Loss history is cut off

[johschnee]

In the notebook workflow.ipynb in the branch documentation after commit 7cde490 one can see an example of consecutive loss curves during the training of the model. There is an accurate loss curve which shows approx. 200 epochs. The next plot of the loss curve shows 100 epochs and the final loss curve shows 160 epochs.

I would expect that all loss curves show the full loss history of the model, e.g. here 300 and 360 epochs.

Most likely the bug happens somewhere during the shifting of binding modes. Maybe the previous loss history is not copied over to the final model.