turtle how to use already trained representation and model to evaluate different dataset?

It seems the precomputed features are specific to the original dataset. It cannot be transferred to new, unseen data using the same representation model. I have gotten the trained representation cifar100.npy and want to use on other data like aircrafts, how to do so?

Oct 17 '24 06:10 easternbun

Dear @easternbun,

If you would like to apply TURTLE on a new dataset, you first need to precompute representations for it as stated in the README, then run TURTLE using these precomputed representations.

Let me know if that has solved your issue.

Best, Artyom

Oct 21 '24 07:10 agadetsky

Thank you so much for answering my question @agadetsky

One more thing, for my dataset, do I need to input the number of classes? Could I just tell it to precompute on my data without the number of labels?

Oct 21 '24 08:10 easternbun

Dear @easternbun,

You don't need to specify the number of classes to precompute representations. However, to execute run_turtle.py you still have to have some labels, although they are just used for the evaluation purposes. You can generate some dummy labels to perform the training in case you don't have ground truth labels. See https://github.com/mlbio-epfl/turtle/issues/1 for the similar issue.

Let me know if that resolves your questions.

Best, Artyom

Oct 21 '24 08:10 agadetsky