numbers. If you are not expect to do it, remove the operations `taskset` in `scripts/run_distribute_train.sh` > diff --git a/official/cv/densenet/README.md b/official/cv/densenet/README.md index 97dca3522..6e7f3532c 100644 --- a/official/cv/densenet/README.md +++ b/official/cv/densenet/README.md @@ -79,7 +79,7 @@ The default configuration of the Dataset are as follows: ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. diff --git a/official/cv/densenet/README_CN.md b/official/cv/densenet/README_CN.md index 74c86403e..8e6e8b9f8 100644 --- a/official/cv/densenet/README_CN.md +++ b/official/cv/densenet/README_CN.md @@ -83,7 +83,7 @@ DenseNet-100使用的数据集： Cifar-10 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/depthnet/README.md b/official/cv/depthnet/README.md index 5af2a58f5..d1868a4e7 100644 --- a/official/cv/depthnet/README.md +++ b/official/cv/depthnet/README.md @@ -74,7 +74,7 @@ ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/dpn/README.md b/official/cv/dpn/README.md index 106d98cd2..f2ec942bb 100644 --- a/official/cv/dpn/README.md +++ b/official/cv/dpn/README.md @@ -67,7 +67,7 @@ All the models in this repository are trained and validated on ImageNet-1K. The ## [Mixed Precision](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/official/cv/east/README.md b/official/cv/east/README.md index 3686d8fdf..76c1e00e6 100644 --- a/official/cv/east/README.md +++ b/official/cv/east/README.md @@ -130,7 +130,7 @@ bash run_eval_gpu.sh [DATASET_PATH] [CKPT_PATH] [DEVICE_ID] ``` > Notes: -> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. +> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. > > This is processor cores binding operation regarding the `device_num` and total processor numbers. If you are not expect to do it, remove the operations `taskset` in `scripts/run_distribute_train.sh` > diff --git a/official/cv/essay-recogination/README_CN.md b/official/cv/essay-recogination/README_CN.md index 6456047af..064425d29 100644 --- a/official/cv/essay-recogination/README_CN.md +++ b/official/cv/essay-recogination/README_CN.md @@ -111,7 +111,7 @@ train.valInterval = 100 #边训练边推 ## 训练过程 -- 在`parameters/hwdb.gin`中设置选项，包括学习率和网络超参数。单击[MindSpore加载数据集教程](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/dataset_sample.html)，了解更多信息。 +- 在`parameters/hwdb.gin`中设置选项，包括学习率和网络超参数。单击[MindSpore加载数据集教程](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset.html)，了解更多信息。 ### 训练 diff --git a/official/cv/googlenet/README.md b/official/cv/googlenet/README.md index 3fb862acb..04708b39f 100644 --- a/official/cv/googlenet/README.md +++ b/official/cv/googlenet/README.md @@ -1,4 +1,4 @@ -# Contents +# Contents [查看中文](./README_CN.md) @@ -71,7 +71,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) @@ -595,7 +595,7 @@ Current batch_ Size can only be set to 1. ### Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following the steps below, this is a simple example: +If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: - Running on Ascend diff --git a/official/cv/googlenet/README_CN.md b/official/cv/googlenet/README_CN.md index d8f0e8885..569ed28fa 100644 --- a/official/cv/googlenet/README_CN.md +++ b/official/cv/googlenet/README_CN.md @@ -73,7 +73,7 @@ GoogleNet由多个inception模块串联起来，可以更加深入。 降维的 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 @@ -596,7 +596,7 @@ python export.py --config_path [CONFIG_PATH] ### 推理 -如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。下面是操作步骤示例： +如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。下面是操作步骤示例： - Ascend处理器环境运行 diff --git a/official/cv/inceptionv3/README.md b/official/cv/inceptionv3/README.md index 3c5bd47fd..e445e7022 100644 --- a/official/cv/inceptionv3/README.md +++ b/official/cv/inceptionv3/README.md @@ -65,7 +65,7 @@ Dataset used: [CIFAR-10](http://www.cs.toronto.edu/~kriz/cifar.html) ## [Mixed Precision(Ascend)](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. diff --git a/official/cv/inceptionv3/README_CN.md b/official/cv/inceptionv3/README_CN.md index ff3189a67..cb1910b17 100644 --- a/official/cv/inceptionv3/README_CN.md +++ b/official/cv/inceptionv3/README_CN.md @@ -69,7 +69,7 @@ InceptionV3的总体网络架构如下： ## 混合精度（Ascend） -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 diff --git a/official/cv/inceptionv4/README.md b/official/cv/inceptionv4/README.md index 4cf8250d4..7cf88d279 100644 --- a/official/cv/inceptionv4/README.md +++ b/official/cv/inceptionv4/README.md @@ -44,7 +44,7 @@ Dataset used can refer to paper. ## [Mixed Precision(Ascend)](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. @@ -263,7 +263,7 @@ bash scripts/run_standalone_train_ascend.sh [DEVICE_ID] [DATA_DIR] ``` > Notes: -> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. +> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. > > This is processor cores binding operation regarding the `device_num` and total processor numbers. If you are not expect to do it, remove the operations `taskset` in `scripts/run_distribute_train.sh` diff --git a/official/cv/maskrcnn/README.md b/official/cv/maskrcnn/README.md index 3689e43e1..cc3be8102 100644 --- a/official/cv/maskrcnn/README.md +++ b/official/cv/maskrcnn/README.md @@ -544,7 +544,7 @@ Usage: bash run_standalone_train.sh [PRETRAINED_MODEL] [DATA_PATH] ## [Training Process](#contents) -- Set options in `config.py`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `config.py`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. ### [Training](#content) diff --git a/official/cv/maskrcnn/README_CN.md b/official/cv/maskrcnn/README_CN.md index cbe608e85..fcca9b9d0 100644 --- a/official/cv/maskrcnn/README_CN.md +++ b/official/cv/maskrcnn/README_CN.md @@ -526,7 +526,7 @@ bash run_eval.sh [VALIDATION_JSON_FILE] [CHECKPOINT_PATH] [DATA_PATH] ## 训练过程 -- 在`config.py`中设置配置项，包括loss_scale、学习率和网络超参。单击[此处](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/dataset_sample.html)获取更多数据集相关信息. +- 在`config.py`中设置配置项，包括loss_scale、学习率和网络超参。单击[此处](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset.html)获取更多数据集相关信息. ### 训练 diff --git a/official/cv/maskrcnn_mobilenetv1/README.md b/official/cv/maskrcnn_mobilenetv1/README.md index 57e5ecd32..3d8b11269 100644 --- a/official/cv/maskrcnn_mobilenetv1/README.md +++ b/official/cv/maskrcnn_mobilenetv1/README.md @@ -1,4 +1,4 @@ -# Contents +# Contents - [MaskRCNN Description](#maskrcnn-description) - [Model Architecture](#model-architecture) @@ -521,7 +521,7 @@ Usage: bash run_distribute_train_gpu.sh [DATA_PATH] [PRETRAINED_PATH] (optional) ## [Training Process](#contents) -- Set options in `default_config.yaml`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `default_config.yaml`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. ### [Training](#content) diff --git a/official/cv/mobilenetv1/README.md b/official/cv/mobilenetv1/README.md index ce1a3c4b0..5f0771154 100644 --- a/official/cv/mobilenetv1/README.md +++ b/official/cv/mobilenetv1/README.md @@ -73,7 +73,7 @@ Dataset used: [CIFAR-10](http://www.cs.toronto.edu/~kriz/cifar.html) ### Mixed Precision(Ascend) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. ## Environment Requirements diff --git a/official/cv/mobilenetv2/README.md b/official/cv/mobilenetv2/README.md index 454cce4bb..e7b2b046a 100644 --- a/official/cv/mobilenetv2/README.md +++ b/official/cv/mobilenetv2/README.md @@ -59,7 +59,7 @@ Dataset used: [imagenet](http://www.image-net.org/) ## [Mixed Precision(Ascend)](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/official/cv/mobilenetv2/README_CN.md b/official/cv/mobilenetv2/README_CN.md index 35af3e3d4..88caa2261 100644 --- a/official/cv/mobilenetv2/README_CN.md +++ b/official/cv/mobilenetv2/README_CN.md @@ -55,7 +55,7 @@ MobileNetV2总体网络架构如下： ## 混合精度（Ascend） -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/nima/README.md b/official/cv/nima/README.md index 0ddce6559..47485334c 100644 --- a/official/cv/nima/README.md +++ b/official/cv/nima/README.md @@ -84,7 +84,7 @@ python ./src/dividing_label.py --config_path=~/config.yaml ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/openpose/README.md b/official/cv/openpose/README.md index 3bf5319a6..43387c844 100644 --- a/official/cv/openpose/README.md +++ b/official/cv/openpose/README.md @@ -69,7 +69,7 @@ In the currently provided training script, the coco2017 data set is used as an e ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/official/cv/patchcore/README_CN.md b/official/cv/patchcore/README_CN.md index 98018773c..353c7beff 100644 --- a/official/cv/patchcore/README_CN.md +++ b/official/cv/patchcore/README_CN.md @@ -93,7 +93,7 @@ PatchCore使用预训练的WideResNet50作为Encoder, 并去除layer3之后的 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/predrnn++/README.md b/official/cv/predrnn++/README.md index 0018d700f..6e569bc90 100644 --- a/official/cv/predrnn++/README.md +++ b/official/cv/predrnn++/README.md @@ -140,7 +140,7 @@ device_id: 0 # id of NPU used ## [Training Process](#contents) -- Set options in `config.py`, including learning rate and other network hyperparameters. Click [MindSpore dataset preparation tutorial](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `config.py`, including learning rate and other network hyperparameters. Click [MindSpore dataset preparation tutorial](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. ### [Training](#contents) diff --git a/official/cv/psenet/README.md b/official/cv/psenet/README.md index 87e844ece..c297a2637 100644 --- a/official/cv/psenet/README.md +++ b/official/cv/psenet/README.md @@ -427,7 +427,7 @@ The `res` folder is generated in the upper-level directory. For details about th ### Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following the steps below, this is a simple example: +If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: ```python # Load unseen dataset for inference diff --git a/official/cv/psenet/README_CN.md b/official/cv/psenet/README_CN.md index 9a3061b8e..9225e8127 100644 --- a/official/cv/psenet/README_CN.md +++ b/official/cv/psenet/README_CN.md @@ -364,7 +364,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [DEVICE_ID] ### 推理 -如果您需要使用已训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考[此处](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。操作示例如下： +如果您需要使用已训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考[此处](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。