Is your feature request related to a problem? Please describe.

Currently there's no way to monitor how individual chanels and loss functions are changing during training. This is critical for debugging multi-stream models during training (e.g., isolating unstable channels, finiding channels that perform worse). We can make this optional via a flag to reduce clutter. This was part of issue #29

Describe the solution you'd like

Loss is computed individually here and averaged over channels to be stored in losses_all.

The solution is to add another dimension to losses_all and save the individual losses there. As this array is used anyway for logging, minimal changes is required within the code. This change would fully backwards compatible.

Additional context

No response

Organisation

AWI