Noise Inspector for Stochastic Gradient Descent

Author: Guney Tombak ([email protected])
Supervisor: Mr. Thomas Allard
Professor: Dr. Helmut Bölcskei
Chair for Mathematical Information Science, D-ITET, ETH Zurich

Setup

All dependencies can be fulfilled by creating a conda environment using environment.yml

conda env create -f environment.yml && conda activate sgdn

For GPU implementation use also:

conda install pytorch torchvision torchaudio cudatoolkit=X.Y -c pytorch -c conda-forge

with a cuda version X.Y. compatible with your GPU.

Before running, you should log in to your wandb (Weights and Biases) account:

wandb login

Usage

The parameters of the run can be configured by editing config.py.
The parameters as a tuple creates a new branch for the search tree.

Results by Weights and Biases

The results are saved in both your local device in the folder named wandb and also the Weights and Biases cloud. You can inspect the results directly on web or to use the local files, please check the documentation and visualization/plot_results.ipynb.