操作示例如下： ```python # 加载未知数据集进行推理 diff --git a/official/cv/pvnet/README.md b/official/cv/pvnet/README.md index 7b5579ebb..c462410d3 100644 --- a/official/cv/pvnet/README.md +++ b/official/cv/pvnet/README.md @@ -62,7 +62,7 @@ PvNet是一种Encode-Decode的网络结构，通过输入一张rgb图，输出 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/resnet/README.md b/official/cv/resnet/README.md index 116a9cc2a..a88ab183a 100644 --- a/official/cv/resnet/README.md +++ b/official/cv/resnet/README.md @@ -107,7 +107,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) @@ -456,7 +456,7 @@ bash run_eval_gpu_resnet_benchmark.sh [DATASET_PATH] [CKPT_PATH] [BATCH_SIZE](op For distributed training, a hostfile configuration needs to be created in advance. -Please follow the instructions in the link [GPU-Multi-Host](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_gpu.html). +Please follow the instructions in the link [GPU-Multi-Host](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_gpu.html). #### Running parameter server mode training diff --git a/official/cv/resnet/README_CN.md b/official/cv/resnet/README_CN.md index 9663c783e..3df9996ec 100644 --- a/official/cv/resnet/README_CN.md +++ b/official/cv/resnet/README_CN.md @@ -103,7 +103,7 @@ ResNet的总体网络架构如下： ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/resnext/README.md b/official/cv/resnext/README.md index 6c5a0985f..d2e356b76 100644 --- a/official/cv/resnext/README.md +++ b/official/cv/resnext/README.md @@ -54,7 +54,7 @@ Dataset used: [imagenet](http://www.image-net.org/) ## [Mixed Precision](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. diff --git a/official/cv/resnext/README_CN.md b/official/cv/resnext/README_CN.md index 09699a0a9..fce417685 100644 --- a/official/cv/resnext/README_CN.md +++ b/official/cv/resnext/README_CN.md @@ -54,7 +54,7 @@ ResNeXt整体网络架构如下： ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 diff --git a/official/cv/retinanet/README_CN.md b/official/cv/retinanet/README_CN.md index 7e6d09379..36b50b7db 100644 --- a/official/cv/retinanet/README_CN.md +++ b/official/cv/retinanet/README_CN.md @@ -189,7 +189,7 @@ bash scripts/run_single_train.sh DEVICE_ID MINDRECORD_DIR PRE_TRAINED(optional) > 注意: - RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html), 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). + RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html), 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). #### 运行 diff --git a/official/cv/semantic_human_matting/README.md b/official/cv/semantic_human_matting/README.md index 68ca05edd..4649c946b 100644 --- a/official/cv/semantic_human_matting/README.md +++ b/official/cv/semantic_human_matting/README.md @@ -78,7 +78,7 @@ ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) 的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索`reduce precision`查看精度降低的算子。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索`reduce precision`查看精度降低的算子。 # 环境要求 diff --git a/official/cv/simple_pose/README.md b/official/cv/simple_pose/README.md index 9a78c5ca0..f22d647e2 100644 --- a/official/cv/simple_pose/README.md +++ b/official/cv/simple_pose/README.md @@ -57,7 +57,7 @@ Dataset used: COCO2017 ## [Mixed Precision](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/official/cv/squeezenet/README.md b/official/cv/squeezenet/README.md index 6b405c2cc..8b4637b8d 100644 --- a/official/cv/squeezenet/README.md +++ b/official/cv/squeezenet/README.md @@ -62,7 +62,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) @@ -687,7 +687,7 @@ Inference result is saved in current path, you can find result like this in acc. ### Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following the steps below, this is a simple example: +If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: - Running on Ascend diff --git a/official/cv/squeezenet/modelarts/README.md b/official/cv/squeezenet/modelarts/README.md index d8136687b..ddb66f9e2 100644 --- a/official/cv/squeezenet/modelarts/README.md +++ b/official/cv/squeezenet/modelarts/README.md @@ -62,7 +62,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) @@ -687,7 +687,7 @@ Inference result is saved in current path, you can find result like this in acc. ### Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following the steps below, this is a simple example: +If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: - Running on Ascend diff --git a/official/cv/srcnn/README_CN.md b/official/cv/srcnn/README_CN.md index 564fa8855..3fb5fd375 100644 --- a/official/cv/srcnn/README_CN.md +++ b/official/cv/srcnn/README_CN.md @@ -71,7 +71,7 @@ SRCNN首先使用双三次(bicubic)插值将低分辨率图像放大成目标尺 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/ssd/README.md b/official/cv/ssd/README.md index 7b4bee4da..3ab719b6b 100644 --- a/official/cv/ssd/README.md +++ b/official/cv/ssd/README.md @@ -306,7 +306,7 @@ Then you can run everything just like on ascend. ### [Training Process](#contents) -To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/docs/programming_guide/en/master/convert_dataset.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** +To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/tutorials/en/master/advanced/dataset/record.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** #### Training on Ascend diff --git a/official/cv/ssd/README_CN.md b/official/cv/ssd/README_CN.md index fdbdd254b..40fed4347 100644 --- a/official/cv/ssd/README_CN.md +++ b/official/cv/ssd/README_CN.md @@ -246,7 +246,7 @@ bash run_eval_gpu.sh [DATASET] [CHECKPOINT_PATH] [DEVICE_ID] [CONFIG_PATH] ## 训练过程 -运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/convert_dataset.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** +运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset/record.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** ### Ascend上训练 diff --git a/official/cv/ssim-ae/README_CN.md b/official/cv/ssim-ae/README_CN.md index 1e954bf4c..6a34cca87 100644 --- a/official/cv/ssim-ae/README_CN.md +++ b/official/cv/ssim-ae/README_CN.md @@ -108,7 +108,7 @@ MVTec AD数据集 ## 混合精度 -采用 [混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) 的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用 [混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/cv/tinydarknet/README_CN.md b/official/cv/tinydarknet/README_CN.md index 12023a750..c648fd1a1 100644 --- a/official/cv/tinydarknet/README_CN.md +++ b/official/cv/tinydarknet/README_CN.md @@ -64,7 +64,7 @@ Tiny-DarkNet是Joseph Chet Redmon等人提出的一个16层的针对于经典的 - + # [环境要求](#目录) diff --git a/official/cv/unet/README.md b/official/cv/unet/README.md index 627c4c714..3093c54e5 100644 --- a/official/cv/unet/README.md +++ b/official/cv/unet/README.md @@ -504,7 +504,7 @@ The above python command will run in the background. You can view the results th ### Inference If you need to use the trained model to perform inference on multiple hardware platforms, such as Ascend 910 or Ascend 310, you -can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following +can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: #### Running on Ascend 310 diff --git a/official/cv/unet/README_CN.md b/official/cv/unet/README_CN.md index fce8e75dd..1bba434d8 100644 --- a/official/cv/unet/README_CN.md +++ b/official/cv/unet/README_CN.md @@ -503,7 +503,7 @@ bash scripts/run_distribute_train_gpu.sh [RANKSIZE] [DATASET] [CONFIG_PATH] #### 推理 -如果您需要使用训练好的模型在Ascend 910、Ascend 310等多个硬件平台上进行推理上进行推理，可参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。下面是一个简单的操作步骤示例： +如果您需要使用训练好的模型在Ascend 910、Ascend 310等多个硬件平台上进行推理上进行推理，可参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。下面是一个简单的操作步骤示例： ##### Ascend 310环境运行 diff --git a/official/cv/unet3d/README.md b/official/cv/unet3d/README.md index 49968f868..ecd8796e5 100644 --- a/official/cv/unet3d/README.md +++ b/official/cv/unet3d/README.md @@ -288,7 +288,7 @@ After training, you'll get some checkpoint files under the `train_parallel_fp[32 #### Distributed training on Ascend > Notes: -> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. +> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. > ```shell diff --git a/official/cv/vgg16/README.md b/official/cv/vgg16/README.md index ea8971d89..e47a112fb 100644 --- a/official/cv/vgg16/README.md +++ b/official/cv/vgg16/README.md @@ -94,7 +94,7 @@ Note that you can run the scripts based on the dataset mentioned in original pap ### Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. @@ -462,7 +462,7 @@ train_parallel1/log:epcoh: 2 step: 97, loss is 1.7133579 ... ``` -> About rank_table.json, you can refer to the [distributed training tutorial](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training.html). +> About rank_table.json, you can refer to the [distributed training tutorial](https://www.mindspore.cn/tutorials/experts/en/master/parallel/introduction.html). > **Attention** This will bind the processor cores according to the `device_num` and total processor numbers. If you don't expect to run pretraining with binding processor cores, remove the operations about `taskset` in `scripts/run_distribute_train.sh` ##### Run vgg16 on GPU diff --git a/official/cv/vgg16/README_CN.md b/official/cv/vgg16/README_CN.md index d1423e1e1..62a469525 100644 --- a/official/cv/vgg16/README_CN.md +++ b/official/cv/vgg16/README_CN.md @@ -95,7 +95,7 @@ VGG 16网络主要由几个基本模块（包括卷积层和池化层）和三 ### 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 @@ -462,7 +462,7 @@ train_parallel1/log:epcoh: 2 step: 97, loss is 1.7133579 ... ``` -> 关于rank_table.json，可以参考[分布式并行训练](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training.html)。 +> 关于rank_table.json，可以参考[分布式并行训练](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/introduction.html)。 > **注意** 将根据`device_num`和处理器总数绑定处理器核。如果您不希望预训练中绑定处理器内核，请在`scripts/run_distribute_train.sh`脚本中移除`taskset`相关操作。 ##### GPU处理器环境运行VGG16 diff --git a/official/cv/vit/README.md b/official/cv/vit/README.md index 7da8ba2bf..0b304c75c 100644 --- a/official/cv/vit/README.md +++ b/official/cv/vit/README.md @@ -65,7 +65,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) @@ -444,7 +444,7 @@ Current batch_ Size can only be set to 1. ### Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following the steps below, this is a simple example: +If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: - Running on Ascend diff --git a/official/cv/vit/README_CN.md b/official/cv/vit/README_CN.md index 12b1b79d1..db969068f 100644 --- a/official/cv/vit/README_CN.md +++ b/official/cv/vit/README_CN.md @@ -68,7 +68,7 @@ Vit是基于多个transformer encoder模块串联起来，由多个inception模 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 @@ -450,7 +450,7 @@ python export.py --config_path=[CONFIG_PATH] ### 推理 -如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。下面是操作步骤示例： +如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。