Question about splitting batches in 'dp' mode

[xiaopi-ouo]

Hello, I faced the same problem as mentioned as #1508.

    def collate_fn(self, samples):
        output = {}

        output['labels'] = [sample['labels'] for sample in samples]

Here the the samples is a list of tensor with different size.
It's unavoidable in my case. (each sample actually has different number of parts to be predicted)
So, I can't solved it by just converting them into a tensor torch.tensor([sample['labels'] for sample in samples])
Is there other solution?

Thanks.

[awaelchli]

I don't think this is the same issue as you linked.
You want to collate samples of different length into a tensor batch, right?
For this you would need some sort of padding. Perhaps you can hack it easily with
https://pytorch.org/docs/stable/generated/torch.nn.utils.rnn.pad_sequence.html#torch.nn.utils.rnn.pad_sequence
not sure, have never used it. Good luck :)

[xiaopi-ouo]

Thank you:) I finally solved it by padding manually