You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When training the Wavenet model on the Mac with Metal Acceleration (device 'mps:0'), the training works but at the end when computing the final tensor it fails.
It works on CPU and on CUDA however.
This is an important Issue because it allows a 5x 6x acceleration.
The text was updated successfully, but these errors were encountered:
Wavenet is not running on the Mac GPU.
It runs the training and it finishes, however, when adjusting the bias at the end of the iteration the MultiOutputWrapper function throws an error.
In particular, in file multioutput.py, function _get_args_kwargs_by_output, at row 110, inside apply_to_collection(args, Tensor,.. the error is
return tuple(function(x, *args, **kwargs) for x in data)
IndexError: Dimension specified as -1 but tensor has no dimensions
Probably something that has to do on some internal definition of the metrics of the multioutput wrapper (?).
I tried to reinstall the libraries but it didn't work.
**Solution**:
since the tensors are small and on Mac there is a unified memory for GPU and CPU, the solution is to send them back to 'cpu' and then back to 'mps:0'
Dirty solution but better then nothing :)
When training the Wavenet model on the Mac with Metal Acceleration (device 'mps:0'), the training works but at the end when computing the final tensor it fails.
It works on CPU and on CUDA however.
This is an important Issue because it allows a 5x 6x acceleration.
The text was updated successfully, but these errors were encountered: