Project for the course "Advanced Deep-Learning", in the 3rd year of ESPCI, 2024.

• The data can be downloaded from this drive.
• This notebook decribes the dataset, provides explanations to the data and how to use and plot them. Moreover, at the end of this notebook, it is showed how to use a simple CNN to make a first classification.
• You can download the final report of this project following this link : https://dl.espci.fr/ticket/947d5c133b5e0973965d7235a5d881ba

To better start with, the data are pre-processed. In the directory 5-cassette you will find the data subsampled at 20 kHZ, segmented in segments of 30s each. We use only the cassette part.

We used dvc in order to track data and to put our trained models in the cloud so as to not have the data on git (in terms of privacy and lightness of the repo). First, an account needs to be created on Google Cloud Computing. Then initialize the gcp CLI while following these instructions: https://cloud.google.com/sdk/docs/install-sdk?hl=fr

If your account is authorized, you can now fetch the latest version of the data using:

dvc pull

To put data in the cloud just run:

dvc add filepath_to_data
dvc push
git add .dvc
git commit -m'added data'
git push

You can find the models for CNN and networks in the corresponding notebooks.

You can find the models in either 'BESTRq_classes/models.py' or 'models'. The trained models are saved in the models folder. The folder BESTRq_classes contains three python files to make the pretraining of the encoder for the BEST-RQ approach. All the training of this approach is on the 'random_projection_pretraining.ipynb' notebook.

We also made a python file 'compute_fft.py' which aims at computing and ploting the spectrograms along several channels as well as handling the random masking of this spectrograms.