ritchieng / wideresnet-tensorlayer Goto Github PK

In cifar_wide_resnet_tl.py:88 you use avgpooling to downsample the feature map, but in the original implementation it is 1x1 conv with stride =2. It is really confusing. This difference will cause performance loss. I hope you can add a warning to your Readme or change it to 1x1 conv. It will save researchers' time. Thank you.

Hi Ritchie, can you change the kwargs to fn_args in the LambdaLayer?

I am using fn_args in the release version now.

you don't write this line :sess.run(tf.global_variables_initializer()),this will cause error .