remove gc flags, test=develop

PaddleCV/PaddleDetection/tools/eval.py

@@ -18,19 +18,6 @@ from __future__ import print_function
 
 import os
 
-
-def set_paddle_flags(**kwargs):
-    for key, value in kwargs.items():
-        if os.environ.get(key, None) is None:
-            os.environ[key] = str(value)
-
-
-# NOTE(paddle-dev): All of these flags should be set before
-# `import paddle`. Otherwise, it would not take any effect.
-set_paddle_flags(
-    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
-)
-
 import paddle.fluid as fluid
 
 from ppdet.utils.eval_utils import parse_fetches, eval_run, eval_results, json_eval_results

PaddleCV/PaddleDetection/tools/infer.py

@@ -22,19 +22,6 @@ import glob
 import numpy as np
 from PIL import Image
 
-
-def set_paddle_flags(**kwargs):
-    for key, value in kwargs.items():
-        if os.environ.get(key, None) is None:
-            os.environ[key] = str(value)
-
-
-# NOTE(paddle-dev): All of these flags should be set before
-# `import paddle`. Otherwise, it would not take any effect.
-set_paddle_flags(
-    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
-)
-
 from paddle import fluid
 
 from ppdet.utils.cli import print_total_cfg

PaddleCV/PaddleDetection/tools/train.py

@@ -22,19 +22,6 @@ import numpy as np
 import datetime
 from collections import deque
 
-
-def set_paddle_flags(**kwargs):
-    for key, value in kwargs.items():
-        if os.environ.get(key, None) is None:
-            os.environ[key] = str(value)
-
-
-# NOTE(paddle-dev): All of these flags should be set before
-# `import paddle`. Otherwise, it would not take any effect.
-set_paddle_flags(
-    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
-)
-
 from paddle import fluid
 
 from ppdet.experimental import mixed_precision_context

PaddleCV/PaddleGAN/cycle_gan/train.py

@@ -35,13 +35,7 @@ else:
 # not take any effect. 
 set_paddle_flags({
     'FLAGS_cudnn_exhaustive_search': use_cudnn_exhaustive_search,
-    'FLAGS_conv_workspace_size_limit': 256,
-    'FLAGS_eager_delete_tensor_gb': 0,  # enable gc 
-    # You can omit the following settings, because the default
-    # value of FLAGS_memory_fraction_of_eager_deletion is 1,
-    # and default value of FLAGS_fast_eager_deletion_mode is 1 
-    'FLAGS_memory_fraction_of_eager_deletion': 1,
-    'FLAGS_fast_eager_deletion_mode': 1
+    'FLAGS_conv_workspace_size_limit': 256
 })
 
 import random

PaddleCV/PaddleGAN/scripts/run_SPADE.sh

-export FLAGS_eager_delete_tensor_gb=0.0
-export FLAGS_fast_eager_deletion_mode=1
 export FLAGS_fraction_of_gpu_memory_to_use=0.01
 CUDA_VISIBLE_DEVICES=0 python train.py --model_net SPADE --dataset cityscapes --train_list train_list --test_list val_list --crop_type Random --batch_size 1 --epoch 200 --load_height 612 --load_width 1124 --crop_height 512 --crop_width 1024 --label_nc 36

PaddleCV/PaddleVideo/models/bmn/README.md

@@ -29,9 +29,7 @@ BMN的训练数据采用ActivityNet1.3提供的数据集，数据下载及准备
 数据准备完毕后，可以通过如下两种方式启动训练：
 
     export CUDA_VISIBLE_DEVICES=0,1,2,3
-    export FLAGS_eager_delete_tensor_gb=0.0
     export FLAGS_fraction_of_gpu_memory_to_use=0.98
-    export FLAGS_fast_eager_deletion_mode=1
     python train.py --model_name=BMN \
                     --config=./configs/bmn.yaml \
                     --log_interval=10 \

