rebuild resnet using blocks

[brettkoonce]

No description provided.

[saeta]

This looks awesome! I'll let @jekbradbury review as well, but thank you very much.

Couple quick thoughts:

1. How does the compile time of swift-models change with the new version of ResNet? :-D
2. To simplify things a bit, could you make imageSize optional (either in this PR or a follow-on PR)? ResNet is fully convolutional and works with multiple image sizes without changing the model. Having "imageSize" be a required parameter (and a single scalar instead of a (height, width) tuple) might imply to people that the ResNet implementation only works for a single (square) image size. (Thus discouraging folks from using techniques like progressive resizing.)
3. In a follow-up PR, let's discuss making the ResNet implementation public so that folks can (in their own Swift projects or notebooks) import the canonical ResNet and use it themselves.

Thank you again. This is super awesome to see.

All the best,
-Brennan

[brettkoonce]

re 1) Not sure if this is just my system (running semi-current tf-swift master), but the existing version in models repo takes 43 seconds to compile here. Using this code takes ~3-4 seconds to compile.
re 2) Not sure how the current codebase handles arbitrary input, but certainly happy to rework things however desired.
re 3) That is fine, but having said that it would probably be good to refactor this a bit more first. If there is a way to parameterize the BasicBlock/ResidualIdentityBlock logic then this file could be cut in half again and everything reduced to a single ResNet class, which I think would be simpler for people new to computer vision. This torchvision logic is nice: https://github.com/pytorch/vision/blob/50d54a82d1479ffb6dd7469ed05fccdf290a1d84/torchvision/models/resnet.py#L216

Let me know!

[rxwei]

re 1) Not sure if this is just my system (running semi-current tf-swift master), but the existing version in models repo takes 43 seconds to compile here. Using this code takes ~3-4 seconds to compile.

This is expected because the compiler is bugged. The implementation detail is that we've turned off a part of the Swift compiler that optimizes large structures in order to suppress other bugs. We are actively working on this and expect this to be fixed within a week or two.

[jekbradbury]

2. could you make imageSize optional

This would be a good change. The reason imageSize exists at all is to distinguish between ResNets designed for CIFAR image sizes and those designed for ImageNet sizes; my impression is that the ImageNet-sized networks also work fine for images of other similar sizes.

[brettkoonce]

I can work on coming up with a way to add an input type to specify cifar, imagenet and generic variants. In the same vein, can come back to refactoring things more down the road.

[rxwei]

If kind conflicts with the first argument label, you can call it inputKind:. Relatedly, would DataKind be a better name for the type? If so, dataKind: would be a good label name. Same for other initializers that take DataKind.

[brettkoonce]

moved to: inputKind: Kind, dataKind: DataKind, let me know what you think!