Revert "remove memory optimize usage in scripts, test=develop (#3242)" (#3247)

This reverts commit 5e41760f.

PaddleCV/image_classification/README.md

@@ -68,6 +68,7 @@ python train.py \
        --class_dim=1000 \
        --image_shape=3,224,224 \
        --model_save_dir=output/ \
+       --with_mem_opt=False \
        --with_inplace=True \
        --lr_strategy=piecewise_decay \
        --lr=0.1
@@ -82,6 +83,7 @@ python train.py \
 * **class_dim**: the class number of the classification task. Default: 1000.
 * **image_shape**: input size of the network. Default: "3,224,224".
 * **model_save_dir**: the directory to save trained model. Default: "output".
+* **with_mem_opt**: whether to use memory optimization or not. Default: False.
 * **with_inplace**: whether to use inplace memory optimization or not. Default: True.
 * **lr_strategy**: learning rate changing strategy. Default: "piecewise_decay".
 * **lr**: initialized learning rate. Default: 0.1.
@@ -152,6 +154,8 @@ Note: Add and adjust other parameters accroding to specific models and tasks.
 
 You may add `--fp16=1` to start train using mixed precisioin training, which the training process will use float16 and the output model ("master" parameters) is saved as float32. You also may need to pass `--scale_loss` to overcome accuracy issues, usually `--scale_loss=8.0` will do.
 
+Note that currently `--fp16` can not use together with `--with_mem_opt`, so pass `--with_mem_opt=0` to disable memory optimization pass.
+
 ### CE
 
 CE is only for internal testing, don't have to set it.

PaddleCV/image_classification/README_cn.md

@@ -64,6 +64,7 @@ python train.py \
        --class_dim=1000 \
        --image_shape=3,224,224 \
        --model_save_dir=output/ \
+       --with_mem_opt=False \
        --with_inplace=True \
        --lr_strategy=piecewise_decay \
        --lr=0.1
@@ -78,6 +79,7 @@ python train.py \
 * **class_dim**: 类别数，默认值: 1000
 * **image_shape**: 图片大小，默认值: "3,224,224"
 * **model_save_dir**: 模型存储路径，默认值: "output/"
+* **with_mem_opt**: 是否开启显存优化，默认值: False
 * **with_inplace**: 是否开启inplace显存优化，默认值: True
 * **lr_strategy**: 学习率变化策略，默认值: "piecewise_decay"
 * **lr**: 初始学习率，默认值: 0.1
@@ -140,6 +142,8 @@ python infer.py \
 
 可以通过开启`--fp16=True`启动混合精度训练，这样训练过程会使用float16数据，并输出float32的模型参数（"master"参数）。您可能需要同时传入`--scale_loss`来解决fp16训练的精度问题，通常传入`--scale_loss=8.0`即可。
 
+注意，目前混合精度训练不能和内存优化功能同时使用，所以需要传`--with_mem_opt=False`这个参数来禁用内存优化功能。
+
 ### CE测试
 
 注意：CE相关代码仅用于内部测试，enable_ce默认设置False。

PaddleCV/image_classification/dist_train/dist_train.py

@@ -46,6 +46,7 @@ def parse_args():
     add_arg('class_dim',        int,   1000,                 "Class number.")
     add_arg('image_shape',      str,   "3,224,224",          "input image size")
     add_arg('model_save_dir',   str,   "output",             "model save directory")
+    add_arg('with_mem_opt',     bool,  False,                "Whether to use memory optimization or not.")
     add_arg('pretrained_model', str,   None,                 "Whether to use pretrained model.")
     add_arg('checkpoint',       str,   None,                 "Whether to resume checkpoint.")
     add_arg('lr',               float, 0.1,                  "set learning rate.")

PaddleCV/image_classification/run.sh

@@ -7,6 +7,7 @@ python train.py \
        --class_dim=1000 \
        --image_shape=3,224,224 \
        --model_save_dir=output/ \
+       --with_mem_opt=True \
        --lr_strategy=cosine_decay \
        --lr=0.1 \
        --num_epochs=200 \
@@ -21,6 +22,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=piecewise_decay \
 #       --num_epochs=120 \
 #       --lr=0.01 \
@@ -37,6 +39,7 @@ python train.py \
 #        --model_save_dir=output/ \
 #        --lr=0.02 \
 #        --num_epochs=120 \
+#        --with_mem_opt=True \
 #        --l2_decay=1e-4
 
 #SqueezeNet1_1
@@ -50,6 +53,7 @@ python train.py \
 #        --model_save_dir=output/ \
 #        --lr=0.02 \
 #        --num_epochs=120 \
+#        --with_mem_opt=True \
 #        --l2_decay=1e-4
 
 #VGG11:
@@ -63,6 +67,7 @@ python train.py \
 #        --model_save_dir=output/ \
 #        --lr=0.1 \
 #        --num_epochs=90 \
+#        --with_mem_opt=True \
 #        --l2_decay=2e-4
 
 #VGG13:
@@ -76,6 +81,7 @@ python train.py \
 #        --lr=0.01 \
 #        --num_epochs=90 \
 #        --model_save_dir=output/ \
+#        --with_mem_opt=True \
 #        --l2_decay=3e-4
 
 #VGG16:
@@ -89,6 +95,7 @@ python train.py \
 #        --model_save_dir=output/ \
 #        --lr=0.01 \
 #        --num_epochs=90 \
+#        --with_mem_opt=True \
 #        --l2_decay=3e-4
 