PaddleCV/PaddleVideo/models/bsn/README.md

@@ -30,9 +30,7 @@ TEM模块以snippet-level的特征序列作为输入，预测每一个时序位
 数据准备完毕后，可以通过如下两种方式启动训练：
 
     export CUDA_VISIBLE_DEVICES=0
-    export FLAGS_eager_delete_tensor_gb=0.0
     export FLAGS_fraction_of_gpu_memory_to_use=0.98
-    export FLAGS_fast_eager_deletion_mode=1
     python train.py --model_name=BsnTem \
                     --config=./configs/bsn_tem.yaml \
                     --log_interval=10 \
@@ -60,9 +58,7 @@ PEM模块以PGM模块输出的BSP特征作为输入，输出proposal包含动作
 数据准备完毕后，可以通过如下两种方式启动训练：
 
     export CUDA_VISIBLE_DEVICES=0
-    export FLAGS_eager_delete_tensor_gb=0.0
     export FLAGS_fraction_of_gpu_memory_to_use=0.98
-    export FLAGS_fast_eager_deletion_mode=1
     python train.py --model_name=BsnPem \
                     --config=./configs/bsn_pem.yaml \
                     --log_interval=10 \

PaddleCV/PaddleVideo/models/ctcn/README.md

@@ -25,8 +25,6 @@ C-TCN的训练数据采用ActivityNet1.3提供的数据集，数据下载及准
 数据准备完毕后，可以通过如下两种方式启动训练：
 
     export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-    export FLAGS_fast_eager_deletion_mode=1
-    export FLAGS_eager_delete_tensor_gb=0.0
     export FLAGS_fraction_of_gpu_memory_to_use=0.98
     python train.py --model_name=CTCN \
                     --config=./configs/ctcn.yaml \

PaddleCV/PaddleVideo/models/tsm/README.md

@@ -35,8 +35,6 @@ TSM的训练数据采用由DeepMind公布的Kinetics-400动作识别数据集。
 数据准备完毕后，可以通过如下两种方式启动训练：
 
     export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-    export FLAGS_fast_eager_deletion_mode=1
-    export FLAGS_eager_delete_tensor_gb=0.0
     export FLAGS_fraction_of_gpu_memory_to_use=0.98
     python train.py --model_name=TSM \
                     --config=./configs/tsm.yaml \

PaddleCV/PaddleVideo/models/tsn/README.md

@@ -26,8 +26,6 @@ TSN的训练数据采用由DeepMind公布的Kinetics-400动作识别数据集。
 数据准备完毕后，可以通过如下两种方式启动训练：
 
     export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-    export FLAGS_fast_eager_deletion_mode=1
-    export FLAGS_eager_delete_tensor_gb=0.0
     export FLAGS_fraction_of_gpu_memory_to_use=0.98
     python train.py --model_name=TSN \
                     --config=./configs/tsn.yaml \

PaddleCV/PaddleVideo/run.sh

@@ -25,8 +25,6 @@ weights="" #set the path of weights to enable eval and predicut, just ignore thi
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 #export CUDA_VISIBLE_DEVICES=0,1,2,3
 #export CUDA_VISIBLE_DEVICES=0
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 if [ "$mode"x == "train"x ]; then

PaddleCV/PaddleVideo/run_ce.sh

 #!/bin/bash
 
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 export CUDA_VISIBLE_DEVICES=0
 

PaddleCV/deeplabv3+/train.py

@@ -27,12 +27,6 @@ def set_paddle_flags(flags):
 # set before `import paddle`. Otherwise, it would
 # not take any effect. 
 set_paddle_flags({
-    'FLAGS_eager_delete_tensor_gb': 0,  # enable GC 
-    # You can omit the following settings, because the default
-    # value of FLAGS_memory_fraction_of_eager_deletion is 1,
-    # and default value of FLAGS_fast_eager_deletion_mode is 1 
-    'FLAGS_memory_fraction_of_eager_deletion': 1,
-    'FLAGS_fast_eager_deletion_mode': 1,
     # Setting the default used gpu memory
     'FLAGS_fraction_of_gpu_memory_to_use': 0.98
 })

PaddleCV/face_detection/train.py

@@ -21,21 +21,6 @@ import numpy as np
 import time
 import argparse
 import functools
-
-
-def set_paddle_flags(**kwargs):
-    for key, value in kwargs.items():
-        if os.environ.get(key, None) is None:
-            os.environ[key] = str(value)
-
-
-# NOTE(paddle-dev): All of these flags should be
-# set before `import paddle`. Otherwise, it would
-# not take any effect. 
-set_paddle_flags(
-    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
-)
-
 import paddle
 import paddle.fluid as fluid
 from pyramidbox import PyramidBox

