[Contribution] Convert cnn output from 'flatten' to 'avg_pool'

[PavlosMelissinos]

Current version of Inception_v3 doesn't have a layer named 'flatten' so the code crashes when extract_image_features is set to True.

In this older version, it's clear that the 'flatten' layer used to be the one right before the dense one at the end, therefore I suggest that L232 of train.py be replaced with output=base_model.get_layer('avg_pool').output)!

Not sure the change is worth a PR to be honest...

[jmtatsch]

I also tried that, however the extracted features with keras 2.0.8 are different from the downloaded inception_image_name_to_features.h5.
Mine are 10mb smaller (maybe a newer hdf5 lib?), the weights are visibly different for the first image and also when training the convergence is worse:
best val_loss 2.01692 after 160 epochs instead of 1.96 after 90 epochs.

[PavlosMelissinos]

I can confirm your observations, @jmtatsch. Have you come up with a better solution?

[oarriaga]

Maybe you can change the CNN feature extractor. You can try using the Xception network also available in keras applications.

[jmtatsch]

I tried ResNet50 already, it performed worse, Xception may well be worth a try.
I just don't understand the performance drop with the inception weights. Your inception_image_name_to_features.h5 was processed by inception v3, correct?