In pytorch 0.5 0-dim tensors will be an error rather than a warning

[whackashoe]

This replaces 0-dim tensors with item() to get a Python number from the tensors containing a single value.