Data reshape issue in _process_signal

[azajala]

At the beginning of this process the data is transpose to have T at the end of the tensor. Not sure how it works for 1D and spectrograms. But for videos, the input shape is (T,H,W) while the output becomes (W,H,T) which leads to error in training.
As a bandaid fix I transpose the output around lines 712 and 744 of data loader. But we needs a more permanent fix.

[renierts]

The input shape as returned from the dataset class is (n_frames, height, width).
The autoencoder has to reconstruct the same shape.
I checked this for both, reconstruction and prediction. In the case of IRTV, the shape is (n_frames, 513, 640).

Can you verify the model outputs?

[renierts]

As we discussed: We are going to call self._apply_processing(...) after loading the raw movie instead of _process_signal(...).