Training the Feature Extractor

The feature extractor you just created works, in the sense that data will flow through the network and result in predictions.

But the predictions aren't accurate, because you haven't yet trained the new classification layer.

In order to do that, you'll need to read in the training dataset and train the network.

Training AlexNet (even just the final layer!) can take a little while, so if you don't have a GPU, running on a subset of the data is a good alternative. As a point of reference one epoch over the training set takes roughly 53-55 seconds with a GTX 970.

Open up train_feature_extraction.py and complete the TODO(s).