mrpep / tp-picml

Trabajo final del curso de Neuro/ML a gran escala

Integrantes:

• Leonardo Pepino
• Joaquín González
• Agustín Sansone

Instructions:

0) Initial setup

Clone this repository and we also recommend executing this script from original repository to download dataset.

1) Extract activations from EnCodecMAE models

There are 2 options:

a) Clone and install EnCodecMAE repository. Then run:

bash extract_activations.sh

This will extract the activations in a model_actv folder.

b) Download precomputed activations from EnCodecMAE and other models from here

2) Run regressions

⚠️ Warning this might take a long time (around 3 days with a 24-core CPU). If you want faster results, calculating only components will be faster (should take a few minutes). Alternatively, precomputed results can be downloaded from here

bash run_regressions.sh

Use this script to replicate regularization parameters of the Ridge and Lasso regressions when predicting all components for EncodecMAE trained in the mixture context.

3) Generate plots

Plots can be generated using the notebooks in /src