torchrecipes
This library is currently under heavy development - if you have suggestions on the API or use-cases you'd like to be covered, please open an github issue or reach out. We'd love to hear about how you're using torchrecipes.
torchrecipes is a prototype is built on top of PyTORCH and provides a set of reproduci-able, re-usable, ready-to-run RECIPES for training different types of models, across multiple domains, on PyTorch Lightning.
It aims to provide reproduci-able "applications" built on top of PyTorch with good performance and easy reproduciability. Because this project builds on the pytorch ecosystem and requires significant investment, we'd love to hear from and work with early adopters to shape the design. Please reach out on the issue tracker if you're interested in using this for your project.
Why torchrecipes?
The primary goal of the torchrecipes is to 10x ML development by providing standard blueprints to easily train production-ready ML models across environemnts (from local development to cluster deployment).
Requirements
PyTorch Recipes (torchrecipes):
- python3 (3.8+)
- torch
Running
The easiest way to run torchrecipes is to use torchx. You can install it directly (if not already included as part of our requirements.txt) with:
pip install torchx
Then go to torchrecipes/launcher/ and create a file torchx_app.py:
# 'torchrecipes/launcher/torchx_app.py'
import torchx.specs as specs
image_classification_args = [
"-m", "run",
"--config-name",
"train_app",
"--config-path",
"torchrecipes/vision/image_classification/conf",
]
def torchx_app(image: str = "run.py:latest", *job_args: str) -> specs.AppDef:
return specs.AppDef(
name="run",
roles=[
specs.Role(
name="run",
image=image,
entrypoint="python",
args=[*image_classification_args, *job_args],
env={
"CONFIG_MODULE": "torchrecipes.vision.image_classification.conf",
"MODE": "prod",
"HYDRA_FULL_ERROR": "1",
}
)
],
)
This app defines the entrypoint, args and image for launching.
Now that we have created a torchx app, we are (almost) ready for launching a job!
Firstly, create a symlink for launcher/run.py at the top level of the repo:
ln -s torchrecipes/launcher/run.py ./run.py
Then we are ready-to-go! Simply launch the image_classification recipe with the following command:
torchx run --scheduler local_cwd torchrecipes/launcher/torchx_app.py:torchx_app trainer.fast_dev_run=True trainer.checkpoint_callback=False +tb_save_dir=/tmp/
Release
# install torchrecipes
pip install torchrecipes
Contributing
We welcome PRs! See the CONTRIBUTING file.
License
torchrecipes is BSD licensed, as found in the LICENSE file.
![[paved path] basic charnn with GPT example](https://avatars.githubusercontent.com/u/3350366?v=4)
184 Feb 10, 2021
20 Dec 15, 2022
873 Dec 15, 2022
220 Dec 30, 2022
15k Jan 2, 2023
40 Sep 19, 2022
581 Dec 21, 2022
513 Jan 3, 2023
5 Oct 21, 2022
3.8k Dec 30, 2022
357 Dec 24, 2022
914 Jan 7, 2023
2k Dec 27, 2022
19 Oct 28, 2022
358 Dec 24, 2022