下面是操作步骤示例： - Ascend处理器环境运行 diff --git a/official/cv/warpctc/README.md b/official/cv/warpctc/README.md index d8554d92e..783c2596b 100644 --- a/official/cv/warpctc/README.md +++ b/official/cv/warpctc/README.md @@ -254,7 +254,7 @@ save_checkpoint_path: "./checkpoint" # path to save checkpoint ### [Training Process](#contents) -- Set options in `default_config.yaml`, including learning rate and other network hyperparameters. Click [MindSpore dataset preparation tutorial](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `default_config.yaml`, including learning rate and other network hyperparameters. Click [MindSpore dataset preparation tutorial](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. #### [Training](#contents) diff --git a/official/cv/warpctc/README_CN.md b/official/cv/warpctc/README_CN.md index 4e8750d10..6ead399ac 100644 --- a/official/cv/warpctc/README_CN.md +++ b/official/cv/warpctc/README_CN.md @@ -257,7 +257,7 @@ save_checkpoint_path: "./checkpoints" # 检查点保存路径，相对于t ## 训练过程 -- 在`default_config.yaml`中设置选项，包括学习率和网络超参数。单击[MindSpore加载数据集教程](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/dataset_sample.html)，了解更多信息。 +- 在`default_config.yaml`中设置选项，包括学习率和网络超参数。单击[MindSpore加载数据集教程](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset.html)，了解更多信息。 ### 训练 diff --git a/official/cv/xception/README.md b/official/cv/xception/README.md index 5ae40e616..6dc790198 100644 --- a/official/cv/xception/README.md +++ b/official/cv/xception/README.md @@ -54,7 +54,7 @@ Dataset used can refer to paper. ## [Mixed Precision](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. @@ -193,7 +193,7 @@ bash run_eval_gpu.sh DEVICE_ID DATASET_PATH CHECKPOINT_PATH bash run_infer_310.sh MINDIR_PATH DATA_PATH LABEL_FILE DEVICE_ID ``` -> Notes: RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html), and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). +> Notes: RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html), and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). ### Launch diff --git a/official/cv/yolov3_resnet18/README.md b/official/cv/yolov3_resnet18/README.md index 46281a40f..14e442557 100644 --- a/official/cv/yolov3_resnet18/README.md +++ b/official/cv/yolov3_resnet18/README.md @@ -270,7 +270,7 @@ After installing MindSpore via the official website, you can start training and ### Training on Ascend -To train the model, run `train.py` with the dataset `image_dir`, `anno_path` and `mindrecord_dir`. If the `mindrecord_dir` is empty, it wil generate [mindrecord](https://www.mindspore.cn/docs/programming_guide/en/master/convert_dataset.html) file by `image_dir` and `anno_path`(the absolute image path is joined by the `image_dir` and the relative path in `anno_path`). **Note if `mindrecord_dir` isn't empty, it will use `mindrecord_dir` rather than `image_dir` and `anno_path`.** +To train the model, run `train.py` with the dataset `image_dir`, `anno_path` and `mindrecord_dir`. If the `mindrecord_dir` is empty, it wil generate [mindrecord](https://www.mindspore.cn/tutorials/en/master/advanced/dataset/record.html) file by `image_dir` and `anno_path`(the absolute image path is joined by the `image_dir` and the relative path in `anno_path`). **Note if `mindrecord_dir` isn't empty, it will use `mindrecord_dir` rather than `image_dir` and `anno_path`.** - Stand alone mode @@ -311,7 +311,7 @@ Note the results is two-classification(person and face) used our own annotations ### Evaluation on Ascend -To eval, run `eval.py` with the dataset `image_dir`, `anno_path`(eval txt), `mindrecord_dir` and `ckpt_path`. `ckpt_path` is the path of [checkpoint](https://www.mindspore.cn/docs/programming_guide/en/master/save_model.html) file. +To eval, run `eval.py` with the dataset `image_dir`, `anno_path`(eval txt), `mindrecord_dir` and `ckpt_path`. `ckpt_path` is the path of [checkpoint](https://www.mindspore.cn/tutorials/en/master/advanced/train/save.html) file. ```bash bash run_eval.sh 0 yolo.ckpt ./Mindrecord_eval ./dataset ./dataset/eval.txt diff --git a/official/cv/yolov3_resnet18/README_CN.md b/official/cv/yolov3_resnet18/README_CN.md index 6b0719df8..6dd798f1a 100644 --- a/official/cv/yolov3_resnet18/README_CN.md +++ b/official/cv/yolov3_resnet18/README_CN.md @@ -269,7 +269,7 @@ YOLOv3整体网络架构如下： ### Ascend上训练 -训练模型运行`train.py`，使用数据集`image_dir`、`anno_path`和`mindrecord_dir`。如果`mindrecord_dir`为空，则通过`image_dir`和`anno_path`（图像绝对路径由`image_dir`和`anno_path`中的相对路径连接）生成[MindRecord](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/convert_dataset.html)文件。**注意，如果`mindrecord_dir`不为空，将使用`mindrecord_dir`而不是`image_dir`和`anno_path`。** +训练模型运行`train.py`，使用数据集`image_dir`、`anno_path`和`mindrecord_dir`。如果`mindrecord_dir`为空，则通过`image_dir`和`anno_path`（图像绝对路径由`image_dir`和`anno_path`中的相对路径连接）生成[MindRecord](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset/record.html)文件。**注意，如果`mindrecord_dir`不为空，将使用`mindrecord_dir`而不是`image_dir`和`anno_path`。** - 单机模式 @@ -310,7 +310,7 @@ YOLOv3整体网络架构如下： ### Ascend评估 -运行`eval.py`，数据集为`image_dir`、`anno_path`(评估TXT)、`mindrecord_dir`和`ckpt_path`。`ckpt_path`是[检查点](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/save_model.html)文件的路径。 +运行`eval.py`，数据集为`image_dir`、`anno_path`(评估TXT)、`mindrecord_dir`和`ckpt_path`。`ckpt_path`是[检查点](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/train/save.html)文件的路径。 ```shell script bash run_eval.sh 0 yolo.ckpt ./Mindrecord_eval ./dataset ./dataset/eval.txt diff --git a/official/nlp/bert/README.md b/official/nlp/bert/README.md index e0f4f38e1..72f6bb9d5 100644 --- a/official/nlp/bert/README.md +++ b/official/nlp/bert/README.md @@ -209,8 +209,6 @@ Please follow the instructions in the link below to create an hccl.json file in For distributed training among multiple machines, training command should be executed on each machine in a small time interval. Thus, an hccl.json is needed on each machine. [merge_hccl](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools#merge_hccl) is a tool to create hccl.json for multi-machine case. -For dataset, if you want to set the format and parameters, a schema configuration file with JSON format needs to be created, please refer to [tfrecord](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_loading.html#tfrecord) format. - ```text For pretraining, schema file contains ["input_ids", "input_mask", "segment_ids", "next_sentence_labels", "masked_lm_positions", "masked_lm_ids", "masked_lm_weights"]. diff --git a/official/nlp/cpm/README.md b/official/nlp/cpm/README.md index a309cd6a6..33afd01d0 100644 --- a/official/nlp/cpm/README.md +++ b/official/nlp/cpm/README.md @@ -309,7 +309,7 @@ After processing, the mindrecord file of training and reasoning is generated in ### Finetune Training Process -- Set options in `src/config.py`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `src/config.py`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. - Run `run_distribute_train_ascend_single_machine.sh` for distributed and single machine training of CPM model. diff --git a/official/nlp/cpm/README_CN.md b/official/nlp/cpm/README_CN.md index bfa87f8ad..f6bc6ad1b 100644 --- a/official/nlp/cpm/README_CN.md +++ b/official/nlp/cpm/README_CN.md @@ -309,7 +309,7 @@ Parameters for dataset and network (Training/Evaluation): ### Finetune训练过程 -- 在`src/config.py`中设置，包括模型并行、batchsize、学习率和网络超参数。点击[这里](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/dataset_sample.html)查看更多数据集信息。 +- 在`src/config.py`中设置，包括模型并行、batchsize、学习率和网络超参数。点击[这里](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset.html)查看更多数据集信息。 - 运行`run_distribute_train_ascend_single_machine.sh`，进行CPM模型的单机8卡分布式训练。 diff --git a/official/nlp/duconv/README_CN.md b/official/nlp/duconv/README_CN.md index 95047b1b3..afb773f9c 100644 --- a/official/nlp/duconv/README_CN.md +++ b/official/nlp/duconv/README_CN.md @@ -85,7 +85,7 @@ Proactive Conversation模型包含四个部分： ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/official/nlp/mass/README.md b/official/nlp/mass/README.md index 98e0c045d..d421207e6 100644 --- a/official/nlp/mass/README.md +++ b/official/nlp/mass/README.md @@ -501,7 +501,7 @@ subword-nmt rouge ``` - + # Get started @@ -563,7 +563,7 @@ Get the log and output files under the path `./train_mass_*/`, and the model fil ## Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). +If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). For inference, config the options in `default_config.yaml` firstly: - Assign the `default_config.yaml` under `data_path` node to the dataset path. diff --git a/official/nlp/mass/README_CN.md b/official/nlp/mass/README_CN.md index 3020cf77c..fc8f203e7 100644 --- a/official/nlp/mass/README_CN.md +++ b/official/nlp/mass/README_CN.md @@ -505,7 +505,7 @@ subword-nmt rouge ``` - + # 快速上手 @@ -567,7 +567,7 @@ bash run_gpu.sh -t t -n 1 -i 1 ## 推理 -如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。 +如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。 推理时，请先配置`config.json`中的选项： - 将`default_config.yaml`节点下的`data_path`配置为数据集路径。 