hop size issue

[Paintako]

Hello,

When I train using a custom dataset, I encounter the following error with the following parameters:

"filter_length": 2048,
"hop_length": 512,
"win_length": 2048,

File "/mnt/Linux_DATA/synthesis/model/vits2_pytorch/train_ms.py", line 441, in train_and_evaluate
loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
File "/home/p76111652/.conda/envs/vits/lib/python3.8/site-packages/torch/nn/functional.py", line 3263, in l1_loss
expanded_input, expanded_target = torch.broadcast_tensors(input, target)
File "/home/p76111652/.conda/envs/vits/lib/python3.8/site-packages/torch/functional.py", line 74, in broadcast_tensors
return _VF.broadcast_tensors(tensors) # type: ignore[attr-defined]
RuntimeError: The size of tensor a (16) must match the size of tensor b (8) at non-singleton dimension 2

It seems that the shape of y_mel does not match the shape of y_hat_mel when the hop_size is increased beyond 256.
Any help would be extremely helpful, thanks!

[lexkoro]

@Paintako You also need to change the segment_size in config.