Transfer learning

[mrocamora]

Another feature we would need is transfer learning, for adapting a model to our particular purposes.

In particular, we would like to train a model, for instance, SB_CNN with URBAN_SED. Then, we would like to remove the last fully connected layers and replace them with new fully connected layers. Finally, we would like to train these layers with a certain dataset (e.g. recordings of the sound classes we want to detect), while keeping the weights fixed for the rest of the model.

I am willing to discuss the different possible ways of implementing this, if you agree.

Thanks!

[pzinemanas]

Hi Martín, nice to hear from you.

Thanks for proposing new features to improve the library. I'm working on this, I added a fine_tuning function in the DCASEModelContainer class. I'm also adding tests/test_fine_tuning.py script as an example on how to use the new function. Please check it out and let me know your opinions.

Regards,
Pablo