diff --git a/official/nlp/pangu_alpha/README.md b/official/nlp/pangu_alpha/README.md index d885a1fea..bed9c6d0e 100644 --- a/official/nlp/pangu_alpha/README.md +++ b/official/nlp/pangu_alpha/README.md @@ -1,4 +1,4 @@ -# Contents +# Contents - [Contents](#contents) - [PanGu-Alpha Description](#pangu-alpha-description) @@ -45,7 +45,7 @@ with our parallel setting. We summarized the training tricks as followings: 2. Pipeline Model Parallelism 3. Optimizer Model Parallelism -The above features can be found [here](https://www.mindspore.cn/docs/programming_guide/en/master/auto_parallel.html). +The above features can be found [here](https://www.mindspore.cn/tutorials/experts/en/master/parallel/introduction.html). More amazing features are still under developing. The technical report and checkpoint file can be found [here](https://git.openi.org.cn/PCL-Platform.Intelligence/PanGu-AIpha). @@ -151,7 +151,7 @@ bash scripts/run_distribute_train.sh /data/pangu_30_step_ba64/ /root/hccl_8p.jso The above command involves some `args` described below: - DATASET: The path to the mindrecord files's parent directory . For example: `/home/work/mindrecord/`. -- RANK_TABLE: The details of the rank table can be found [here](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html). It's a json file describes the `device id`, `service ip` and `rank`. +- RANK_TABLE: The details of the rank table can be found [here](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html). It's a json file describes the `device id`, `service ip` and `rank`. - RANK_SIZE: The device number. This can be your total device numbers. For example, 8, 16, 32 ... - TYPE: The param init type. The parameters will be initialized with float32. Or you can replace it with `fp16`. This will save a little memory used on the device. - MODE: The configure mode. This mode will set the `hidden size` and `layers` to make the parameter number near 2.6 billions. The other mode can be `13B` (`hidden size` 5120 and `layers` 40, which needs at least 16 cards to train.) and `200B`. @@ -189,7 +189,7 @@ device0/log0.log). The script will launch the GPU training through `mpirun`, the user can run the following command on any machine to start training. Note when start training multi-node, the variables `NCCL_SOCKET_IFNAME` `NCCL_IB_HCA` may be different on some servers. If you meet some errors and -strange phenomenon, please unset or set the NCCL variables. Details can be checked on this [link](https://www.mindspore.cn/docs/faq/zh-CN/master/distributed_configure.html). +strange phenomenon, please unset or set the NCCL variables. Details can be checked on this [link](https://www.mindspore.cn/docs/zh-CN/master/faq/distributed_configure.html). ```bash # The following variables are optional. @@ -200,7 +200,7 @@ bash scripts/run_distributed_train_gpu.sh RANK_SIZE HOSTFILE DATASET PER_BATCH M ``` - RANK_SIZE: The device number. This can be your total device numbers. For example, 8, 16, 32 ... -- HOSTFILE: It's a text file describes the host ip and its devices. Please see our [tutorial](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_gpu.html) or [OpenMPI](https://www.open-mpi.org/) for more details. +- HOSTFILE: It's a text file describes the host ip and its devices. Please see our [tutorial](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_gpu.html) or [OpenMPI](https://www.open-mpi.org/) for more details. - DATASET: The path to the mindrecord files's parent directory . For example: `/home/work/mindrecord/`. - PER_BATCH: The batch size for each data parallel-way. - MODE: Can be `1.3B` `2.6B`, `13B` and `200B`. @@ -222,7 +222,7 @@ bash scripts/run_distribute_train_moe_host_device.sh DATASET RANK_TABLE RANK_SIZ The above command involves some `args` described below: - DATASET: The path to the mindrecord files's parent directory . For example: `/home/work/mindrecord/`. -- RANK_TABLE: The details of the rank table can be found [here](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html). It's a json file describes the `device id`, `service ip` and `rank`. +- RANK_TABLE: The details of the rank table can be found [here](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html). It's a json file describes the `device id`, `service ip` and `rank`. - RANK_SIZE: The device number. This can be your total device numbers. For example, 8, 16, 32 ... - TYPE: The param init type. The parameters will be initialized with float32. Or you can replace it with `fp16`. This will save a little memory used on the device. - MODE: The configure mode. This mode will set the `hidden size` and `layers` to make the parameter number near 2.6 billions. The other mode can be `13B` (`hidden size` 5120 and `layers` 40, which needs at least 16 cards to train.) and `200B`. diff --git a/official/nlp/transformer/README.md b/official/nlp/transformer/README.md index 3e35c3784..4fec4896d 100644 --- a/official/nlp/transformer/README.md +++ b/official/nlp/transformer/README.md @@ -342,7 +342,7 @@ Parameters for learning rate: ## [Training Process](#contents) -- Set options in `default_config.yaml`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `default_config.yaml`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. - Run `run_standalone_train.sh` for non-distributed training of Transformer model. diff --git a/official/nlp/transformer/README_CN.md b/official/nlp/transformer/README_CN.md index be21a0a8c..913aafe56 100644 --- a/official/nlp/transformer/README_CN.md +++ b/official/nlp/transformer/README_CN.md @@ -341,7 +341,7 @@ Parameters for learning rate: ### 训练过程 -- 在`default_config.yaml`中设置选项，包括loss_scale、学习率和网络超参数。点击[这里](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/dataset_sample.html)查看更多数据集信息。 +- 在`default_config.yaml`中设置选项，包括loss_scale、学习率和网络超参数。点击[这里](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset.html)查看更多数据集信息。 - 运行`run_standalone_train.sh`，进行Transformer模型的非分布式训练。 diff --git a/official/recommend/ncf/README.md b/official/recommend/ncf/README.md index f12d20935..72b829054 100644 --- a/official/recommend/ncf/README.md +++ b/official/recommend/ncf/README.md @@ -73,7 +73,7 @@ In both datasets, the timestamp is represented in seconds since midnight Coordin ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) @@ -335,9 +335,9 @@ Inference result is saved in current path, you can find result like this in acc. ### Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following the steps below, this is a simple example: +If you need to use the trained model to perform inference on multiple hardware platforms, such as Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: - + ```python # Load unseen dataset for inference diff --git a/research/audio/fcn-4/README.md b/research/audio/fcn-4/README.md index 663a5e305..34e07d0c6 100644 --- a/research/audio/fcn-4/README.md +++ b/research/audio/fcn-4/README.md @@ -41,7 +41,7 @@ FCN-4 is a convolutional neural network architecture, its name FCN-4 comes from ### Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. ## [Environment Requirements](#contents) diff --git a/research/audio/speech_transformer/README.md b/research/audio/speech_transformer/README.md index 246ff4370..e665fba09 100644 --- a/research/audio/speech_transformer/README.md +++ b/research/audio/speech_transformer/README.md @@ -187,7 +187,7 @@ Dataset is preprocessed using `Kaldi` and converts kaldi binaries into Python pi ## [Training Process](#contents) -- Set options in `default_config.yaml`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `default_config.yaml`, including loss_scale, learning rate and network hyperparameters. Click [here](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. - Run `run_standalone_train_gpu.sh` for non-distributed training of Transformer model. diff --git a/research/cv/3D_DenseNet/README.md b/research/cv/3D_DenseNet/README.md index e3ad419a9..68a648a46 100644 --- a/research/cv/3D_DenseNet/README.md +++ b/research/cv/3D_DenseNet/README.md @@ -222,7 +222,7 @@ Dice Coefficient (DC) for 9th subject (9 subjects for training and 1 subject for |-------------------|:-------------------:|:---------------------:|:-----:|:--------------:| |3D-SkipDenseSeg | 93.66| 90.80 | 90.65 | 91.70 | -Notes: RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools) For large models like InceptionV4, it's better to export an external environment variable export HCCL_CONNECT_TIMEOUT=600 to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. To avoid ops error，you should change the code like below: +Notes: RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools) For large models like InceptionV4, it's better to export an external environment variable export HCCL_CONNECT_TIMEOUT=600 to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. To avoid ops error，you should change the code like below: in train.py： diff --git a/research/cv/3D_DenseNet/README_CN.md b/research/cv/3D_DenseNet/README_CN.md index 2f81477c8..e9d5bd111 100644 --- a/research/cv/3D_DenseNet/README_CN.md +++ b/research/cv/3D_DenseNet/README_CN.md @@ -1,5 +1,3 @@ - - # 目录 [View English](./README.md) @@ -214,7 +212,7 @@ bash run_eval.sh 3D-DenseSeg-20000_36.ckpt data/data_val |-------------------|:-------------------:|:---------------------:|:-----:|:--------------:| |3D-SkipDenseSeg | 93.66| 90.80 | 90.65 | 91.70 | -Notes: 分布式训练需要一个RANK_TABLE_FILE，文件的删除方式可以参考该链接[Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) ,device_ip的设置参考该链接 [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools) 对于像InceptionV4这样的大模型来说, 最好导出一个外部环境变量，export HCCL_CONNECT_TIMEOUT=600，以将hccl连接检查时间从默认的120秒延长到600秒。否则，连接可能会超时，因为编译时间会随着模型大小的增加而增加。在1.3.0版本下，3D算子可能存在一些问题，您可能需要更改context.set_auto_parallel_context的部分代码: +Notes: 分布式训练需要一个RANK_TABLE_FILE，文件的删除方式可以参考该链接[Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) ,device_ip的设置参考该链接 [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools) 对于像InceptionV4这样的大模型来说, 最好导出一个外部环境变量，export HCCL_CONNECT_TIMEOUT=600，以将hccl连接检查时间从默认的120秒延长到600秒。否则，连接可能会超时，因为编译时间会随着模型大小的增加而增加。在1.3.0版本下，3D算子可能存在一些问题，您可能需要更改context.set_auto_parallel_context的部分代码: in train.py： diff --git a/research/cv/APDrawingGAN/README_CN.md b/research/cv/APDrawingGAN/README_CN.md index 292a446ab..15d62ce12 100644 --- a/research/cv/APDrawingGAN/README_CN.md +++ b/research/cv/APDrawingGAN/README_CN.md @@ -86,7 +86,7 @@ auxiliary.ckpt文件获取：从 https://cg.cs.tsinghua.edu.cn/people/~Yongjin/A ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/AlignedReID++/README_CN.md b/research/cv/AlignedReID++/README_CN.md index b44ae7167..53bdce3e0 100644 --- a/research/cv/AlignedReID++/README_CN.md +++ b/research/cv/AlignedReID++/README_CN.md @@ -61,7 +61,7 @@ AlignedReID++采用resnet50作为backbone，重新命名了AlignedReID中提出 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 @@ -403,7 +403,7 @@ market1501上评估AlignedReID++ ### 推理 -如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。下面是操作步骤示例： +如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。下面是操作步骤示例： 在进行推理之前我们需要先导出模型，mindir可以在本地环境上导出。batch_size默认为1。 