Move benchmark changing of cycle gan to models/develop (#2409)

* move benchmark changing of cycle gan to models/develop, test=develop

* add flag exist check, test=develop

PaddleCV/gan/cycle_gan/layers.py

@@ -4,10 +4,9 @@ import numpy as np
 import os
 
 # cudnn is not better when batch size is 1.
-use_cudnn = False
-if 'ce_mode' in os.environ:
-    use_cudnn = False
-
+use_cudnn_conv2d_transpose = False
+use_cudnn_conv2d = True
+use_layer_norm = True
 
 def cal_padding(img_size, stride, filter_size, dilation=1):
     """Calculate padding size."""
@@ -21,7 +20,9 @@ def cal_padding(img_size, stride, filter_size, dilation=1):
 
 def instance_norm(input, name=None):
     # TODO(lvmengsi@baidu.com): Check the accuracy when using fluid.layers.layer_norm.
-    # return fluid.layers.layer_norm(input, begin_norm_axis=2) 
+    if use_layer_norm:
+        return fluid.layers.layer_norm(input, begin_norm_axis=2) 
+
     helper = fluid.layer_helper.LayerHelper("instance_norm", **locals())
     dtype = helper.input_dtype()
     epsilon = 1e-5
@@ -90,7 +91,7 @@ def conv2d(input,
         name=name,
         stride=stride,
         padding=padding,
-        use_cudnn=use_cudnn,
+        use_cudnn=use_cudnn_conv2d,
         param_attr=param_attr,
         bias_attr=bias_attr)
     if need_crop:
@@ -145,7 +146,7 @@ def deconv2d(input,
         filter_size=filter_size,
         stride=stride,
         padding=padding,
-        use_cudnn=use_cudnn,
+        use_cudnn=use_cudnn_conv2d_transpose,
         param_attr=param_attr,
         bias_attr=bias_attr)
 

PaddleCV/gan/cycle_gan/train.py

@@ -2,6 +2,28 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 import os
+
+
+def set_paddle_flags(flags):
+    for key, value in flags.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_cudnn_exhaustive_search': 1,
+    '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
+})
+
 import random
 import sys
 import paddle
@@ -150,18 +172,30 @@ def train(args):
     build_strategy.enable_inplace = False
     build_strategy.memory_optimize = False
 
+    exec_strategy = fluid.ExecutionStrategy()
+    exec_strategy.num_threads = 1
+    exec_strategy.use_experimental_executor = True
+
     g_A_trainer_program = fluid.CompiledProgram(
         g_A_trainer.program).with_data_parallel(
-            loss_name=g_A_trainer.g_loss_A.name, build_strategy=build_strategy)
+            loss_name=g_A_trainer.g_loss_A.name,
+            build_strategy=build_strategy,
+            exec_strategy=exec_strategy)
     g_B_trainer_program = fluid.CompiledProgram(
         g_B_trainer.program).with_data_parallel(
-            loss_name=g_B_trainer.g_loss_B.name, build_strategy=build_strategy)
+            loss_name=g_B_trainer.g_loss_B.name,
+            build_strategy=build_strategy,
+            exec_strategy=exec_strategy)
     d_B_trainer_program = fluid.CompiledProgram(
         d_B_trainer.program).with_data_parallel(
-            loss_name=d_B_trainer.d_loss_B.name, build_strategy=build_strategy)
+            loss_name=d_B_trainer.d_loss_B.name,
+            build_strategy=build_strategy,
+            exec_strategy=exec_strategy)
     d_A_trainer_program = fluid.CompiledProgram(
         d_A_trainer.program).with_data_parallel(
-            loss_name=d_A_trainer.d_loss_A.name, build_strategy=build_strategy)
+            loss_name=d_A_trainer.d_loss_A.name,
+            build_strategy=build_strategy,
+            exec_strategy=exec_strategy)
     for epoch in range(args.epoch):
         batch_id = 0
         for i in range(max_images_num):