Weights & Biases (wandb)

Weights & Biases (wandb) is a meta machine learning platform designed to help AI practitioners and teams build reliable machine learning models for real-world applications by streamlining the machine learning model lifecycle. By using wandb, you can track, compare, explain and reproduce machine learning experiments.

Using wandb on Alliance clusters

Availability on compute nodes

Since it requires an internet connection, wandb has restricted availability on compute nodes, depending on the cluster:

Cluster	Availability	Note
Béluga	No ❌	wandb requires access to Google Cloud Storage, which is not accessible from the compute nodes
Narval
Cedar	Yes ✅	internet access is enabled
Graham	No ❌	internet access is disabled on compute nodes

Béluga and Narval

While it is possible to upload basic metrics to Weights&Biases during a job on Béluga, the wandb package automatically uploads information about your environment to a Google Cloud Storage bucket, resulting in a crash during or at the very end of a training run. It is not currently possible to disable this behaviour. Uploading artifacts to W&B with wandb.save() also requires access to Google Cloud Storage, which is not available on Béluga's compute nodes.

You can still use wandb on Béluga by enabling the offline mode. In this mode, wandb will write all metrics, logs and artifacts to the local disk and will not attempt to sync anything to the Weights&Biases service on the internet. After your jobs finish running, you can sync their wandb content to the online service by running the command wandb sync on the login node.

Note that Comet.ml is a product very similar to Weights & Biases, and works on Béluga.

Example

The following is an example of how to use wandb to track experiments on Béluga. To reproduce this on Cedar, it is not necessary to enable the offline mode.

File : wandb-test.sh

#!/bin/bash
#SBATCH --account=YOUR_ACCOUNT
#SBATCH --cpus-per-task=2 # At least two cpus is recommended - one for the main process and one for the wandB process
#SBATCH --mem=4G       
#SBATCH --time=0-03:00
#SBATCH --output=%N-%j.out


module load StdEnv/2020 python/3.8
virtualenv --no-download $SLURM_TMPDIR/env
source $SLURM_TMPDIR/env/bin/activate
pip install --no-index wandb

### Save your wandb API key in your .bash_profile or replace $API_KEY with your actual API key. Uncomment the line below and comment out "wandb offline" if running on Cedar ###

#wandb login $API_KEY 

wandb offline

python wandb-test.py

The script wandb-test.py is a simple example of metric logging. See W&B's full documentation for more options.

File : wandb-test.py

import wandb

wandb.init(project="wandb-pytorch-test", settings=wandb.Settings(start_method="fork"))

for my_metric in range(10):
    wandb.log({'my_metric': my_metric})

After a training run in offline mode, there will be a new folder ./wandb/offline-run*. You can send the metrics to the server using the command wandb sync ./wandb/offline-run*. Note that using * will sync all runs.