Extremely low accuracy after finetuning InceptionV3 (not overfitting)

[gkarampatsakis]

Hello everyone!

I'm using Keras 2.1.6 and TF 1.8. I want to fine-tune the pre-trained InceptionV3 on a new set of classes as shown here: https://keras.io/applications/

My code is literally copy-paste from the documentation. While doing that, I noticed that the accuracy on my training dataset is very high while the one on the validation set is very low. Immediately I suspected over-fitting but after wasting several hours I could not find anything. Then I tried to fit/evaluate on the same dataset. To my surprise the accuracy reported by fit() was 87.2% while the one reported by evaluate() was 62.9% on the SAME dataset.

How is this possible? Is there a bug or something? I tried also Keras 2.1.5 and 2.1.4 but I get the same problem. Am I doing something wrong? I following the documentation.

[gkarampatsakis]

After digging through the issues and googling around I found out that this is related to Batch Normalization. I tried @datumbox's temporary fix (described at #9214) and now the accuracy on training/validation is the same.

@Dref360 @farizrahman4u @taehoonlee @fchollet Any plans to fix this on master? It's really annoying!

[gkarampatsakis]

Also it seems there was a pull-request (#9965) to fix this problem but it was not merged. Anyone know what's the status of that?