Error when working with multimodal data: 'dict' object has no attribute 'to'

[spartanjoax]

🐛 Describe the bug I am working with a multimodal signal dataset. It is composed of time-series data (signals) and structured data (process parameters). The samples are returned as a dict with two keys (x and proc_data), which are then fed to the model. I am getting an error during training, as Avalanche attempts to move the minibatches to the GPU device. The error I am getting is the following:

File "/workspace/example/continual_training.py", line 325, in main strategy.train(scenario.train_stream) File "/opt/conda/lib/python3.11/site-packages/avalanche/training/supervised/joint_training.py", line 152, in train self._before_training(**kwargs) File "/opt/conda/lib/python3.11/site-packages/avalanche/training/templates/base.py", line 326, in _before_training trigger_plugins(self, "before_training", **kwargs) File "/opt/conda/lib/python3.11/site-packages/avalanche/training/utils.py", line 75, in trigger_plugins getattr(p, event)(strategy, **kwargs) File "/opt/conda/lib/python3.11/site-packages/avalanche/training/templates/base_sgd.py", line 616, in before_training self._peval(strategy, **kwargs) File "/opt/conda/lib/python3.11/site-packages/avalanche/training/templates/base_sgd.py", line 637, in _peval strategy.eval(el, **kwargs) File "/opt/conda/lib/python3.11/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/lib/python3.11/site-packages/avalanche/training/templates/base_sgd.py", line 228, in eval super().eval(exp_list, **kwargs) File "/opt/conda/lib/python3.11/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/lib/python3.11/site-packages/avalanche/training/templates/base.py", line 212, in eval self._eval_exp(**kwargs) File "/opt/conda/lib/python3.11/site-packages/avalanche/training/templates/base_sgd.py", line 232, in _eval_exp self.eval_epoch(**kwargs) File "/opt/conda/lib/python3.11/site-packages/avalanche/training/templates/base_sgd.py", line 270, in eval_epoch self._unpack_minibatch() File "/opt/conda/lib/python3.11/site-packages/avalanche/training/templates/problem_type/supervised_problem.py", line 63, in _unpack_minibatch mbatch[i] = mbatch[i].to(self.device, non_blocking=True) # type: ignore ^^^^^^^^^^^^ AttributeError: 'dict' object has no attribute 'to'

🐜 To Reproduce The following repository has a minimal example to reproduce the error: https://github.com/spartanjoax/continual_example. It can be run with the following command:

python example.py

🐝 Expected behavior I would expect to be able to work with multimodal data. I would appreciate it if you could provide me with some guidance on how to edit the code so that it can do something like the following:

if isinstance(mbatch[i], dict):
    # Move all tensor values in the dictionary to the specified device
    mbatch[i] = {key: value.to(self.device, non_blocking=True) for key, value in mbatch[i].items()}
elif isinstance(mbatch[i], torch.Tensor):
    # Directly move the tensor to the specified device
    mbatch[i] = mbatch[i].to(self.device, non_blocking=True)
else:
    # Optionally handle other cases (e.g., raise an error if unsupported types are encountered)
    raise TypeError(f"Unsupported type in batch: {type(mbatch[i])}")