PaddleCV/image_classification/scripts/train/AlexNet.sh

 #Training details
 #GPU: NVIDIA® Tesla® P40  8cards 120epochs 55h
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #AlexNet:

PaddleCV/image_classification/scripts/train/EfficientNetB0.sh

 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.96
 
 

PaddleCV/image_classification/scripts/train/GoogLeNet.sh

 #Training details
 #GPU: NVIDIA® Tesla® V100 4cards 200epochs 132h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #GoogLeNet:

PaddleCV/image_classification/scripts/train/InceptionV4.sh

 #Training details
 #GPU: NVIDIA® Tesla® V100 8cards 200epochs 367h
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #InceptionV4

PaddleCV/image_classification/scripts/train/MobileNetV1.sh

 #Training details
 #GPU: NVIDIA® Tesla® V100 4cards 120epochs 55h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 

PaddleCV/image_classification/scripts/train/MobileNetV1_x0_25.sh

 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 

PaddleCV/image_classification/scripts/train/MobileNetV1_x0_5.sh

 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 

PaddleCV/image_classification/scripts/train/MobileNetV1_x0_75.sh

 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 

PaddleCV/image_classification/scripts/train/MobileNetV2.sh

 #Training details
 #GPU: NVIDIA® Tesla® V100 4cards 240epochs 135h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 

PaddleCV/image_classification/scripts/train/MobileNetV2_x0_75.sh

 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 

PaddleCV/image_classification/scripts/train/ResNeXt101_32x4d.sh

 #Training details
 #Missed
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNeXt101_32x4d

PaddleCV/image_classification/scripts/train/ResNeXt101_64x4d.sh

 #Training details
 #Missed
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNeXt101_64x4d

PaddleCV/image_classification/scripts/train/ResNeXt101_vd_32x4d.sh

 #Training details
 #Missed
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNeXt101_vd_32x4d

PaddleCV/image_classification/scripts/train/ResNeXt101_vd_64x4d.sh

 #Training details
 #Missed
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNeXt101_vd_64x4d

PaddleCV/image_classification/scripts/train/ResNeXt152_64x4d.sh

 #Training details
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNeXt152_64x4d

PaddleCV/image_classification/scripts/train/ResNet101.sh

 ##Training details
 #GPU: NVIDIA® Tesla® V100 4cards 120epochs 100h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNet101:

PaddleCV/image_classification/scripts/train/ResNet101_vd.sh

 #Training details
 #GPU: NVIDIA® Tesla® V100  4cards 200epochs 182h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNet101_vd

PaddleCV/image_classification/scripts/train/ResNet152.sh

 ##Training details
 #GPU: NVIDIA® Tesla® P40 8cards 120epochs 200h
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNet152:

PaddleCV/image_classification/scripts/train/ResNet152_vd.sh

 ##Training details
 #GPU: NVIDIA® Tesla® P40 8cards 200epochs 346h
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 python train.py \

PaddleCV/image_classification/scripts/train/ResNet18.sh

 ##Training details
 #GPU: NVIDIA® Tesla® V100 4cards 120epochs 67h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNet18:

PaddleCV/image_classification/scripts/train/ResNet18_vd.sh

 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 python train.py \

PaddleCV/image_classification/scripts/train/ResNet200_vd.sh

 #Training details
 #Machine: Missed
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNet200_vd

PaddleCV/image_classification/scripts/train/ResNet34.sh

 #Training details
 #GPU: NVIDIA® Tesla® V100 4cards 120epochs 73h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 #ResNet34:
 python train.py \

PaddleCV/image_classification/scripts/train/ResNet34_vd.sh

 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 python train.py \

PaddleCV/image_classification/scripts/train/ResNet50.sh

 ##Training details
 #GPU: NVIDIA® Tesla® V100 4cards 120epochs 67h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNet50:

PaddleCV/image_classification/scripts/train/ResNet50_vc.sh

 ##Training details
 #GPU: NVIDIA® Tesla® V100 4cards 200epochs 141h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #ResNet50_vc

PaddleCV/image_classification/scripts/train/ResNet50_vd.sh

 #Training details
 #GPU: NVIDIA® Tesla® V100 4cards 200epochs 120h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 python train.py \

PaddleCV/image_classification/scripts/train/SENet154_vd.sh

 #Training details
 #GPU: NVIDIA® Tesla® P40 8cards 200epochs 916h
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #SE_154

