Pytorch-adapt

Many changes, may update this with details

Debugging messages are appended to the traceback when an exception occurs inside a hook (24).

For example:

Old behavior:

Traceback (most recent call last):
...
TypeError: forward() takes 2 positional arguments but 3 were given


New behavior:

Traceback (most recent call last):
...
TypeError: in GVBHook: __call__
in ChainHook: __call__
in OptimizerHook: __call__
in ChainHook: __call__
in FeaturesLogitsAndGBridge: __call__
in GBridgeAndLogitsHook: __call__
GBridgeAndLogitsHook: Getting src
GBridgeAndLogitsHook: Getting output: ['src_imgs_features_logits', 'src_imgs_features_gbridge']
GBridgeAndLogitsHook: Using model C with inputs: src_imgs_features, return_bridge
forward() takes 2 positional arguments but 3 were given
C.forward() signature is (input: torch.Tensor) -> torch.Tensor

Swapped order of input and output argument of hooks.

| Before | After |
| - | - |
losses, output = hook(losses, inputs) | output, losses = hook(inputs, losses)

The loss input argument is now optional, which makes the top level syntax cleaner:
python
old
hook({}, {**models, **data})

new
hook({**models, **data})