diff --git a/research/cv/AlphaPose/README_CN.md b/research/cv/AlphaPose/README_CN.md index eb2809997..39c521465 100644 --- a/research/cv/AlphaPose/README_CN.md +++ b/research/cv/AlphaPose/README_CN.md @@ -55,7 +55,7 @@ AlphaPose的总体网络架构如下： ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/DDRNet/README_CN.md b/research/cv/DDRNet/README_CN.md index e0ef16c49..0723bcef7 100644 --- a/research/cv/DDRNet/README_CN.md +++ b/research/cv/DDRNet/README_CN.md @@ -53,7 +53,7 @@ ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # [环境要求](#目录) diff --git a/research/cv/EDSR/README_CN.md b/research/cv/EDSR/README_CN.md index 1af53d936..cec64c247 100644 --- a/research/cv/EDSR/README_CN.md +++ b/research/cv/EDSR/README_CN.md @@ -97,7 +97,7 @@ EDSR是由多个优化后的residual blocks串联而成，相比原始版本的r ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html?highlight=%E6%B7%B7%E5%90%88%E7%B2%BE%E5%BA%A6)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html?highlight=%E6%B7%B7%E5%90%88%E7%B2%BE%E5%BA%A6)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/EGnet/README_CN.md b/research/cv/EGnet/README_CN.md index e945d0a66..8f9c9f0e5 100644 --- a/research/cv/EGnet/README_CN.md +++ b/research/cv/EGnet/README_CN.md @@ -359,7 +359,7 @@ bash run_standalone_train_gpu.sh bash run_distribute_train.sh 8 [RANK_TABLE_FILE] ``` -线下运行分布式训练请参照[mindspore分布式并行训练基础样例（Ascend）](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html) +线下运行分布式训练请参照[mindspore分布式并行训练基础样例（Ascend）](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html) - 线上modelarts分布式训练 diff --git a/research/cv/GENet_Res50/README_CN.md b/research/cv/GENet_Res50/README_CN.md index 726cf0d2f..0a7736873 100644 --- a/research/cv/GENet_Res50/README_CN.md +++ b/research/cv/GENet_Res50/README_CN.md @@ -64,7 +64,7 @@ Imagenet 2017和Imagenet 2012 数据集一致 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/LightCNN/README.md b/research/cv/LightCNN/README.md index c2d524a5a..21f59b7fb 100644 --- a/research/cv/LightCNN/README.md +++ b/research/cv/LightCNN/README.md @@ -119,7 +119,7 @@ Dataset structure: ## [Mixed Precision](#mixedprecision) -The [mixed-precision](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) training +The [mixed-precision](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) training method uses single-precision and half-precision data to improve the training speed of deep learning neural networks, while maintaining the network accuracy that can be achieved by single-precision training. Mixed-precision training increases computing speed and reduces memory usage, while supporting training larger models or achieving larger batches @@ -139,7 +139,7 @@ reduce precision" to view the operators with reduced precision. - Generate config json file for 8-card training - [Simple tutorial](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools) - For detailed configuration method, please refer to - the [official website tutorial](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html#configuring-distributed-environment-variables). + the [official website tutorial](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html#configuring-distributed-environment-variables). # [Quick start](#Quickstart) @@ -637,7 +637,7 @@ Please check the official [homepage](https://gitee.com/mindspore/models). [5]: https://pan.baidu.com/s/1eR6vHFO -[6]: https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html +[6]: https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html [7]: http://www.cbsr.ia.ac.cn/users/scliao/projects/blufr/BLUFR.zip diff --git a/research/cv/LightCNN/README_CN.md b/research/cv/LightCNN/README_CN.md index 97e91e010..4866de2f2 100644 --- a/research/cv/LightCNN/README_CN.md +++ b/research/cv/LightCNN/README_CN.md @@ -107,7 +107,7 @@ LightCNN适用于有大量噪声的人脸识别数据集，提出了maxout 的 - [MindSpore Python API](https://www.mindspore.cn/docs/api/zh-CN/master/index.html) - 生成config json文件用于8卡训练。 - [简易教程](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools) - - 详细配置方法请参照[官网教程](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html#配置分布式环境变量)。 + - 详细配置方法请参照[官网教程](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html#配置分布式环境变量)。 # 快速入门 @@ -516,7 +516,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [DEVICE_ID] [3]: https://drive.google.com/file/d/0ByNaVHFekDPRbFg1YTNiMUxNYXc/view?usp=sharing [4]: https://hyper.ai/datasets/5543 [5]: https://pan.baidu.com/s/1eR6vHFO -[6]: https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html +[6]: https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html [7]: http://www.cbsr.ia.ac.cn/users/scliao/projects/blufr/BLUFR.zip [8]: https://github.com/AlfredXiangWu/face_verification_experiment/blob/master/code/lfw_pairs.mat [9]: https://github.com/AlfredXiangWu/face_verification_experiment/blob/master/results/LightenedCNN_B_lfw.mat diff --git a/research/cv/ManiDP/Readme.md b/research/cv/ManiDP/Readme.md index 2f4302712..403094c0f 100644 --- a/research/cv/ManiDP/Readme.md +++ b/research/cv/ManiDP/Readme.md @@ -40,7 +40,7 @@ Dataset used: [CIFAR-10](https://www.cs.toronto.edu/~kriz/cifar.html) ## [Mixed Precision(Ascend)](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/research/cv/NFNet/README_CN.md b/research/cv/NFNet/README_CN.md index d46125b2b..ee4fdad5e 100644 --- a/research/cv/NFNet/README_CN.md +++ b/research/cv/NFNet/README_CN.md @@ -57,7 +57,7 @@ ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # [环境要求](#目录) diff --git a/research/cv/RefineDet/README_CN.md b/research/cv/RefineDet/README_CN.md index 3645326d5..92353c907 100644 --- a/research/cv/RefineDet/README_CN.md +++ b/research/cv/RefineDet/README_CN.md @@ -211,7 +211,7 @@ sh run_eval_gpu.sh [DATASET] [CHECKPOINT_PATH] [DEVICE_ID] ## 训练过程 -运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/convert_dataset.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** +运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset/record.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** ### Ascend上训练 diff --git a/research/cv/RefineNet/README.md b/research/cv/RefineNet/README.md index 413b7e363..fb8c1e4db 100644 --- a/research/cv/RefineNet/README.md +++ b/research/cv/RefineNet/README.md @@ -84,7 +84,7 @@ Pascal VOC数据集和语义边界数据集（Semantic Boundaries Dataset，SBD ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 diff --git a/research/cv/SE-Net/README.md b/research/cv/SE-Net/README.md index dd8993efe..6c981e56c 100644 --- a/research/cv/SE-Net/README.md +++ b/research/cv/SE-Net/README.md @@ -67,7 +67,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/research/cv/SE_ResNeXt50/README_CN.md b/research/cv/SE_ResNeXt50/README_CN.md index e4d3136ab..e9e54a23c 100644 --- a/research/cv/SE_ResNeXt50/README_CN.md +++ b/research/cv/SE_ResNeXt50/README_CN.md @@ -56,7 +56,7 @@ SE-ResNeXt的总体网络架构如下： [链接](https://arxiv.org/abs/1709.015 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # 环境要求 diff --git a/research/cv/TNT/README_CN.md b/research/cv/TNT/README_CN.md index bf21f0efc..cf8699f34 100644 --- a/research/cv/TNT/README_CN.md +++ b/research/cv/TNT/README_CN.md @@ -53,7 +53,7 @@ Transformer是一种最初用于NLP任务的基于自注意力的神经网络。 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # [环境要求](#目录) diff --git a/research/cv/cct/README_CN.md b/research/cv/cct/README_CN.md index f67e02896..b61064ab0 100644 --- a/research/cv/cct/README_CN.md +++ b/research/cv/cct/README_CN.md @@ -51,7 +51,7 @@ ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # [环境要求](#目录) diff --git a/research/cv/convnext/README_CN.md b/research/cv/convnext/README_CN.md index eec99f773..09a902628 100644 --- a/research/cv/convnext/README_CN.md +++ b/research/cv/convnext/README_CN.md @@ -53,7 +53,7 @@ ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # [环境要求](#目录) diff --git a/research/cv/dcgan/README.md b/research/cv/dcgan/README.md index 5fd8c0e66..cca854467 100644 --- a/research/cv/dcgan/README.md +++ b/research/cv/dcgan/README.md @@ -137,7 +137,7 @@ dcgan_cifar10_cfg { ## [Training Process](#contents) -- Set options in `config.py`, including learning rate, output filename and network hyperparameters. Click [here](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `config.py`, including learning rate, output filename and network hyperparameters. Click [here](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. ### [Training](#content) diff --git a/research/cv/deeplabv3plus/README_CN.md b/research/cv/deeplabv3plus/README_CN.md index 38a80416f..a404b5c28 100644 --- a/research/cv/deeplabv3plus/README_CN.md +++ b/research/cv/deeplabv3plus/README_CN.md @@ -85,7 +85,7 @@ Pascal VOC数据集和语义边界数据集（Semantic Boundaries Dataset，SBD ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/dlinknet/README.md b/research/cv/dlinknet/README.md index 06cf9bded..d272c06d9 100644 --- a/research/cv/dlinknet/README.md +++ b/research/cv/dlinknet/README.md @@ -316,7 +316,7 @@ bash scripts/run_distribute_train.sh [RANK_TABLE_FILE] [DATASET] [CONFIG_PATH] #### inference If you need to use the trained model to perform inference on multiple hardware platforms, such as Ascend 910 or Ascend 310, you -can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following +can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: ##### running-on-ascend-310 diff --git a/research/cv/dlinknet/README_CN.md b/research/cv/dlinknet/README_CN.md index 2cdb3ed0c..2e43b9cb7 100644 --- a/research/cv/dlinknet/README_CN.md +++ b/research/cv/dlinknet/README_CN.md @@ -320,7 +320,7 @@ bash scripts/run_distribute_train.sh [RANK_TABLE_FILE] [DATASET] [CONFIG_PATH] #### 推理 -如果您需要使用训练好的模型在Ascend 910、Ascend 310等多个硬件平台上进行推理上进行推理，可参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。下面是一个简单的操作步骤示例： +如果您需要使用训练好的模型在Ascend 910、Ascend 310等多个硬件平台上进行推理上进行推理，可参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。下面是一个简单的操作步骤示例： ##### Ascend 310环境运行 diff --git a/research/cv/efficientnetv2/README_CN.md b/research/cv/efficientnetv2/README_CN.md index 9e90c4a99..75dea2a67 100644 --- a/research/cv/efficientnetv2/README_CN.md +++ b/research/cv/efficientnetv2/README_CN.md @@ -51,7 +51,7 @@ ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # [环境要求](#目录) diff --git a/research/cv/fairmot/README.md b/research/cv/fairmot/README.md index ff3f565fa..c75f79a65 100644 --- a/research/cv/fairmot/README.md +++ b/research/cv/fairmot/README.md @@ -46,7 +46,7 @@ Dataset used: ETH, CalTech, MOT17, CUHK-SYSU, PRW, CityPerson ## [Mixed Precision](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/research/cv/fishnet99/README_CN.md b/research/cv/fishnet99/README_CN.md index 7129785a4..86aae8c34 100644 --- a/research/cv/fishnet99/README_CN.md +++ b/research/cv/fishnet99/README_CN.md @@ -63,7 +63,7 @@ ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # 环境要求 diff --git a/research/cv/glore_res/README_CN.md b/research/cv/glore_res/README_CN.md index ead07cb53..4a7afb8ba 100644 --- a/research/cv/glore_res/README_CN.md +++ b/research/cv/glore_res/README_CN.md @@ -81,7 +81,7 @@ glore_res200网络模型的backbone是ResNet200, 在Stage2, Stage3中分别均 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/glore_res200/README_CN.md b/research/cv/glore_res200/README_CN.md index 6c81a1be6..cd0bf75fa 100644 --- a/research/cv/glore_res200/README_CN.md +++ b/research/cv/glore_res200/README_CN.md @@ -72,7 +72,7 @@ ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/glore_res50/README.md b/research/cv/glore_res50/README.md index 39e47cae9..bc80ce7d1 100644 --- a/research/cv/glore_res50/README.md +++ b/research/cv/glore_res50/README.md @@ -61,7 +61,7 @@ glore_res的总体网络架构如下： ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/hardnet/README_CN.md b/research/cv/hardnet/README_CN.md index d1b901770..7b44eef55 100644 --- a/research/cv/hardnet/README_CN.md +++ b/research/cv/hardnet/README_CN.md @@ -60,7 +60,7 @@ HarDNet指的是Harmonic DenseNet: A low memory traffic network，其突出的 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 @@ -419,7 +419,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [DEVICE_ID] ### 推理 -如果您需要使用此训练模型在Ascend 910上进行推理，可参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。