Question) Fine tunning a transfer learned model

[SyedUmairHassanKazmi]

I know this is not a forum for help but I couldn't find any forum where I could reach for help.
I have 2 questions.
Q1) How can I fine tune a transfer learned model by unfreezing some layer , or adding new layers and retraining the model in Darts?
Q2) Is their any functionality through which I can plot train and val accuracy with no of epochs. Also something like tensorboard?
Thanks

[madtoinou]

Hi,

Q1)

1. A PR covering the loading of the weights from a checkpoint to a freshly instantiated model (with new optimizer/learning rate scheduler, discussion can be found here) has been merged yesterday and should cover the retraining aspect.
2. Darts does not provide a lot of flexibility as for the architecture of a model (new layer, custom activation function) and you would have to implement a new model if you want to be able to add layers (especially after pre-training the model).
3. As far as I know, it should be possible to (un)freeze layers of darts model as in any PyTorch model.

Q2)

1. By default, darts models are already generating tensor-board compatible logs (using PyTorch Lightning) in the 'darts_logs/model_name/logs/' folder that can be visualize with the app. You tweak this using the logger keyword of the pl_trainer_kwargs argument, in the models' constructor.

Writing a tutorial notebook for "advanced" model fine-tuning is in the backlog, we are trying to focus on more urgent bug/feature for the moment. If you want to give it a try, contributions are welcome!

PS: You could try to ask your questions on the project gitter, which can be considered as some kind of forum.