hydra-torch

Configuration classes enabling type-safe PyTorch configuration for Hydra apps.
This repo is work in progress.

The config dataclasses are generated using configen, check it out if you want to generate config dataclasses for your own project.

Install:

# For now, please obtain through github. Soon, versioned (per-project) dists will be on PyPI.
pip install git+https://github.com/pytorch/hydra-torch

Example config:

Here is one of many configs available. Notice it uses the defaults defined in the torch function signatures:

@dataclass
class TripletMarginLossConf:
    _target_: str = "torch.nn.modules.loss.TripletMarginLoss"
    margin: float = 1.0
    p: float = 2.0
    eps: float = 1e-06
    swap: bool = False
    size_average: Any = None
    reduce: Any = None
    reduction: str = "mean"

Importing Convention:

from hydra_configs.<package_name>.path.to.module import <ClassName>Conf

where <package_name> is the package being configured and path.to.module is the path in the original package.

Inferring where the package is located is as simple as prepending hydra_configs. and postpending Conf to the original class import: e.g.

#module to be configured
from torch.optim.adam import Adam

#config for the module
from hydra_configs.torch.optim.adam import AdamConf

Getting Started:

Take a look at our tutorial series:

Basic Tutorial
Intermediate Tutorial (coming soon)
Advanced Tutorial (coming soon)

Other Config Projects:

A list of projects following the hydra_configs convention (please notify us if you have one!):

Pytorch Lightning

License

hydra-torch is licensed under MIT License.

pytorch / hydra-torch

readme