To me, the new file structure looks great. After the cyclic import issue is fixed, we can merge this PR.

Thank you @lucadellalib for this PR!
Here are my comments:

1. recipe tests are failing (recipes/ESC50/interpret/hparams/amt_focalnet.yaml, recipes/ESC50/interpret/hparams/amt_vit.yaml, recipes/ESC50/interpret/hparams/piq_focalnet.yaml, recipes/ESC50/interpret/hparams/piq_vit.yaml) => "ValueError: 'embedding_model_path' is a !PLACEHOLDER and must be replaced." Please, run them with:
   python -c 'from tests.utils.recipe_tests import run_recipe_tests; print("TEST FAILED!") if not(run_recipe_tests(filters_fields=["Dataset"], filters=[["ESC50"]], do_checks=True, run_opts="--device=cuda")) else print("TEST PASSED")'

2. Please, share the models with me such that I can upload them on the official dropbox.

3. Just like the other methods, I recommend putting the model on HF as well. This way, users can directly run the interpretation script (nterpret_amt.py) before training the classifier.

Thank you @lucadellalib for this PR. Everything now works properly.