Fix callback in multiartifact CNN fine-tuning

[markus-hinsche]

Problem: We see that when fine tuning starts, the loss jumps up.

Details:

• Nadam has a LR per layer
• when we fine-tune (we make base network's layers trainable), then these LR might be too big

Solution (besides considering a smaller LR)

Task: Improve code:

• Add callback for unfreezing (e.g. after 1000 layers) instead of calling model.fit() again.
• This will have the effect that the Adam statistics (momentum etc) can be continued to use