From 8566e01af97eb4a6234253962c6c6162b73d3c75 Mon Sep 17 00:00:00 2001 From: jieliang cai <975092674@qq.com> Date: Thu, 7 Jul 2022 12:41:20 +0000 Subject: [PATCH 1/7] update TensorFlow2/built-in/cv/image_classification/MNIST_ID2481_for_TensorFlow2.X/README.md. --- .../MNIST_ID2481_for_TensorFlow2.X/README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/TensorFlow2/built-in/cv/image_classification/MNIST_ID2481_for_TensorFlow2.X/README.md b/TensorFlow2/built-in/cv/image_classification/MNIST_ID2481_for_TensorFlow2.X/README.md index 42959f171..837522b1a 100644 --- a/TensorFlow2/built-in/cv/image_classification/MNIST_ID2481_for_TensorFlow2.X/README.md +++ b/TensorFlow2/built-in/cv/image_classification/MNIST_ID2481_for_TensorFlow2.X/README.md @@ -94,7 +94,7 @@ pip3 install requirements.txt ## 数据集准备 -1. 用户需自行下载ml-1m训练数据集，应有如下结构 +1. 用户需自行下载MNIST训练数据集，应有如下结构 ``` dataset/ ├── mnist -- Gitee From 0f03b93b7f4941a60dba1a7281dc6d6edc49bb92 Mon Sep 17 00:00:00 2001 From: jieliang cai <975092674@qq.com> Date: Fri, 8 Jul 2022 06:44:07 +0000 Subject: [PATCH 2/7] update TensorFlow2/built-in/nlp/Word2vec_ID2350_for_TensorFlow2.X/README.md. --- .../built-in/nlp/Word2vec_ID2350_for_TensorFlow2.X/README.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/TensorFlow2/built-in/nlp/Word2vec_ID2350_for_TensorFlow2.X/README.md b/TensorFlow2/built-in/nlp/Word2vec_ID2350_for_TensorFlow2.X/README.md index 378741945..be1ce0246 100644 --- a/TensorFlow2/built-in/nlp/Word2vec_ID2350_for_TensorFlow2.X/README.md +++ b/TensorFlow2/built-in/nlp/Word2vec_ID2350_for_TensorFlow2.X/README.md @@ -32,7 +32,8 @@ ## 简述 -本项目是基于TensorFlow2.X的文本分类任务，通过直接配置可以支持：TextCNN/TextRNN/TextRCNN/Transformer/Bert/Albert/DistilBert基本分类模型；TextCNN/TextRNN/TextRCNN/Transformer的token可选用词粒度/子粒度；Word2Vec特征增强后接TextCNN/TextRNN/TextRCNN/Transformer；支持Attention-TextCNN/TextRNN；FGM和PGD两种对抗方法的引入训练；对比学习方法R-drop引入；支持二分类和多分类，支持FocalLoss；保存为pb文件可供部署；项目代码支持交互式测试和批量测试。 +本项目是基于TensorFlow2.X的文本分类任务，使用Word2vec词向量训练模型进行文本分类。Word2vec是一种将词转化成向量的方法，其中包含两种算法，分别是skip-gram和CBOW，它们最大的区别是skip-gram是通过中心词去预测中心词周围的词，而CBOW是通过周围的词去预测中心词。 + - 参考论文： -- Gitee From 0fcf419f8e22ba087704cb2f77f9016c9e731378 Mon Sep 17 00:00:00 2001 From: jieliang cai <975092674@qq.com> Date: Tue, 12 Jul 2022 06:09:12 +0000 Subject: [PATCH 3/7] modify --- .../test/train_full_1p_4096bs_dynamic_noeval.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh index 632d4525f..e5e95a7b0 100644 --- a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh +++ b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh @@ -22,7 +22,7 @@ Network="Transformer_ID0633_for_TensorFlow2.X" #训练batch_size batch_size=4096 #训练step -train_steps=400000 +train_steps=10000 #TF2.X独有，不需要修改 #export NPU_ENABLE_PERF=true -- Gitee From 0f1fee3679d97af592179ef9df6d2cb9fa9ddb35 Mon Sep 17 00:00:00 2001 From: jieliang cai <975092674@qq.com> Date: Tue, 12 Jul 2022 06:30:54 +0000 Subject: [PATCH 4/7] modify --- .../test/train_full_1p_4096bs_dynamic_noeval.