search

wenxinxu / resnet-in-tensorflow

Re-implement Kaiming He's deep residual networks in tensorflow. Can be trained with cifar10.
MIT License
828 stars 276 forks source link

reuse should be 'True' in test phase. Shouldn't it? #15

Open lokhande-vishnu opened 6 years ago

lokhande-vishnu commented 6 years ago

https://github.com/wenxinxu/resnet-in-tensorflow/blob/c1ef9f4c50ad296c7a33ea3ac5c6e306e28f6d85/cifar10_train.py#L225

wenxinxu commented 6 years ago

It should be False. In line 225, the graph has not been built yet and no variables existing in the scope.

Note conceptually it's different from evaluation, where you want to use the existing graph to do the inference