PaddleCV/image_classification/scripts/train/SE_ResNeXt101_32x4d.sh

 ##Training details
 #GPU: NVIDIA® Tesla® P40 8cards 120epochs 566h
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #SE_ResNeXt101_32x4d:

PaddleCV/image_classification/scripts/train/SE_ResNeXt50_32x4d.sh

 #Training details
 #Machine:V100 4cards 200epochs 282h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 

PaddleCV/image_classification/scripts/train/ShuffleNetV2.sh

 ##Training details
 #GPU: NVIDIA® Tesla® K40 4cards 240epochs 156h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 python train.py \

PaddleCV/image_classification/scripts/train/ShuffleNetV2_swish.sh

 ##Training details
 #GPU: NVIDIA® Tesla® K40 4cards 240epochs 156h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 python train.py \

PaddleCV/image_classification/scripts/train/VGG11.sh

 #Training details
 #GPU: NVIDIA® Tesla® P40 8cards 90epochs 52h
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #VGG11:

PaddleCV/image_classification/scripts/train/VGG13.sh

 #Training details
 #GPU: NVIDIA® Tesla® V100 4cards 90epochs 58h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #VGG13:

PaddleCV/image_classification/scripts/train/VGG16.sh

 #Training details
 #GPU: NVIDIA® Tesla® P40 8cards 90epochs 72h
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #VGG16:

PaddleCV/image_classification/scripts/train/VGG19.sh

 #Training details
 #GPU: NVIDIA® Tesla® V100 4cards 150epochs 173h
 export CUDA_VISIBLE_DEVICES=0,1,2,3
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export FLAGS_fraction_of_gpu_memory_to_use=0.98
 
 #VGG19:

PaddleCV/image_classification/train.py

@@ -31,7 +31,6 @@ def set_paddle_flags(flags):
 # set before `import paddle`. Otherwise, it would
 # not take any effect. 
 set_paddle_flags({
-    'FLAGS_eager_delete_tensor_gb': 0,  # enable gc 
     'FLAGS_fraction_of_gpu_memory_to_use': 0.98
 })
 

PaddleCV/rcnn/train.py

@@ -26,8 +26,6 @@ def set_paddle_flags(flags):
 
 set_paddle_flags({
     'FLAGS_conv_workspace_size_limit': 500,
-    'FLAGS_eager_delete_tensor_gb': 0,  # enable gc
-    'FLAGS_memory_fraction_of_eager_deletion': 1,
     'FLAGS_fraction_of_gpu_memory_to_use': 0.98
 })
 

PaddleCV/ssd/train.py

@@ -20,20 +20,6 @@ import shutil
 import math
 import multiprocessing
 
-
-def set_paddle_flags(**kwargs):
-    for key, value in kwargs.items():
-        if os.environ.get(key, None) is None:
-            os.environ[key] = str(value)
-
-
-# NOTE(paddle-dev): All of these flags should be
-# set before `import paddle`. Otherwise, it would
-# not take any effect. 
-set_paddle_flags(
-    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
-)
-
 import paddle
 import paddle.fluid as fluid
 import reader

PaddleCV/yolov3/train.py

@@ -25,8 +25,6 @@ def set_paddle_flags(flags):
 
 
 set_paddle_flags({
-    'FLAGS_eager_delete_tensor_gb': 0,  # enable gc
-    'FLAGS_memory_fraction_of_eager_deletion': 1,
     'FLAGS_fraction_of_gpu_memory_to_use': 0.98
 })
 

PaddleNLP/Research/ACL2018-DAM/run.sh

 export CUDA_VISIBLE_DEVICES=3
-export FLAGS_eager_delete_tensor_gb=0.0
 
 #train on ubuntu 
 python -u main.py \

PaddleNLP/Research/ACL2018-DAM/run_CPU.sh

 export CPU_NUM=1
-export FLAGS_eager_delete_tensor_gb=0.0
 
 #train on ubuntu 
 python -u main.py \

PaddleNLP/Research/MRQA2019-BASELINE/run_finetuning.sh

@@ -18,7 +18,6 @@
 set -xe
 
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1
 
 # set CUDA_VISIBLE_DEVICES
 export CUDA_VISIBLE_DEVICES=0

PaddleNLP/Research/MRQA2019-BASELINE/run_predict.sh

@@ -18,7 +18,6 @@
 set -xe
 
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1
 
 # set CUDA_VISIBLE_DEVICES
 export CUDA_VISIBLE_DEVICES=0

PaddleNLP/Research/MRQA2019-D-NET/knowledge_distillation/run_distill.sh

 #!/bin/bash
 
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1
 
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 if  [ ! "$CUDA_VISIBLE_DEVICES" ]

