Pipeline Runtime
python -m pipe.main ... # train models (+eval) (+preprocess)
Do use the --help
and examples to explore.
python pipe/data/download/datasets/download_datasets.py
Data for T5 tasks is obtained by using --mode perprocess
cmd option.
See configs for config examples.
To choose a spesific config, add it to command line:
mpirun -np 2 python -m pipe.main --config $PATH_TO_CONFIG
without doing so, it will run the dummy config (created for dev usage).
if data preprocessing is needed, run the selected config with:
python -m pipe.main --mode preproc --config $PATH_TO_CONFIG ...
cuda aware openmpi:
mpirun -np 2 python -m pipe.main --config $PATH_TO_CONFIG
A PoC runtime which can be used with the --mode mp
cmd option.
Supposed to work for very simple stright pipelines only (mostly- torchvision models, vit), and is BUGGY when configuration gets more exotic. (e.g, Tied Wieghts)
python -m pipe.main --nprocs 2 --mode mp --config $PATH_TO_CONFIG