15_autoencoders - Unsupervised pretraining is not implemented as book figure 15-8

[ageron]

15_autoencoders - Unsupervised pretraining is not implemented as book figure 15-8

[ericchencth]

hi @ageron ,
Thanks for your quick responding!
I misunderstand the code below:
pretrain_saver.restore(sess, "./my_model_cache_frozen.ckpt")
This saver is from the 'Cache the frozen layer outputs' paragraph, not from the code above this cell.
So this code is absolutely correct 👍
Thanks again!

[ageron]

Hi @ericchencth , thanks for your feedback.

Suppose you have 1 million images of handwritten digits (dataset A), plus only 55,000 labeled digit images (dataset B), and another 5,000 labeled images for the test set (dataset C).

The idea is to train the autoencoder on datasets A+B (ignoring B's labels), then reuse the lower layers of the autoencoder to build a new neural net for classification, and train this net on dataset B, then finally evaluate its performance on dataset C.

Unfortunately, I don't actually have a million unlabeled digit images to demonstrate this (I don't have dataset A), so I faked one by using B (without their labels) instead. Perhaps that's what's confusing. It might be clearer if I took out, say, 35,000 instances out of B to make A, just for the purpose of explaining this whole process. I'm not sure that it would make things clearer, however, I'm afraid that some people might wonder why I'm not using perfectly good labels for 35,000 instances.

What do you think?