About the loss calculation in audio.

[musicrainie]

Maybe this line

data2vec-pytorch/audio/trainer.py

Line 61 in 61c0aa8

loss = self.criterion(x.float(), y.float()).sum(dim=-1).div(x.size(0))

should be fixed to
loss = self.criterion(x.float(), y.float()).sum(dim=-1).sum().div(x.size(0))
to convert MATRIX to SCALAR, just like the text and vision case:

data2vec-pytorch/text/trainer.py

Line 69 in 61c0aa8

loss = self.criterion(x.float(), y.float()).sum(dim=-1).sum().div(x.size(0))

data2vec-pytorch/vision/trainer.py

Line 57 in 61c0aa8

loss = self.criterion(x.float(), y.float()).sum(dim=-1).sum().div(x.size(0))

.