下面是操作步骤示例： +如果您需要使用此训练模型在Ascend 910上进行推理，可参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。下面是操作步骤示例： - Ascend处理器环境运行 @@ -456,7 +456,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATASET_PATH] [DEVICE_ID] print("==============Acc: {} ==============".format(acc)) ``` -如果您需要使用此训练模型在GPU上进行推理，可参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。下面是操作步骤示例： +如果您需要使用此训练模型在GPU上进行推理，可参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。下面是操作步骤示例： - GPU处理器环境运行 diff --git a/research/cv/inception_resnet_v2/README.md b/research/cv/inception_resnet_v2/README.md index cf199c606..3852562de 100644 --- a/research/cv/inception_resnet_v2/README.md +++ b/research/cv/inception_resnet_v2/README.md @@ -44,7 +44,7 @@ The dataset used is [ImageNet](https://image-net.org/download.php). ## [Mixed Precision(Ascend)](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. @@ -122,7 +122,7 @@ bash scripts/run_standalone_train_ascend.sh DEVICE_ID DATA_DIR ``` > Notes: -> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. +> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. > > This is processor cores binding operation regarding the `device_num` and total processor numbers. If you are not expect to do it, remove the operations `taskset` in `scripts/run_distribute_train.sh` diff --git a/research/cv/inception_resnet_v2/README_CN.md b/research/cv/inception_resnet_v2/README_CN.md index 9fadab9ea..8be29bd5c 100644 --- a/research/cv/inception_resnet_v2/README_CN.md +++ b/research/cv/inception_resnet_v2/README_CN.md @@ -56,7 +56,7 @@ Inception_ResNet_v2的总体网络架构如下： ## 混合精度（Ascend） -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 @@ -133,7 +133,7 @@ bash scripts/run_distribute_train_ascend.sh RANK_TABLE_FILE DATA_DIR bash scripts/run_standalone_train_ascend.sh DEVICE_ID DATA_DIR ``` -> 注：RANK_TABLE_FILE可参考[链接]( https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html)。device_ip可以通过[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools)获取 +> 注：RANK_TABLE_FILE可参考[链接]( https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html)。device_ip可以通过[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools)获取 - GPU: diff --git a/research/cv/mae/README_CN.md b/research/cv/mae/README_CN.md index 5c8f9a266..ee67f0afa 100644 --- a/research/cv/mae/README_CN.md +++ b/research/cv/mae/README_CN.md @@ -63,7 +63,7 @@ This is a MindSpore/NPU re-implementation of the paper [Masked Autoencoders Are ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 @@ -390,7 +390,7 @@ This is a MindSpore/NPU re-implementation of the paper [Masked Autoencoders Are ### 推理 -如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/multi_platform_inference.html)。下面是操作步骤示例： +如果您需要使用此训练模型在GPU、Ascend 910、Ascend 310等多个硬件平台上进行推理，可参考此[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/infer/inference.html)。下面是操作步骤示例： - Ascend处理器环境运行 diff --git a/research/cv/metric_learn/README_CN.md b/research/cv/metric_learn/README_CN.md index 6c95794fd..1588e0afa 100644 --- a/research/cv/metric_learn/README_CN.md +++ b/research/cv/metric_learn/README_CN.md @@ -80,7 +80,7 @@ cd Stanford_Online_Products && head -n 1048 test.txt > test_tiny.txt ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/midas/README.md b/research/cv/midas/README.md index 4ce4c8ffd..b55353383 100644 --- a/research/cv/midas/README.md +++ b/research/cv/midas/README.md @@ -55,7 +55,7 @@ Midas的总体网络架构如下： ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/nas-fpn/README_CN.md b/research/cv/nas-fpn/README_CN.md index 1fd7d580f..8706b3600 100644 --- a/research/cv/nas-fpn/README_CN.md +++ b/research/cv/nas-fpn/README_CN.md @@ -161,7 +161,7 @@ bash scripts/run_single_train.sh DEVICE_ID MINDRECORD_DIR PRE_TRAINED(optional) ``` > 注意: -RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html), 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). +RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html), 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). #### 运行 diff --git a/research/cv/ntsnet/README.md b/research/cv/ntsnet/README.md index 20a42e7c3..f53854d33 100644 --- a/research/cv/ntsnet/README.md +++ b/research/cv/ntsnet/README.md @@ -133,7 +133,7 @@ Usage: bash run_standalone_train_ascend.sh [DATA_URL] [TRAIN_URL] ## [Training Process](#contents) -- Set options in `config.py`, including learning rate, output filename and network hyperparameters. Click [here](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_sample.html) for more information about dataset. +- Set options in `config.py`, including learning rate, output filename and network hyperparameters. Click [here](https://www.mindspore.cn/tutorials/en/master/advanced/dataset.html) for more information about dataset. - Get ResNet50 pretrained model from [Mindspore Hub](https://www.mindspore.cn/resources/hub/details?MindSpore/ascend/v1.2/resnet50_v1.2_imagenet2012) ### [Training](#content) diff --git a/research/cv/osnet/README.md b/research/cv/osnet/README.md index 15c7ea6a6..f449b8b59 100644 --- a/research/cv/osnet/README.md +++ b/research/cv/osnet/README.md @@ -155,7 +155,7 @@ bash run_eval_ascend.sh [DATASET] [CHECKPOINT_PATH] [DEVICE_ID] ``` > Notes: -> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. +> RANK_TABLE_FILE can refer to [Link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html) , and the device_ip can be got as [Link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). For large models like InceptionV4, it's better to export an external environment variable `export HCCL_CONNECT_TIMEOUT=600` to extend hccl connection checking time from the default 120 seconds to 600 seconds. Otherwise, the connection could be timeout since compiling time increases with the growth of model size. > > This is processor cores binding operation regarding the `device_num` and total processor numbers. If you are not expect to do it, remove the operations `taskset` in `scripts/run_train_distribute_ascend.sh` > diff --git a/research/cv/ras/README.md b/research/cv/ras/README.md index c2a18eb1b..1dc300f6e 100644 --- a/research/cv/ras/README.md +++ b/research/cv/ras/README.md @@ -73,7 +73,7 @@ RAS总体网络架构如下: ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) 的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/renas/Readme.md b/research/cv/renas/Readme.md index 3a862f987..f76c2c8ed 100644 --- a/research/cv/renas/Readme.md +++ b/research/cv/renas/Readme.md @@ -39,7 +39,7 @@ An effective and efficient architecture performance evaluation scheme is essenti ## [Mixed Precision(Ascend)](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/research/cv/res2net/README.md b/research/cv/res2net/README.md index d971e0ee4..199f6fa24 100644 --- a/research/cv/res2net/README.md +++ b/research/cv/res2net/README.md @@ -82,7 +82,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/research/cv/res2net_deeplabv3/README.md b/research/cv/res2net_deeplabv3/README.md index 4632c1d4f..478034d00 100644 --- a/research/cv/res2net_deeplabv3/README.md +++ b/research/cv/res2net_deeplabv3/README.md @@ -85,7 +85,7 @@ You can also generate the list file automatically by run script: `python get_dat ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/research/cv/resnet3d/README_CN.md b/research/cv/resnet3d/README_CN.md index 3410ec5ca..5ed6d25aa 100644 --- a/research/cv/resnet3d/README_CN.md +++ b/research/cv/resnet3d/README_CN.md @@ -105,7 +105,7 @@ python3 generate_video_jpgs.py --video_path ~/dataset/hmdb51/videos/ --target_pa ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/resnet50_bam/README.md b/research/cv/resnet50_bam/README.md index 170f2124c..9367f89f3 100644 --- a/research/cv/resnet50_bam/README.md +++ b/research/cv/resnet50_bam/README.md @@ -56,7 +56,7 @@ Data set used: [ImageNet2012](http://www.image-net.org/) ## Mixed precision -The [mixed-precision](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) training method uses single-precision and half-precision data to improve the training speed of deep learning neural networks, while maintaining the network accuracy that can be achieved by single-precision training. Mixed-precision training increases computing speed and reduces memory usage, while supporting training larger models or achieving larger batches of training on specific hardware. +The [mixed-precision](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) training method uses single-precision and half-precision data to improve the training speed of deep learning neural networks, while maintaining the network accuracy that can be achieved by single-precision training. Mixed-precision training increases computing speed and reduces memory usage, while supporting training larger models or achieving larger batches of training on specific hardware. # Environmental requirements diff --git a/research/cv/resnet50_bam/README_CN.md b/research/cv/resnet50_bam/README_CN.md index 5b7ea5b26..d4a8c28f6 100644 --- a/research/cv/resnet50_bam/README_CN.md +++ b/research/cv/resnet50_bam/README_CN.md @@ -56,7 +56,7 @@ resnet50_bam的作者提出了一个简单但是有效的Attention模型——BA ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # 环境要求 diff --git a/research/cv/resnext152_64x4d/README.md b/research/cv/resnext152_64x4d/README.md index 3320bcf35..61fb3a324 100644 --- a/research/cv/resnext152_64x4d/README.md +++ b/research/cv/resnext152_64x4d/README.md @@ -54,7 +54,7 @@ Dataset used: [imagenet](http://www.image-net.org/) ## [Mixed Precision](#contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. diff --git a/research/cv/resnext152_64x4d/README_CN.md b/research/cv/resnext152_64x4d/README_CN.md index 8b6a05b0e..2a6e58096 100644 --- a/research/cv/resnext152_64x4d/README_CN.md +++ b/research/cv/resnext152_64x4d/README_CN.md @@ -54,7 +54,7 @@ ResNeXt整体网络架构如下： ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 diff --git a/research/cv/retinanet_resnet101/README.md b/research/cv/retinanet_resnet101/README.md index d20bef0b3..c2e896644 100644 --- a/research/cv/retinanet_resnet101/README.md +++ b/research/cv/retinanet_resnet101/README.md @@ -287,7 +287,7 @@ bash run_distribute_train.sh [DEVICE_NUM] [EPOCH_SIZE] [LR] [DATASET] [RANK_TABL bash run_single_train.sh [DEVICE_ID] [EPOCH_SIZE] [LR] [DATASET] [PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional) ``` -> Note: