/wandb_minimalist_example

Primary LanguagePython

CIFAR-10 Training Showcase with Weights & Biases

This repo demonstrates a simple example on how to use Weights & Biases (wandb) for managing and tracking machine learning experiments. You will find a train.py script in this repository, which is designed to train a ResNet-18 model on the CIFAR-10 dataset with full integration of wandb for logging and sweeps.

Prerequisites

To run this project, you will need Python 3.6+ with the following libraries :

  • torch
  • torchvision
  • wandb
  • omegaconf : it's used to simply convert a yaml config file into a python object (instead of using a python dict) which allows for example to write config.lr (instead of config['lr'])

Install the necessary packages using pip:

pip install torch torchvision wandb omegaconf

Create an account on Weights & Biases and set up the wandb CLI on your machine.

Wandb classical run

To start a standard training run with your configuration:

python train.py --config config.yaml

Wandb Sweep run

For a hyperparameter sweep, use the config_sweep.yaml file instead :

python train.py --config config_sweep.yaml

As it's implemented in train.py, the string 'sweep' has to be inside the config sweep file name for the training script to know that it should use a sweep run. This allows some flexibility by using the same script train.py for both classical trainning with wandb or wandb sweeps.

Additional Resources

For a detailed explanation see the official doc of wandb :