 #VGG19:
@@ -101,6 +108,7 @@ python train.py \
 #        --lr_strategy=cosine_decay \
 #        --lr=0.01 \
 #        --num_epochs=90 \
+#        --with_mem_opt=True \
 #        --model_save_dir=output/ \
 #        --l2_decay=3e-4
 
@@ -112,6 +120,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=piecewise_decay \
 #       --num_epochs=120 \
 #       --lr=0.1 \
@@ -125,6 +134,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_decay \
 #       --num_epochs=240 \
 #       --lr=0.1 \
@@ -140,6 +150,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_decay \
 #       --num_epochs=240 \
 #       --lr=0.1 \
@@ -155,6 +166,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_decay \
 #       --num_epochs=240 \
 #       --lr=0.1 \
@@ -168,6 +180,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_decay \
 #       --num_epochs=240 \
 #       --lr=0.1 \
@@ -181,6 +194,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_decay \
 #       --num_epochs=240 \
 #       --lr=0.1 \
@@ -194,6 +208,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_warmup_decay \
 #       --num_epochs=240 \
 #       --lr=0.5 \
@@ -210,6 +225,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_warmup_decay \
 #       --num_epochs=240 \
 #       --lr=0.5 \
@@ -226,6 +242,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_warmup_decay \
 #       --num_epochs=240 \
 #       --lr=0.5 \
@@ -242,6 +259,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_warmup_decay \
 #       --num_epochs=240 \
 #       --lr=0.5 \
@@ -256,6 +274,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_warmup_decay \
 #       --num_epochs=240 \
 #       --lr=0.25 \
@@ -271,6 +290,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_warmup_decay \
 #       --num_epochs=240 \
 #       --lr=0.25 \
@@ -284,6 +304,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_warmup_decay \
 #       --lr=0.5 \
 #       --num_epochs=240 \
@@ -297,6 +318,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
@@ -310,6 +332,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
@@ -323,6 +346,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=piecewise_decay \
 #       --num_epochs=120 \
 #       --lr=0.1 \
@@ -338,6 +362,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 
@@ -351,6 +376,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=7e-5 \
 #       --use_mixup=True \
@@ -365,6 +391,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=piecewise_decay \
 #       --num_epochs=120 \
 #       --lr=0.1 \
@@ -380,6 +407,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --use_mixup=True \
@@ -395,6 +423,7 @@ python train.py \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
 #       --lr_strategy=piecewise_decay \
+#       --with_mem_opt=True \
 #       --lr=0.1 \
 #       --num_epochs=120 \
 #       --l2_decay=1e-4
@@ -409,6 +438,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --use_mixup=True \
@@ -425,6 +455,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --use_mixup=True \
@@ -441,6 +472,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4
 
@@ -454,6 +486,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --use_mixup=True \
@@ -470,6 +503,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4
 
@@ -483,6 +517,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --use_mixup=True \
@@ -499,6 +534,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4
 
@@ -512,6 +548,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=15e-5
 
@@ -525,6 +562,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --use_mixup=True \
@@ -541,6 +579,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4
 
@@ -554,6 +593,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=18e-5
 
@@ -567,6 +607,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4
 
@@ -580,6 +621,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4
 
@@ -593,6 +635,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4
 
@@ -606,6 +649,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4
 
@@ -619,6 +663,7 @@ python train.py \
 #       --lr_strategy=piecewise_decay \
 #       --lr=0.1 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4
 
@@ -633,6 +678,7 @@ python train.py \
 #       --model_save_dir=output/ \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --l2_decay=1.2e-4
 
 #SE_ResNeXt101_32x4d:
@@ -646,6 +692,7 @@ python train.py \
 #       --model_save_dir=output/ \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --l2_decay=1.5e-5
 
 # SE_154
@@ -658,6 +705,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --use_mixup=True \
@@ -672,6 +720,7 @@ python train.py \
 #       --class_dim=1000 \
 #       --image_shape=3,224,224 \
 #       --model_save_dir=output/ \
+#       --with_mem_opt=True \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.01 \
 #       --num_epochs=200 \
@@ -687,6 +736,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.045 \
 #       --num_epochs=120 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --resize_short_size=320
@@ -701,6 +751,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.045 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --use_mixup=True \
@@ -718,6 +769,7 @@ python train.py \
 #       --lr_strategy=cosine_decay \
 #       --lr=0.1 \
 #       --num_epochs=200 \
+#       --with_mem_opt=True \
 #       --model_save_dir=output/ \
 #       --l2_decay=1e-4 \
 #       --use_mixup=True \
@@ -735,6 +787,7 @@ python train.py \
 #       --image_shape=3,224,224 \
 #       --lr=0.001 \
 #       --num_epochs=120 \
+#       --with_mem_opt=False \
 #       --model_save_dir=output/ \
 #       --lr_strategy=adam \
 #       --use_gpu=False

PaddleCV/image_classification/train.py

@@ -443,6 +443,8 @@ def train(args):
     use_ngraph = os.getenv('FLAGS_use_ngraph')
     if not use_ngraph:
         build_strategy = fluid.BuildStrategy()
+        # memopt may affect GC results
+        #build_strategy.memory_optimize = args.with_mem_opt
         build_strategy.enable_inplace = args.with_inplace
         #build_strategy.fuse_all_reduce_ops=1
 

PaddleNLP/Research/ACL2019-JEMT/README.md

@@ -81,6 +81,7 @@ beta代表发音信息的权重。这表明，即使将绝大部分权重放在
   --sort_type pool \
   --pool_size 200000 \
   --use_py_reader False \
+  --use_mem_opt False \
   --enable_ce False \
   --fetch_steps 1 \
   pass_num 100 \