Input batch size does not match target batch size when using more than one positive context when training DPR

[Raldir]

Describe the bug When fine-tuning DPR on a custom dataset with parameter num_positives being larger than one, I get a batch size mismatch error when starting training. Training works perfectly fine when num_positives=1. The list of 'positive_ctxs' is set to always be equal to num_postives. I have also tried out the default dataset (i.e. biencoder-nq) -- same issue. So when I train with a batch size of 16 and num_positives=2 I get the error below.

Error message File "/home/XXX/miniconda3/envs/python3.7/lib/python3.7/site-packages/torch/nn/modules/loss.py", line 213, in forward return F.nll_loss(input, target, weight=self.weight, ignore_index=self.ignore_index, reduction=self.reduction) File "/home/XXX/miniconda3/envs/python3.7/lib/python3.7/site-packages/torch/nn/functional.py", line 2262, in nll_loss .format(input.size(0), target.size(0))) ValueError: Expected input batch_size (16) to match target batch_size (32).

Additional context Code is essentially identical to the tutorial. I am using the latest version of the repository.

[ottozastrow]

I have the same issue. Also using code from tutorial. only works for num_positives=1. Did you find a way to solve the issue? error still occurs when using the json datasets from the tutorial

[ottozastrow]

thanks for reopening - happy to provide more info, but since its the code from the tutorial I assume thats all you need

[sjrl]

Hi @ottozastrow just to confirm you are referring to this tutorial Tutorial9_DPR_training.ipynb?

And could you tell me which version of Haystack are you using?

[ottozastrow]

sure - tested with 1.5.0 and 1.6.0 and occurs on both and yes that is the tutorial I followed (though locally not on colab) the link I used for the tutorial is this

  File "/home/otto/anaconda3/lib/python3.9/site-packages/haystack/modeling/training/base.py", line 291, in train
    loss = self.compute_loss(batch, step)
  File "/home/otto/anaconda3/lib/python3.9/site-packages/haystack/modeling/training/base.py", line 376, in compute_loss
    per_sample_loss = self.model.logits_to_loss(logits=logits, global_step=self.global_step, **batch)
  File "/home/otto/anaconda3/lib/python3.9/site-packages/haystack/modeling/model/biadaptive_model.py", line 194, in logits_to_loss
    all_losses = self.logits_to_loss_per_head(logits, **kwargs)
  File "/home/otto/anaconda3/lib/python3.9/site-packages/haystack/modeling/model/biadaptive_model.py", line 181, in logits_to_loss_per_head
    all_losses.append(head.logits_to_loss(logits=logits_for_one_head, **kwargs))
  File "/home/otto/anaconda3/lib/python3.9/site-packages/haystack/modeling/model/prediction_head.py", line 1071, in logits_to_loss
    loss = self.loss_fct(softmax_scores, targets)
  File "/home/otto/anaconda3/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/otto/anaconda3/lib/python3.9/site-packages/torch/nn/modules/loss.py", line 211, in forward
    return F.nll_loss(input, target, weight=self.weight, ignore_index=self.ignore_index, reduction=self.reduction)
  File "/home/otto/anaconda3/lib/python3.9/site-packages/torch/nn/functional.py", line 2532, in nll_loss
    return torch._C._nn.nll_loss_nd(input, target, weight, _Reduction.get_enum(reduction), ignore_index)
ValueError: Expected input batch_size (8) to match target batch_size (32).
``

[sjrl]

Hi @ottozastrow. Unfortunately, this doesn't look like a quick fix. I'll be adding it to our sprint backlog so we can dedicate time to this issue in the near future.

[ottozastrow]

ok thanks! (just incase this is helpful: for my usecase DPR with only 1 positive is not usable)

[julian-risch]

Maybe @bogdankostic knows what's causing the problem or even has an idea for quick fix? He's a DPR expert! 🙂