Removing eval in model weight API

[calebrob6]

Removing use of eval.

@nilsleh -- do you remember what the use case for eval was? If not, I say we replace all instances of get_weight with get_model_weights and remove get_weight. Nevermind, I understand it now.

get_weight(...) allows a user to pass a string "ResNet18_Weights.LANDSAT_ETM_SR_MOCO" and get back the object ResNet18_Weights.LANDSAT_ETM_SR_MOCO WeightEnum. Previously we did this with eval("ResNet18_Weights.LANDSAT_ETM_SR_MOCO") which is bad. This proposed fix simply iterates through all available WeightEnums to look for matches and has the benefit of raising an error that makes sense if there is no match (vs. the eval(...) which would do whatever).

[adamjstewart]

Most of this API was copied from torchvision, let me check what their code looks like these days.

[adamjstewart]

I wonder if we can use torchvision's @register_model decorator and remove this entire file. Timm also has a way to register models, although I don't know how compatible it is. But it would let us easily use our custom models in our trainers.

[calebrob6]

I looked at the torchvision version (https://github.com/pytorch/vision/blob/main/torchvision/models/_api.py#L108) and it seems more hacky than what I've implemented here. I updated the PR description with more details.

[adamjstewart]

More lines of code, but also more secure. This is also better from a type checking perspective, as eval can return any type. We can think about how to delete this file and properly register our models in a separate PR.

[adamjstewart]

Published CVE for reference: https://www.cve.org/CVERecord?id=CVE-2024-49048

[adamjstewart]

And published GitHub advisory: GHSA-ghq9-vc6f-8qjf