PaddleNLP/Research/MRQA2019-D-NET/multi_task_learning/run_multi_task.sh

@@ -2,7 +2,6 @@
 
 # for gpu memory optimization
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1
 
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 

PaddleNLP/Research/MRQA2019-D-NET/server/xlnet_server/serve.sh

 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1
 export FLAGS_fraction_of_gpu_memory_to_use=0.1
 
 python serve.py ./infer_model_800_bs128 5001 &

PaddleNLP/dialogue_model_toolkit/auto_dialogue_evaluation/.run_ce.sh

 #!/bin/bash
 
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1.0
 
 export CUDA_VISIBLE_DEVICES=0
 

PaddleNLP/dialogue_model_toolkit/auto_dialogue_evaluation/README.md

@@ -156,7 +156,6 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 ```
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1  #开启显存优化
 
 export CUDA_VISIBLE_DEVICES=0  #GPU单卡训练
 #export CUDA_VISIBLE_DEVICES=0,1,2,3  #GPU多卡训练
@@ -222,7 +221,6 @@ task_type: train、predict、evaluate、inference, 选择4个参数选项中任
 
 ```
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1  #开启显存优化
 
 export CUDA_VISIBLE_DEVICES=0  #GPU单卡训练
 #export CUDA_VISIBLE_DEVICES=0,1,2,3  #GPU多卡训练
@@ -298,7 +296,6 @@ export CUDA_VISIBLE_DEVICES=0 #用户可自行指定空闲的卡
 
 ```
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1  #开启显存优化
 
 export CUDA_VISIBLE_DEVICES=0  #单卡预测
 #export CUDA_VISIBLE_DEVICES=  #CPU预测
@@ -346,7 +343,6 @@ task_type: train、predict、evaluate、inference, 选择4个参数选项中任
 
 ```
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1  #开启显存优化
 
 export CUDA_VISIBLE_DEVICES=0  #单卡预测
 #export CUDA_VISIBLE_DEVICES=  #CPU预测

PaddleNLP/dialogue_model_toolkit/auto_dialogue_evaluation/run.sh

 #!/bin/bash
 
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1.0
 
 export CUDA_VISIBLE_DEVICES=0
 

PaddleNLP/dialogue_model_toolkit/dialogue_general_understanding/.run_ce.sh

@@ -52,12 +52,6 @@ train_mrda(){
   --enable_ce=store_true 
 }
 
-# FIXME(zjl): this model would fail when GC is enabled,
-# but it seems that this error is from the model itself.
-# See issue here: https://github.com/PaddlePaddle/Paddle/issues/18994#event-2532039900       
-# To fix ce, disable gc in this model temporarily.  
-export FLAGS_eager_delete_tensor_gb=1
-
 cudaid=${multi:=0,1,2,3}
 export CUDA_VISIBLE_DEVICES=$cudaid
 train_atis_slot | python _ce.py

PaddleNLP/dialogue_model_toolkit/dialogue_general_understanding/README.md

@@ -177,7 +177,6 @@ export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 ```
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1  #开启显存优化
 
 export CUDA_VISIBLE_DEVICES=0  #GPU单卡训练
 #export CUDA_VISIBLE_DEVICES=0,1,2,3  #GPU多卡训练
@@ -259,7 +258,6 @@ export CUDA_VISIBLE_DEVICES=0
 
 ```
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1  #开启显存优化
 
 export CUDA_VISIBLE_DEVICES=0  #单卡预测
 #export CUDA_VISIBLE_DEVICES=  #CPU预测

PaddleNLP/dialogue_model_toolkit/dialogue_general_understanding/run.sh

 #!/bin/bash
 
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1
 
 export CUDA_VISIBLE_DEVICES=0
 if  [ ! "$CUDA_VISIBLE_DEVICES" ]

