wandbtocsv
allows you to quickly export a flat csv file of the top
level metrics, config, etc. from your Weights &
Biases runs in an easy to use CLI interface. This
utility will also get peak gpu and system memory usage for each run.
pip install wandbtocsv
!wandbtocsv -h
usage: wandbtocsv [-h] [--entity ENTITY] [--project PROJECT]
[--output_file OUTPUT_FILE]
options:
-h, --help show this help message and exit
--entity ENTITY Entity for the WANDB runs. If not provided,
attempts to fetch from environment variable
'WANDB_ENTITY'.
--project PROJECT Project for the WANDB runs. If not provided,
attempts to fetch from environment variable
'WANDB_PROJECT'.
--output_file OUTPUT_FILE Path to save the output CSV. Default name format:
{entity}-{project}-{mmddyy}.csv
!wandbtocsv --entity hamelsmu --project deepspeed-data
Fetching runs... ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:00% 0:00:01[36m0:00:03
Saved 480 runs to hamelsmu-deepspeed-data-100823.csv
Bonus, explore your wandb runs in Datasette
csvs-to-sqlite hamelsmu-deepspeed-data-100823.csv hamelsmu-deepspeed-data-100823.db
datasette hamelsmu-deepspeed-data-100823.db -o