Create python implementation of torch.nn.Conv2d and torch.nn.AvgPool2d

[iamtrask]

For the same reasons we have python implementations of LSTM and GRU, I have created a basic implementation of Conv2d and AvgPool2d

It does not include all functionality, but it does include assertions to cause errors for settings which are not supported.

[Jasopaum]

We already have our own conv2d implementation, it’s not in the nn module but in the FixedPrecisionTensor class, I think. It was implemented before the rnn so it’s not at the right place (I know we should have moved it before).
Maybe it can be improved but you can at least start from there 🙂

[iamtrask]

Hi @Jasopaum - upon further review of that class it does indeed look like a convolutional layer exists there. However, unbeknownst to me the class has an enormous amount of functionality which is unrelated to FixedPrecision representations. I suspect that this has been done for coding convenience and possibly justified for performance reasons. If so, this is done at the expense of code re-use and good organization and is in urgent need of refactoring.

At present, I will not refer to it and have created a Github issue describing extensive refactoring required (#2897). Once the refactoring is completed, we can adjust the classes created in this PR to reference wherever the new functionality exists.

[Jasopaum]

I do agree that a refactor would do some good! 👍

[karlhigley]

Looks good, maybe a few things to clean up before merging though.