To train e.g. SequenceModelTree one needs a specific type of combined dataloaders. These are currently built from the task dicts which are attached to the model dynamically on self.build_tree.
Would it be feasible to create these combined dataloaders from numpy arrays? I imagine something like
train_dataloader, val_dataloader = model.get_dataloaders_from_numpy(x, y, **other_kwargs)
where x.shape == [dataset_size, sequence_length] and y.shape = [dataset_size, num_tasks], and the sizes of the dataloaders, batch sizes, and so on are specified through the **other_kwargs.
To train e.g.
SequenceModelTree
one needs a specific type of combined dataloaders. These are currently built from the task dicts which are attached to the model dynamically onself.build_tree
.Would it be feasible to create these combined dataloaders from
numpy
arrays? I imagine something likewhere
x.shape == [dataset_size, sequence_length]
andy.shape = [dataset_size, num_tasks]
, and the sizes of the dataloaders, batch sizes, and so on are specified through the**other_kwargs
.