namedtuple is incompatible with num_workers>0 in PyTorch DataLoader

[KonstantinWilleke]

The __getitem__ of the datasets can not handle a namedtuple when there are multiple parallel workers. The parallel workers are needed to reach high data loading speeds for powerful compute nodes. This is a fundamental pytorch issue: because each worker instantiates a Dataset object, and the namedtuple is instantiated in each one, the parallel workers can't collate the batches in a "custom" named tuple.

Potential workarounds:

• write a custom collate_fn for the DataLoader which converts the entire batch into a namedtuple.
• each batch returns a dictionary, as in {device_name: value_tensor}

[pollytur]

Another potential solution is to do same as in neuralpredictors and define the namedtuple globally, then it works with the num_workers>0