未验证 提交 ff7af219 编写于 作者: H Huihuang Zheng 提交者: GitHub

Fix a Cudnn Deterministic Random Failure, test=develop (#25389)

As the title
上级 6de75082
...@@ -22,6 +22,7 @@ from simple_nets import init_data ...@@ -22,6 +22,7 @@ from simple_nets import init_data
import math import math
import os import os
os.environ['CPU_NUM'] = str(4) os.environ['CPU_NUM'] = str(4)
os.environ['FLAGS_cudnn_deterministic'] = str(1)
# FIXME(zcd): If the neural net has dropout_op, the output of ParallelExecutor # FIXME(zcd): If the neural net has dropout_op, the output of ParallelExecutor
# and Executor is different. Because, for ParallelExecutor, the dropout_op of # and Executor is different. Because, for ParallelExecutor, the dropout_op of
...@@ -35,14 +36,10 @@ remove_dropout = False ...@@ -35,14 +36,10 @@ remove_dropout = False
# and Executor is different. # and Executor is different.
remove_bn = False remove_bn = False
# FIXME(huihuangzheng): Temporarily disable cudnn of conv2d in unit test because
# it will cause random test failure. We have to re-enable it after someone fixs
# cudnn_conv
remove_cudnn_conv = False remove_cudnn_conv = False
remove_dropout = True remove_dropout = True
remove_bn = True remove_bn = True
remove_cudnn_conv = True
def squeeze_excitation(input, num_channels, reduction_ratio): def squeeze_excitation(input, num_channels, reduction_ratio):
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册