PaddleNLP/language_representations_kit/BERT/README.md

@@ -147,7 +147,6 @@ export current_endpoint=192.168.0.17:9185
 
 ```shell
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 
 BERT_BASE_PATH="chinese_L-12_H-768_A-12"
@@ -209,7 +208,6 @@ SQuAD v1.1
 
  ```shell
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1
 export CUDA_VISIBLE_DEVICES=0,1,2,3
 
 BERT_BASE_PATH="uncased_L-12_H-768_A-12"
@@ -255,7 +253,6 @@ python ${SQUAD_PATH}/evaluate-v1.1.py ${SQUAD_PATH}/dev-v1.1.json ${CHECKPOINT_P
 
 ```shell
 export FLAGS_sync_nccl_allreduce=0
-export FLAGS_eager_delete_tensor_gb=1
 export CUDA_VISIBLE_DEVICES=0,1,2,3
 BERT_BASE_PATH="uncased_L-12_H-768_A-12"
 CHECKPOINT_PATH=/path/to/save/checkpoints/

PaddleNLP/language_representations_kit/ELMo/LAC_demo/run.sh

 export FLAGS_fraction_of_gpu_memory_to_use=0.5
-export FLAGS_eager_delete_tensor_gb=0.0 
-export FLAGS_fast_eager_deletion_mode=1
 export CUDA_VISIBLE_DEVICES=0
 
 python train.py \

PaddleNLP/lexical_analysis/.run_ce.sh

 #!/bin/bash
 export FLAGS_fraction_of_gpu_memory_to_use=0.5
-export FLAGS_eager_delete_tensor_gb=0.0
-export FLAGS_fast_eager_deletion_mode=1
 
 
 train()

PaddleNLP/lexical_analysis/run.sh

 #!/bin/bash
 export FLAGS_fraction_of_gpu_memory_to_use=0.02
-export FLAGS_eager_delete_tensor_gb=0.0
-export FLAGS_fast_eager_deletion_mode=1
 export CUDA_VISIBLE_DEVICES=0,1,2,3     #   which GPU to use
 
 function run_train() {

PaddleNLP/lexical_analysis/run_ernie.sh

 #set -eux
 export FLAGS_fraction_of_gpu_memory_to_use=0.02
-export FLAGS_eager_delete_tensor_gb=0.0
-export FLAGS_fast_eager_deletion_mode=1
 # export FLAGS_sync_nccl_allreduce=1
 # export NCCL_DEBUG=INFO
 # export NCCL_IB_GID_INDEX=3

PaddleNLP/neural_machine_translation/transformer/README.md

@@ -68,8 +68,6 @@
 以提供的英德翻译数据为例，可以执行以下命令进行模型训练：
 
 ```sh
-# open garbage collection to save memory
-export FLAGS_eager_delete_tensor_gb=0.0
 # setting visible devices for training
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 
@@ -86,8 +84,6 @@ python -u main.py \
 以上命令中传入了执行训练（`do_train`）、训练轮数（`epoch`）和训练数据文件路径（注意请正确设置，支持通配符）等参数，更多参数的使用以及支持的模型超参数可以参见 `transformer.yaml` 配置文件，其中默认提供了 Transformer base model 的配置，如需调整可以在配置文件中更改或通过命令行传入（命令行传入内容将覆盖配置文件中的设置）。可以通过以下命令来训练 Transformer 论文中的 big model：
 
 ```sh
-# open garbage collection to save memory
-export FLAGS_eager_delete_tensor_gb=0.0
 # setting visible devices for training
 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 
@@ -124,8 +120,6 @@ python -u main.py \
 以英德翻译数据为例，模型训练完成后可以执行以下命令对指定文件中的文本进行翻译：
 
 ```sh
-# open garbage collection to save memory
-export FLAGS_eager_delete_tensor_gb=0.0
 # setting visible devices for prediction
 export CUDA_VISIBLE_DEVICES=0
 