RANK_TABLE_FILE related reference materials see in this [link](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_ascend.html), for details on how to get device_ip check this [link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). +> Note: RANK_TABLE_FILE related reference materials see in this [link](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_ascend.html), for details on how to get device_ip check this [link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). - GPU diff --git a/research/cv/retinanet_resnet101/README_CN.md b/research/cv/retinanet_resnet101/README_CN.md index 8f86a05d8..d62df255a 100644 --- a/research/cv/retinanet_resnet101/README_CN.md +++ b/research/cv/retinanet_resnet101/README_CN.md @@ -292,7 +292,7 @@ bash run_distribute_train.sh [DEVICE_NUM] [EPOCH_SIZE] [LR] [DATASET] [RANK_TABL bash run_single_train.sh [DEVICE_ID] [EPOCH_SIZE] [LR] [DATASET] [PRE_TRAINED](optional) [PRE_TRAINED_EPOCH_SIZE](optional) ``` -> 注意: RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html), 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). +> 注意: RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html), 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). - GPU diff --git a/research/cv/retinanet_resnet152/README.md b/research/cv/retinanet_resnet152/README.md index 23e04a27d..d1be441cb 100644 --- a/research/cv/retinanet_resnet152/README.md +++ b/research/cv/retinanet_resnet152/README.md @@ -291,7 +291,7 @@ bash run_distribute_train.sh DEVICE_NUM EPOCH_SIZE LR DATASET RANK_TABLE_FILE PR bash run_distribute_train.sh DEVICE_ID EPOCH_SIZE LR DATASET PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional) ``` -> Note: RANK_TABLE_FILE related reference materials see in this [link](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html), +> Note: RANK_TABLE_FILE related reference materials see in this [link](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html), > for details on how to get device_ip check this [link](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). - GPU: diff --git a/research/cv/retinanet_resnet152/README_CN.md b/research/cv/retinanet_resnet152/README_CN.md index 1dda1c52d..f3c709496 100644 --- a/research/cv/retinanet_resnet152/README_CN.md +++ b/research/cv/retinanet_resnet152/README_CN.md @@ -285,7 +285,7 @@ bash run_distribute_train.sh DEVICE_NUM EPOCH_SIZE LR DATASET RANK_TABLE_FILE PR bash run_distribute_train.sh DEVICE_ID EPOCH_SIZE LR DATASET PRE_TRAINED(optional) PRE_TRAINED_EPOCH_SIZE(optional) ``` -> 注意: RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html), +> 注意: RANK_TABLE_FILE相关参考资料见[链接](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html), > 获取device_ip方法详见[链接](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). - GPU: diff --git a/research/cv/siamRPN/README_CN.md b/research/cv/siamRPN/README_CN.md index d7937fa68..3376c0cd3 100644 --- a/research/cv/siamRPN/README_CN.md +++ b/research/cv/siamRPN/README_CN.md @@ -51,7 +51,7 @@ Siam-RPN提出了一种基于RPN的孪生网络结构。由孪生子网络和RPN ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/simple_baselines/README_CN.md b/research/cv/simple_baselines/README_CN.md index 1a228c7e7..3eb48ea04 100644 --- a/research/cv/simple_baselines/README_CN.md +++ b/research/cv/simple_baselines/README_CN.md @@ -53,7 +53,7 @@ simple_baselines的总体网络架构如下： ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html))的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html))的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/single_path_nas/README.md b/research/cv/single_path_nas/README.md index f4899b7ae..ae660b649 100644 --- a/research/cv/single_path_nas/README.md +++ b/research/cv/single_path_nas/README.md @@ -70,7 +70,7 @@ Dataset used：[ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed-precision](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) +The [mixed-precision](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) training method uses single-precision and half-precision data to improve the training speed of deep learning neural networks, while maintaining the network accuracy that can be achieved by single-precision training. Mixed-precision training increases computing speed and reduces memory usage, while supporting training larger models or diff --git a/research/cv/single_path_nas/README_CN.md b/research/cv/single_path_nas/README_CN.md index 3c71cfe53..62c6d04c6 100644 --- a/research/cv/single_path_nas/README_CN.md +++ b/research/cv/single_path_nas/README_CN.md @@ -57,7 +57,7 @@ single-path-nas的作者用一个7x7的大卷积，来代表3x3、5x5和7x7的 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # 环境要求 diff --git a/research/cv/sknet/README.md b/research/cv/sknet/README.md index 60f7315da..6e581761a 100644 --- a/research/cv/sknet/README.md +++ b/research/cv/sknet/README.md @@ -74,7 +74,7 @@ Dataset used: [CIFAR10](https://www.kaggle.com/c/cifar-10) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/research/cv/squeezenet/README.md b/research/cv/squeezenet/README.md index 7a045e948..de1f902c6 100644 --- a/research/cv/squeezenet/README.md +++ b/research/cv/squeezenet/README.md @@ -74,7 +74,7 @@ Dataset used: [ImageNet2012](http://www.image-net.org/) ## Mixed Precision -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) @@ -512,7 +512,7 @@ result: {'top_1_accuracy': 0.6094950384122919, 'top_5_accuracy': 0.8263244238156 ### Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following the steps below, this is a simple example: +If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: - Running on Ascend diff --git a/research/cv/squeezenet1_1/README.md b/research/cv/squeezenet1_1/README.md index 5042d64b3..ee112140d 100644 --- a/research/cv/squeezenet1_1/README.md +++ b/research/cv/squeezenet1_1/README.md @@ -304,7 +304,7 @@ Inference result is saved in current path, you can find result like this in acc. ### Inference -If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/docs/programming_guide/en/master/multi_platform_inference.html). Following the steps below, this is a simple example: +If you need to use the trained model to perform inference on multiple hardware platforms, such as GPU, Ascend 910 or Ascend 310, you can refer to this [Link](https://www.mindspore.cn/tutorials/experts/en/master/infer/inference.html). Following the steps below, this is a simple example: - Running on Ascend diff --git a/research/cv/ssd_ghostnet/README.md b/research/cv/ssd_ghostnet/README.md index cbc408763..1e8b82af2 100644 --- a/research/cv/ssd_ghostnet/README.md +++ b/research/cv/ssd_ghostnet/README.md @@ -210,7 +210,7 @@ If you want to run in modelarts, please check the official documentation of [mod ### Training on Ascend -To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/docs/programming_guide/en/master/convert_dataset.html) files by `coco_root`(coco dataset) or `iamge_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** +To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/tutorials/en/master/advanced/dataset/record.html) files by `coco_root`(coco dataset) or `iamge_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** - Distribute mode diff --git a/research/cv/ssd_inception_v2/README.md b/research/cv/ssd_inception_v2/README.md index 0a55f1663..cd0916115 100644 --- a/research/cv/ssd_inception_v2/README.md +++ b/research/cv/ssd_inception_v2/README.md @@ -213,7 +213,7 @@ bash scripts/docker_start.sh ssd:20.1.0 [DATA_DIR] [MODEL_DIR] ### [Training Process](#contents) -To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/docs/programming_guide/en/master/convert_dataset.html) files by `coco_root`(coco dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** +To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/tutorials/en/master/advanced/dataset/record.html) files by `coco_root`(coco dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** #### Training on GPU diff --git a/research/cv/ssd_inceptionv2/README_CN.md b/research/cv/ssd_inceptionv2/README_CN.md index f1b0298ee..fcf4a26d8 100644 --- a/research/cv/ssd_inceptionv2/README_CN.md +++ b/research/cv/ssd_inceptionv2/README_CN.md @@ -171,7 +171,7 @@ bash run_eval.sh [DEVICE_ID] [DATASET] [DATASET_PATH] [CHECKPOINT_PATH] [MINDREC ## 训练过程 -运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/convert_dataset.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** +运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset/record.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** ### Ascend上训练 diff --git a/research/cv/ssd_mobilenetV2/README.md b/research/cv/ssd_mobilenetV2/README.md index 3987cbddd..7b2ca8caf 100644 --- a/research/cv/ssd_mobilenetV2/README.md +++ b/research/cv/ssd_mobilenetV2/README.md @@ -221,7 +221,7 @@ bash scripts/run_eval_gpu.sh [DATASET] [CHECKPOINT_PATH] [DEVICE_ID] ### [Training Process](#contents) -To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/docs/programming_guide/en/master/convert_dataset.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** +To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/tutorials/en/master/advanced/dataset/record.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** #### Training on Ascend diff --git a/research/cv/ssd_mobilenetV2_FPNlite/README.md b/research/cv/ssd_mobilenetV2_FPNlite/README.md index 0190650aa..6f2cdd299 100644 --- a/research/cv/ssd_mobilenetV2_FPNlite/README.md +++ b/research/cv/ssd_mobilenetV2_FPNlite/README.md @@ -233,7 +233,7 @@ bash run_eval_gpu.sh [CONFIG_FILE] [DATASET] [CHECKPOINT_PATH] [DEVICE_ID] ### [Training Process](#contents) -To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/docs/programming_guide/en/master/convert_dataset.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** +To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/tutorials/en/master/advanced/dataset/record.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** #### Training on Ascend diff --git a/research/cv/ssd_resnet34/README.md b/research/cv/ssd_resnet34/README.md index e3938abec..8ce22a5ff 100644 --- a/research/cv/ssd_resnet34/README.md +++ b/research/cv/ssd_resnet34/README.md @@ -202,7 +202,7 @@ bash run_infer_310.sh [MINDIR_PATH] [DATA_PATH] [DEVICE_ID] ### [Training Process](#contents) -To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/convert_dataset.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** +To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset/record.