# Multilabel sequence classification with Roberta value error expected input batch size to match target batch size

**URL:** <https://discuss.huggingface.co/t/multilabel-sequence-classification-with-roberta-value-error-expected-input-batch-size-to-match-target-batch-size/1653>\
**Category:** 🤗Transformers\
**Created:** [October 20, 2020, 1:52pm UTC](https://discuss.huggingface.co/t/multilabel-sequence-classification-with-roberta-value-error-expected-input-batch-size-to-match-target-batch-size/1653 "2020-10-20T13:52:27Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![laurb](https://avatars.discourse-cdn.com/v4/letter/l/e47c2d/32.png) [@laurb](https://discuss.huggingface.co/u/laurb)\
**Post date:** [October 20, 2020, 1:52pm UTC](https://discuss.huggingface.co/t/multilabel-sequence-classification-with-roberta-value-error-expected-input-batch-size-to-match-target-batch-size/1653/1 "2020-10-20T13:52:27Z")

</div>

Trying to tune a multilabel (4 labels) model based on roberta-base. I’ve followed the examples in [https://huggingface.co/transformers/custom\_datasets.html](https://huggingface.co/transformers/custom_datasets.html).

Trying to debug this value error:  
Traceback (most recent call last):  
trainer.train()  
File “transformers/trainer.py”, line 762, in train  
tr\_loss += self.training\_step(model, inputs)  
File “transformers/trainer.py”, line 1112, in training\_step  
loss = self.compute\_loss(model, inputs)  
File “transformers/trainer.py”, line 1136, in compute\_loss  
outputs = model(\*\*inputs)  
File “torch/nn/modules/module.py”, line 532, in **call**  
result = self.forward(\*input, \*\*kwargs)  
File “transformers/modeling\_roberta.py”, line 1015, in forward  
loss = loss\_fct(logits.view(-1, self.num\_labels), labels.view(-1))  
File “torch/nn/modules/module.py”, line 532, in **call**  
result = self.forward(\*input, \*\*kwargs)  
File “torch/nn/modules/loss.py”, line 916, in forward  
ignore\_index=self.ignore\_index, reduction=self.reduction)  
File “torch/nn/functional.py”, line 2021, in cross\_entropy  
return nll\_loss(log\_softmax(input, 1), target, weight, None, ignore\_index, None, reduction)  
File “torch/nn/functional.py”, line 1836, in nll\_loss  
.format(input.size(0), target.size(0)))  
ValueError: Expected input batch\_size (16) to match target batch\_size (64).

I see this in modeling\_roberta at the point of error. I looks like the labels for each of the batch results have been flattened into a single tensor, while the batch has the labels separately for each example of the 16. Seems like this might be the cause of the ValueError? but I’m not sure, and don’t know where the labels would have been flattened. Any ideas?  
tensor([[0.1793, 0.1338, -0.2123, -0.0945],  
[0.0498, 0.0472, -0.1983, -0.0353],  
[0.1932, 0.1970, -0.2003, -0.0471],  
[0.0913, 0.1411, -0.1835, -0.1387],  
[0.0770, -0.0101, -0.1017, -0.0149],  
[0.1980, 0.0772, -0.1894, -0.0487],  
[0.0161, 0.0107, -0.0100, 0.0067],  
[0.1063, 0.1120, -0.1842, -0.0567],  
[0.1610, 0.0769, -0.1609, -0.0883],  
[0.1866, 0.0182, -0.1137, -0.1047],  
[0.1132, 0.0587, -0.2452, -0.0698],  
[0.1680, -0.0125, -0.2019, -0.0674],  
[-0.0282, 0.1099, -0.1637, -0.1112],  
[0.1620, 0.1197, -0.2099, 0.0236],  
[0.1197, 0.1232, -0.2318, -0.0955],  
[0.3232, 0.1935, -0.3226, -0.0547]], device=‘cuda:0’,  
grad\_fn=)  
labels view  
tensor([0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0,  
0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1,  
0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0], device=‘cuda:0’)

---

<div class="post-metadata">

**Author:** ![jwa018](https://avatars.discourse-cdn.com/v4/letter/j/9de0a6/32.png) [@jwa018](https://discuss.huggingface.co/u/jwa018)\
**Post date:** [March 2, 2021, 1:05pm UTC](https://discuss.huggingface.co/t/multilabel-sequence-classification-with-roberta-value-error-expected-input-batch-size-to-match-target-batch-size/1653/2 "2021-03-02T13:05:29Z")

</div>

I had the same problem. The problem lies in nll\_loss. For multilabel problems BCEWithLogitsLoss is the most common I think. You can subclass Trainer and overwrite the compute\_loss function in your custom trainer to make things work. This worked for me:

```python

    class CustomTrainer(Trainer):
        def compute_loss(self, model, inputs, return_outputs=False):
            outputs = model(
                input_ids=inputs['input_ids'],
                attention_mask=inputs['attention_mask'],
                token_type_ids=inputs['token_type_ids']
            )
            loss = th.nn.BCEWithLogitsLoss()(outputs['logits'],
                                             inputs['labels'])
            return (loss, outputs) if return_outputs else loss

```