keras-team/keras-applications

Imagenet pretrained weights not match the current implementation of ResNext, thanks

junyongyou opened this issue · 2 comments

junyongyou commented

I am using ResNext50 in tensorflow.keras.application.resnet and also the ImageNet pretrained weights downloaded from https://github.com/keras-team/keras-applications/releases/download/resnet/resnext50_weights_tf_dim_ordering_tf_kernels_notop.h5. However, it seems the latest ResNext implementation does not match the weights. For example, if no top layers are included, an error "a weight file containing 106 layers into a model with 122 layers" is thrown. Does anybody have ImageNet pretrained weights matching the current implementation of ResNext50? Thanks a lot.

chqiwang commented

@junyongyou Did you solve this problem? I also have this problem.

junyongyou commented

@junyongyou Did you solve this problem? I also have this problem.

No. I didn't and I didn't work on this since then.