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** #### Training on Ascend diff --git a/research/cv/ssd_resnet34/README_CN.md b/research/cv/ssd_resnet34/README_CN.md index 963267753..2aab91733 100644 --- a/research/cv/ssd_resnet34/README_CN.md +++ b/research/cv/ssd_resnet34/README_CN.md @@ -169,7 +169,7 @@ sh scripts/run_eval.sh [DEVICE_ID] [DATASET] [DATASET_PATH] [CHECKPOINT_PATH] [M ## 训练过程 -运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/convert_dataset.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** +运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset/record.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** ### Ascend上训练 diff --git a/research/cv/ssd_resnet50/README.md b/research/cv/ssd_resnet50/README.md index 116c1abb0..9075c7e95 100644 --- a/research/cv/ssd_resnet50/README.md +++ b/research/cv/ssd_resnet50/README.md @@ -204,7 +204,7 @@ Then you can run everything just like on ascend. ### [Training Process](#contents) -To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/docs/programming_guide/en/master/convert_dataset.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** +To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/tutorials/en/master/advanced/dataset/record.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** #### Training on Ascend diff --git a/research/cv/ssd_resnet50/README_CN.md b/research/cv/ssd_resnet50/README_CN.md index 0f2d4067e..4f7d5d167 100644 --- a/research/cv/ssd_resnet50/README_CN.md +++ b/research/cv/ssd_resnet50/README_CN.md @@ -163,7 +163,7 @@ bash run_eval.sh [DATASET] [CHECKPOINT_PATH] [DEVICE_ID] ## 训练过程 -运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/convert_dataset.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** +运行`train.py`训练模型。如果`mindrecord_dir`为空，则会通过`coco_root`（coco数据集）或`image_dir`和`anno_path`（自己的数据集）生成[MindRecord](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset/record.html)文件。**注意，如果mindrecord_dir不为空，将使用mindrecord_dir代替原始图像。** ### Ascend上训练 diff --git a/research/cv/ssd_resnet_34/README.md b/research/cv/ssd_resnet_34/README.md index 6fde21e6f..1704cb8eb 100644 --- a/research/cv/ssd_resnet_34/README.md +++ b/research/cv/ssd_resnet_34/README.md @@ -204,7 +204,7 @@ Major parameters in train.py and config.py for Multi GPU train: ### [Training Process](#contents) -To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/convert_dataset.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** +To train the model, run `train.py`. If the `mindrecord_dir` is empty, it will generate [mindrecord](https://www.mindspore.cn/tutorials/zh-CN/master/advanced/dataset/record.html) files by `coco_root`(coco dataset), `voc_root`(voc dataset) or `image_dir` and `anno_path`(own dataset). **Note if mindrecord_dir isn't empty, it will use mindrecord_dir instead of raw images.** #### Training on GPU diff --git a/research/cv/swin_transformer/README_CN.md b/research/cv/swin_transformer/README_CN.md index 7d0a842c9..23ed2d54c 100644 --- a/research/cv/swin_transformer/README_CN.md +++ b/research/cv/swin_transformer/README_CN.md @@ -53,7 +53,7 @@ SwinTransformer是新型的视觉Transformer，它可以用作计算机视觉的 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html) +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html) 的训练方法，使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 # [环境要求](#目录) diff --git a/research/cv/tsm/README_CN.md b/research/cv/tsm/README_CN.md index 1df30c8db..9f66040d2 100644 --- a/research/cv/tsm/README_CN.md +++ b/research/cv/tsm/README_CN.md @@ -59,7 +59,7 @@ TSM应用了一种通用而有效的时间转移模块。 时间转移模块将 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/cv/vgg19/README.md b/research/cv/vgg19/README.md index 35bad1b67..b48fc9c79 100644 --- a/research/cv/vgg19/README.md +++ b/research/cv/vgg19/README.md @@ -440,7 +440,7 @@ train_parallel1/log:epcoh: 2 step: 97, loss is 1.7133579 ... ``` -> About rank_table.json, you can refer to the [distributed training tutorial](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training.html). +> About rank_table.json, you can refer to the [distributed training tutorial](https://www.mindspore.cn/tutorials/experts/en/master/parallel/introduction.html). > **Attention** This will bind the processor cores according to the `device_num` and total processor numbers. If you don't expect to run pretraining with binding processor cores, remove the operations about `taskset` in `scripts/run_distribute_train.sh` ##### Run vgg19 on GPU diff --git a/research/cv/vgg19/README_CN.md b/research/cv/vgg19/README_CN.md index b4afd312f..9c99d010a 100644 --- a/research/cv/vgg19/README_CN.md +++ b/research/cv/vgg19/README_CN.md @@ -87,7 +87,7 @@ VGG 19网络主要由几个基本模块（包括卷积层和池化层）和三 ### 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 @@ -459,7 +459,7 @@ train_parallel1/log:epcoh: 2 step: 97, loss is 1.7133579 ... ``` -> 关于rank_table.json，可以参考[分布式并行训练](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training.html)。 +> 关于rank_table.json，可以参考[分布式并行训练](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/introduction.html)。 > **注意** 将根据`device_num`和处理器总数绑定处理器核。如果您不希望预训练中绑定处理器内核，请在`scripts/run_distribute_train.sh`脚本中移除`taskset`相关操作。 ##### GPU处理器环境运行VGG19 diff --git a/research/cv/vnet/README_CN.md b/research/cv/vnet/README_CN.md index 4e8c4148c..dd25398ea 100644 --- a/research/cv/vnet/README_CN.md +++ b/research/cv/vnet/README_CN.md @@ -101,7 +101,7 @@ VNet适用于医学图像分割，使用3D卷积，能够处理3D MR图像数据 - [MindSpore Python API](https://www.mindspore.cn/docs/api/zh-CN/master/index.html) - 生成config json文件用于多卡训练。 - [简易教程](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools) - - 详细配置方法请参照[官网教程](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/distributed_training_ascend.html#配置分布式环境变量)。 + - 详细配置方法请参照[官网教程](https://www.mindspore.cn/tutorials/experts/zh-CN/master/parallel/train_ascend.html#配置分布式环境变量)。 # 快速入门 diff --git a/research/cv/wideresnet/README.md b/research/cv/wideresnet/README.md index 80b40d4bb..f6defab27 100644 --- a/research/cv/wideresnet/README.md +++ b/research/cv/wideresnet/README.md @@ -208,7 +208,7 @@ bash run_standalone_train_gpu.sh [DATASET_PATH] [CONFIG_PATH] [EXPERIMENT_LABEL] For distributed training, a hostfile configuration needs to be created in advance. -Please follow the instructions in the link [GPU-Multi-Host](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_gpu.html). +Please follow the instructions in the link [GPU-Multi-Host](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_gpu.html). ##### Evaluation while training diff --git a/research/cv/wideresnet/README_CN.md b/research/cv/wideresnet/README_CN.md index 634064756..12e2ea27e 100644 --- a/research/cv/wideresnet/README_CN.md +++ b/research/cv/wideresnet/README_CN.md @@ -211,7 +211,7 @@ bash run_standalone_train_gpu.sh [DATASET_PATH] [CONFIG_PATH] [EXPERIMENT_LABEL] 对于分布式培训，需要提前创建主机文件配置。 -请按照链接中的说明操作 [GPU-Multi-Host](https://www.mindspore.cn/docs/programming_guide/en/master/distributed_training_gpu.html). +请按照链接中的说明操作 [GPU-Multi-Host](https://www.mindspore.cn/tutorials/experts/en/master/parallel/train_gpu.html). ## 培训时的评估 diff --git a/research/hpc/pinns/README.md b/research/hpc/pinns/README.md index 9ad24330a..6ea8de4f7 100644 --- a/research/hpc/pinns/README.md +++ b/research/hpc/pinns/README.md @@ -1,4 +1,4 @@ -# Contents +# Contents [查看中文](./README_CN.md) @@ -72,7 +72,7 @@ Dataset used：[cylinder nektar wake](https://github.com/maziarraissi/PINNs/tree ## [Mixed Precision](#Contents) -The [mixed precision](https://www.mindspore.cn/docs/programming_guide/en/master/enable_mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. +The [mixed precision](https://www.mindspore.cn/tutorials/experts/en/master/others/mixed_precision.html) training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data formats, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’. # [Environment Requirements](#contents) diff --git a/research/hpc/pinns/README_CN.md b/research/hpc/pinns/README_CN.md index d080e0adf..79cf1a900 100644 --- a/research/hpc/pinns/README_CN.md +++ b/research/hpc/pinns/README_CN.md @@ -70,7 +70,7 @@ Navier-Stokes方程是流体力学中描述粘性牛顿流体的方程。针对N ## [混合精度](#目录) -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # [环境要求](#目录) diff --git a/research/nlp/albert/README.md b/research/nlp/albert/README.md index 303e363ba..4943392e7 100644 --- a/research/nlp/albert/README.md +++ b/research/nlp/albert/README.md @@ -181,10 +181,9 @@ If you want to run in modelarts, please check the official documentation of [mod ``` For distributed training, an hccl configuration file with JSON format needs to be created in advance. -Please follow the instructions in the link below: -https:gitee.com/mindspore/mindspore/tree/master/model_zoo/utils/hccl_tools. -For dataset, if you want to set the format and parameters, a schema configuration file with JSON format needs to be created, please refer to [tfrecord](https://www.mindspore.cn/docs/programming_guide/en/master/dataset_loading.html#tfrecord) format. +Please follow the instructions in the link below: +[https://gitee.com/mindspore/models/tree/master/utils/hccl_tools](https://gitee.com/mindspore/models/tree/master/utils/hccl_tools). ```text For pretraining, schema file contains ["input_ids", "input_mask", "segment_ids", "next_sentence_labels", "masked_lm_positions", "masked_lm_ids", "masked_lm_weights"]. diff --git a/research/nlp/atae_lstm/README.md b/research/nlp/atae_lstm/README.md index 34aadc780..59a313be5 100644 --- a/research/nlp/atae_lstm/README.md +++ b/research/nlp/atae_lstm/README.md @@ -54,7 +54,7 @@ AttentionLSTM模型的输入由aspect和word向量组成，输入部分输入单 ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html)的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 diff --git a/research/nlp/rotate/README_CN.md b/research/nlp/rotate/README_CN.md index 2c8ae6dd1..dc5e24da7 100644 --- a/research/nlp/rotate/README_CN.md +++ b/research/nlp/rotate/README_CN.md @@ -86,7 +86,7 @@ bash run_infer_310.sh [MINDIR_HEAD_PATH] [MINDIR_TAIL_PATH] [DATASET_PATH] [NEED 在裸机环境（本地有Ascend 910 AI 处理器）进行分布式训练时，需要配置当前多卡环境的组网信息文件。 请遵循一下链接中的说明创建json文件： - + - GPU处理器环境运行 diff --git a/research/nlp/seq2seq/README_CN.md b/research/nlp/seq2seq/README_CN.md index 99c995590..45dc01a07 100644 --- a/research/nlp/seq2seq/README_CN.md +++ b/research/nlp/seq2seq/README_CN.md @@ -33,7 +33,7 @@ bash wmt14_en_fr.sh ## 混合精度 -采用[混合精度](https://www.mindspore.cn/docs/programming_guide/zh-CN/master/enable_mixed_precision.html))的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 +采用[混合精度](https://www.mindspore.cn/tutorials/experts/zh-CN/master/others/mixed_precision.html))的训练方法使用支持单精度和半精度数据来提高深度学习神经网络的训练速度，同时保持单精度训练所能达到的网络精度。混合精度训练提高计算速度、减少内存使用的同时，支持在特定硬件上训练更大的模型或实现更大批次的训练。 以FP16算子为例，如果输入数据类型为FP32，MindSpore后台会自动降低精度来处理数据。用户可打开INFO日志，搜索“reduce precision”查看精度降低的算子。 # 环境要求 -- Gitee