support the cutoff training regime

[bertsky]

How about aiding users in cut-off training, besides finetuning and from scratch?

We already have START_MODEL and NET_SPEC, so IIUC we would minimally only need to add some additional variable, say KEEP_LAYERS, for the --append_index parameter. If undefined, everything is kept (i.e. START_MODEL or NET_SPEC is reused as-is, which amounts to finetuning or from-scratch). But if defined, it would add --append_index $(KEEP_LAYERS) (and thus keep only that many layers of START_MODEL and then append new layers from NET_SPEC).

Of course, additional facilities like extracting a traineddata's net-spec, or preconfigured KEEP_LAYERS/NET_SPEC settings for typical scenarios could be offered, too.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.