Skip to content

Latest commit

 

History

History
47 lines (29 loc) · 1.86 KB

README.md

File metadata and controls

47 lines (29 loc) · 1.86 KB
Raw

LRA Benchmark

CREDITS

Adapted from the Nystromformer authors' repo and the original long-range-arena repo

Initial setup

  • install the needed dependencies, listed in the requirements-lra.txt file. Can be done from the root of the repo with pip install -r requirements-lra.txt

  • Run the install script, which can take a little while. bash install.sh

Evaluating

To run a LRA experiment, run the following command in LRA root folder

python3 run_tasks.py --attention <attention> --task <task> --config <config_path> --world_size N

where <attention> can be set to any attention defined in the attached config file, and world size is the number of GPUs on which the benchmark should be run. The best models and log files will be saved LRA/logs/ folder. More parts can be tested, not just attention, the model configuration can be freely altered in the config file (as long as the dimensions agree).

A list and explanation of all the options is available via

python run_tasks.py --help

Submitit

If you have access to a SLURM enabled cluster, you can submit distributed jobs via this script. python3 run_with_submitit.py --attentions <your attention> --tasks <your task> --nodes 1 --config code/config.json --world_size <how many ranks> --checkpoint_dir <shared folder where to write the logs and best models>

Batch submissions are possible, for instance as follows

python3 batch_submit.py -c code/config.json -ck <your checkpoing and log path> -a lambda

Collecting all the results at once can be done with another small script, as follows:

python3 batch_fetch_results.py -ck <your checkpoing and log path>

This will synthetise the logs and show the score matrix for all the tasks and variants in this run.