How to train on the dataset which is sequential?

[vsainteuf]

How to train on the dataset which is sequential?

[VSainteuf]

HI @manapshymyr-OB ,
can you have a look at this issue and see if it helps you?

[manapshymyr-OB]

I executed following your comments and getting this error:

Traceback (most recent call last):
  File "/agrospace/lightweight-temporal-attention-pytorch/train.py", line 363, in <module>
    main(config)
  File "/agrospace/lightweight-temporal-attention-pytorch/train.py", line 256, in main
    train_metrics = train_epoch(model, optimizer, criterion, train_loader, device=device, config=config)
  File "/agrospace/lightweight-temporal-attention-pytorch/train.py", line 26, in train_epoch
    for i, (x, y) in enumerate(data_loader):
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 632, in __next__
    data = self._next_data()
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 1350, in _next_data
    return self._process_data(data)
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 1376, in _process_data
    data.reraise()
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/_utils.py", line 694, in reraise
    raise exception
RuntimeError: Caught RuntimeError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/_utils/worker.py", line 308, in _worker_loop
    data = fetcher.fetch(index)
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 55, in fetch
    return self.collate_fn(data)
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 280, in default_collate
    return collate(batch, collate_fn_map=default_collate_fn_map)
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 143, in collate
    return [collate(samples, collate_fn_map=collate_fn_map) for samples in transposed]  # Backwards compatibility.
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 143, in <listcomp>
    return [collate(samples, collate_fn_map=collate_fn_map) for samples in transposed]  # Backwards compatibility.
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 143, in collate
    return [collate(samples, collate_fn_map=collate_fn_map) for samples in transposed]  # Backwards compatibility.
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 143, in <listcomp>
    return [collate(samples, collate_fn_map=collate_fn_map) for samples in transposed]  # Backwards compatibility.
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 120, in collate
    return collate_fn_map[elem_type](batch, collate_fn_map=collate_fn_map)
  File "/agrospace/lightweight-temporal-attention-pytorch/vevn/lib/python3.10/site-packages/torch/utils/data/_utils/collate.py", line 177, in collate_tensor_fn
    return torch.stack(batch, 0, out=out)
RuntimeError: stack expects each tensor to be equal size, but got [123, 8, 64] at entry 0 and [117, 8, 64] at entry 1

and when I print the batch shapes with the following code:

for b in batch:
        print(b.shape)
    return torch.stack(batch, 0, out=out)

I see this:

which is correct (the first number represents sequence)
Any ideas how to work correctly with the sequential data (so I do not have equal data in terms of the time)?

[VSainteuf]

yes if you have time series of varying length, you need to implement a "collate" function with a padding strategy to form batches of equal size.
You can have a look at how we did it for the PASTIS dataset here