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh index e5e95a7b0..80074912a 100644 --- a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh +++ b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh @@ -120,7 +120,7 @@ do --param_set=big \ --train_steps=${train_steps} \ --batch_size=${batch_size} \ - --steps_between_evals=10000 \ + --steps_between_evals=1000 \ --max_length=64 \ --mode=train \ --decode_batch_size=32 \ -- Gitee From dc5afe55d4b5dcb32e845175b2bf5cdd6bf2c32e Mon Sep 17 00:00:00 2001 From: jieliang cai <975092674@qq.com> Date: Tue, 12 Jul 2022 07:56:31 +0000 Subject: [PATCH 5/7] modify --- .../test/train_full_1p_4096bs_dynamic_noeval.sh | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh index 80074912a..4e00200ac 100644 --- a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh +++ b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh @@ -4,7 +4,8 @@ cur_path=`pwd` #集合通信参数,不需要修改 - +export DUMP_GE_GRAPH=2 +export DUMP_GRAPH_LEVEL=3 export RANK_SIZE=1 export JOB_ID=10087 export RANK_ID_START=0 @@ -30,7 +31,7 @@ train_steps=10000 #维测参数，precision_mode需要模型审视修改 precision_mode="allow_mix_precision" #维持参数，以下不需要修改 -over_dump=False +over_dump=True data_dump_flag=False data_dump_step="10" profiling=False @@ -120,7 +121,7 @@ do --param_set=big \ --train_steps=${train_steps} \ --batch_size=${batch_size} \ - --steps_between_evals=1000 \ + --steps_between_evals=10000 \ --max_length=64 \ --mode=train \ --decode_batch_size=32 \ -- Gitee From 28567d282d9dfd26f67edceda0f590dbc8ce782a Mon Sep 17 00:00:00 2001 From: jieliang cai <975092674@qq.com> Date: Tue, 12 Jul 2022 08:04:16 +0000 Subject: [PATCH 6/7] modify --- .../test/train_full_1p_4096bs_dynamic_noeval.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh index 4e00200ac..70e04b160 100644 --- a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh +++ b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh @@ -31,7 +31,7 @@ train_steps=10000 #维测参数，precision_mode需要模型审视修改 precision_mode="allow_mix_precision" #维持参数，以下不需要修改 -over_dump=True +over_dump=False data_dump_flag=False data_dump_step="10" profiling=False -- Gitee From b6614ee98939cf6b3b09833f3178b1a58050961a Mon Sep 17 00:00:00 2001 From: jieliang cai <975092674@qq.com> Date: Wed, 13 Jul 2022 06:11:32 +0000 Subject: [PATCH 7/7] modify --- .../test/train_full_1p_4096bs_dynamic_noeval.sh | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh index 70e04b160..1f48774f0 100644 --- a/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh +++ b/TensorFlow2/built-in/nlp/Transformer_ID0633_for_TensorFlow2.X/test/train_full_1p_4096bs_dynamic_noeval.sh @@ -23,7 +23,7 @@ Network="Transformer_ID0633_for_TensorFlow2.X" #训练batch_size batch_size=4096 #训练step -train_steps=10000 +train_steps=300 #TF2.X独有，不需要修改 #export NPU_ENABLE_PERF=true @@ -121,7 +121,7 @@ do --param_set=big \ --train_steps=${train_steps} \ --batch_size=${batch_size} \ - --steps_between_evals=10000 \ + --steps_between_evals=100 \ --max_length=64 \ --mode=train \ --decode_batch_size=32 \ @@ -131,7 +131,7 @@ do --dtype=fp16 \ --distribution_strategy='one_device' \ --enable_time_history=true \ - --log_steps=1000 \ + --log_steps=100 \ --loss_scale='dynamic' \ --precision_mode=${precision_mode} \ --over_dump=${over_dump} \ -- Gitee