Implement mnist-cpc training in pytorch

[jusjusjus]

We add '/torch_train_model.py' that reproduces the training in '/train_model.py'. Each element in the cpc network are wrapped in torch.nn.Module subclasses:

• NetworkEncoder : Transforms a batch of images into an encoded representation with code_size elements.
• Autoregressive : Processes a series of encoded images to summarize their content in a context vector of length hidden_size.
• Predictor : Predict a number future encodings given a context vector.

These are in turn wrapped into a CPCNetwork convenience class to manage training.

[martinmamql]

Hi jusjusjus,

Greetings! Sincerely appreciate your great help for writing up this pytorch version!

I have tried to reproduce using this version, but failed to do so. After batch 1000, the train loss is still 0.69 and accuracy is around 0.5. I just typed python torch_train_model.py. Did I do something wrong?

Thank you so much!

Sincerely, Martin