Add list to torch.Tensor injection in yaml config

[fguiotte]

Description & Motivation

I'd like to configure the weight of the loss in the config.yaml.

Pitch

For example with this configuration file:

model:
  out_channels: 3
  criterion:
    class_path: torch.nn.CrossEntropyLoss
    init_args:
      reduction: 'mean'
      weight: [0.0, 1.0, 1.0]

The weight parameter should be a torch.Tensor and not a Python list. So with my current config, instead of the loss instance, it returns a dictionary {'class_path': 'torch.nn.CrossEntropyLoss', 'init_args': {'reduction': 'sum', 'weight': [0.0, 1.0, 1.0]}} for the criterion parameter. If we remove the weight parameter in the example configuration, it creates the loss instance as expected.

It would be helpful to have a Python list to `torch.Tensor injection to automatically create the tensors when needed.

Alternatives

In the mean time, I can create the loss myself in the model __init__ with the returned dictionary.

Additional context

No response

cc @Borda