Hi. I’m working on some code on NLP where I require to create specific masks. However, when i do this operation: `a[a==1] = 5`

, it tells me that `a`

and `a==1`

are not of same shape. This happens after the 8th epoch, meaning that it was working for 8 epochs but not after? Moreover, it happens randomly, maybe at the half or beginning of the training in the 8th epoch or maybe it doesn’t happen at all. If I clear the output and run the epoch again, it dosen’t happen, but happens at the following epoch. This is a showcase of my problem:

```
original_tensor = torch.randn(batch_size,max_sen_len)
tensor_c = original_tensor.detach()
max_value, _ = torch.max(tensor_c,1)
max_value = max_value.unsqueeze(1)
mask = torch.ones_like(tensor_c)
mask[tensor_c<max_value] = 0
mask[tensor_c>=max_value] = 1
mask2 = mask.clone()
values_batch = max_value.clone().squeeze(1)
mask2[mask2 == 1] = 5
```

When i get the error, it’s like this:

`shape mismatch: value tensor of shape [80] cannot be broadcast to indexing result of shape [81]`

Even if i place `assert mask2.shape == (mask2==1).shape`

, the assertion will not run, as they are correct, but PyTorch thinks they aren’t at a particular point. So I have no clue what’s wrong. Can anybody advice on this strange phenomenon?