How to implement transfer learning in unseen dataset .? about super-gradients HOT 8 CLOSED

Join the discussion on DagsHub!

from super-gradients.

Hello, could you please rephrase your question?

from super-gradients.

@NatanBagrov Lets my 1st model train on A ,B,C class...In next time i want to train My pre-train 1st model by adding D class ...(Note: In second time training I do not want to mix all data...i just want to train new D class). Final model must detect A, B, C, D class.

from super-gradients.

I understand.

Please close the duplicate issue you created, #922 .

To achieve that you will need to modify the head to include additional new classes (=kernels), in your example, it has original 3 kernels, then you need to change to 4 kernels. Also, I would freeze the entire backbone, neck and original part of the head, and fine-tune only the "new" kernels.

This is somehow similar to #892

from super-gradients.

@NatanBagrov Can you guide me How to modify the head to include additional new classes (=kernels) i found only kernel_size=3 ...i have to change self.num_classes am i right ? is there any config yml file to change self.num_classes from yml file ?

Also how to freeze the entire backbone, neck and original part of the head can you guide me i am new here from where i have to freeze dynamically....shall we need to unfeeze also?

And how to finetune only the "new" kernels class after freeze the entire backbone, neck and original part of the head?

I am new here guide me with the help of code point i would be a great help.

from super-gradients.

Hello, I will elaborate with general guidelines, and will leave the implementation to you.

• If you change number of classes, you will not be able to re-use the checkpoint, so you need to do the modification after you have loaded the pre-trained checkpoint with the original number of classes.
• I would create a wrapper, that will take the model (already loaded), and will expand cls_preds with the extended filters (=kernels), that is, use the filters of the original classes, but also add uninitialized ones for the new classes. Then it is fine to change num_classes.
• To freeze parts of the model, you can read Torch documentation on how to achieve that. Here's a reference to a forum thread with some info.

Good luck!

from super-gradients.

@NatanBagrov after freeze the entire backbone, neck and original part of the head....how my fine-tune "new" kernels learns without backbone, neck and original part of the head ? Shall i need to add some layer for that?

from super-gradients.

You have the suggested fine-tune recipes in our documentation. You just launch training. The backbone and the neck already learned to extract valuable features, so only your "new" filters need to learn how to "classify" the new classes based on these features.

from super-gradients.