fewfox: creating an efficient model using very few data

Setup

Install the requirements

To install the packages we use the poetry package manager.

poetry install

Get some data

At the moment, the pipeline works with the ESC50 dataset. You can download it using the script /data/get_ESC50.sh. The ESC50 dataset is composed of 50 semantical classes with 40 sounds per class.

Once you have downloaded the ESC50 dataset, we can create a "miniESC50" dataset. The miniESC50 dataset is composed of 5 classes. Each class contains 5 sounds for training, 5 sounds for validating and the remaining sounds (30) for testing the model.

After setting the correct MINI_ESC50_PATH in configs/paths/default.yaml, create the miniESC50 dataset using the following command:

poetry run python miniesc50.py 

Change the config

The most important is to adapt /configs/paths/default.yaml using your paths

Run the pipeline

You can train the prototypical model using the following command:

poetry run python src/protopipeline.py

The model weights will be stored in the lightning_logs folder.

Evaluate your model

You can evaluate the performance of your model using the command:

poetry run python src/evaluate.py

This should return model performance and a 2D image displaying the embeddings and the prototypes