@@ -145,8 +139,6 @@ python -u main.py \
  由 `predict_file` 指定的文件中文本的翻译结果会输出到 `output_file` 指定的文件。执行预测时需要设置 `init_from_params` 来给出模型所在目录，更多参数的使用可以在 `transformer.yaml` 文件中查阅注释说明并进行更改设置。注意若在执行预测时设置了模型超参数，应与模型训练时的设置一致，如若训练时使用 big model 的参数设置，则预测时对应类似如下命令：
 
 ```sh
-# open garbage collection to save memory
-export FLAGS_eager_delete_tensor_gb=0.0
 # setting visible devices for prediction
 export CUDA_VISIBLE_DEVICES=0
 

PaddleNLP/neural_machine_translation/transformer/train.py

@@ -32,8 +32,6 @@ import dist_utils
 import reader
 from transformer import create_net, position_encoding_init
 
-if os.environ.get('FLAGS_eager_delete_tensor_gb', None) is None:
-    os.environ['FLAGS_eager_delete_tensor_gb'] = '0'
 # num_trainers is used for multi-process gpu training
 num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
 

PaddleNLP/unarchived/deep_attention_matching_net/.run_ce.sh

@@ -3,7 +3,6 @@
 
 export CE_MODE_X=1
 export CUDA_VISIBLE_DEVICES=0
-export FLAGS_eager_delete_tensor_gb=0.0
 if  [ ! -e data_small.pkl ]; then
     wget -c http://dam-data.bj.bcebos.com/data_small.pkl
 fi

PaddleNLP/unarchived/deep_attention_matching_net/douban/train.sh

 export CUDA_VISIBLE_DEVICES=0
-export FLAGS_eager_delete_tensor_gb=0.0
 python -u ../train_and_evaluate.py --use_cuda \
                 --data_path ./data/data.pkl \
                 --ext_eval \

PaddleNLP/unarchived/deep_attention_matching_net/ubuntu/train.sh

 export CUDA_VISIBLE_DEVICES=0
-export FLAGS_eager_delete_tensor_gb=0.0
 python -u ../train_and_evaluate.py --use_cuda \
                 --data_path ./data/data.pkl \
                 --word_emb_init ./data/word_embedding.pkl \

PaddleNLP/unarchived/neural_machine_translation/transformer/README_cn.md

@@ -98,8 +98,6 @@ python train.py --help
 ```sh
 # 显存使用的比例，显存不足可适当增大，最大为1
 export FLAGS_fraction_of_gpu_memory_to_use=1.0
-# 显存清理的阈值，显存不足可适当减小，最小为0，为负数时不启用
-export FLAGS_eager_delete_tensor_gb=0.8
 python -u train.py \
   --src_vocab_fpath gen_data/wmt16_ende_data_bpe/vocab_all.bpe.32000 \
   --trg_vocab_fpath gen_data/wmt16_ende_data_bpe/vocab_all.bpe.32000 \

PaddleSlim/.run_ce.sh

 #!/bin/bash
 # This file is only used for continuous evaluation.
 
-export FLAGS_eager_delete_tensor_gb=0.0
 export CUDA_VISIBLE_DEVICES=3
 
 if [ ! -d 'pretrain' ]; then

PaddleSlim/classification/distillation/run.sh

@@ -18,10 +18,6 @@ fi
 
 cd -
 
-# enable GC strategy
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
-
 # for distillation
 #-----------------
 export CUDA_VISIBLE_DEVICES=0,1,2,3

PaddleSlim/classification/quantization/run.sh

@@ -30,10 +30,6 @@ fi
 
 cd -
 
-# enable GC strategy
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
-
 export CUDA_VISIBLE_DEVICES=0
 
 ## for quantization for mobilenet_v1

PaddleSlim/light_nas/run.sh

-# enable GC strategy
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
 export CUDA_VISIBLE_DEVICES=0,1,2,3
 python search.py

PaddleSlim/run.sh

@@ -24,10 +24,6 @@ fi
 
 cd -
 
-# enable GC strategy
-export FLAGS_fast_eager_deletion_mode=1
-export FLAGS_eager_delete_tensor_gb=0.0
-
 # for distillation
 #-----------------
 export CUDA_VISIBLE_DEVICES=0,1,2,3

PaddleSlim/ssd/train.py

@@ -21,20 +21,6 @@ import math
 import multiprocessing
 from paddle.fluid.contrib.slim import Compressor
 
-
-def set_paddle_flags(**kwargs):
-    for key, value in kwargs.items():
-        if os.environ.get(key, None) is None:
-            os.environ[key] = str(value)
-
-
-# NOTE(paddle-dev): All of these flags should be
-# set before `import paddle`. Otherwise, it would
-# not take any effect. 
-set_paddle_flags(
-    FLAGS_eager_delete_tensor_gb=0,  # enable GC to save memory
-)
-
 import paddle
 import paddle.fluid as fluid
 import reader