Auto-configure (not only) torch experiments from the CLI.

Parsonaut makes your experiments

1. Configurable - configure any parameter of your experiment from CLI
2. Reproducible - easily store your full experiment configuration to disk
3. Boilerplate-free - make model checkpointing seampless

To install the library, clone the repository and use pip:

pip install git+https://github.com/janvainer/parsonaut.git

Let's supercharge a simple torch experiment with automatic CLI configuration

"""
usage: script.py [-h] [--in_channels int] [--out_channels int]

options:
 -h, --help show this help message and exit
 --in_channels int
 --out_channels int
"""
import torch.nn as nn
from parsonaut import Parsable
class Model(nn.Module, Parsable):
 def __init__(
 self,
 in_channels: int = 4,
 out_channels: int = 2,
 ):
 super().__init__()
# Parse user CLI args - ta partially initialized model
partial_model = Model.parse_args()
# Serialize model configuration
partial_model.to_file("model_config.yaml")

Now we can do some training. We instantiate the model configuration into a torch model.

model = partial_model.to_eager()
# Training code here ...

Finally, serialize model configuration AND weights.

model.to_checkpoint("ckpt_dir")

We can now load the experiment configuration and model weights later:

model_with_weights = Model.from_checkpoint("ckpt_dir")
just_config = Model.from_file("model_config.yaml")

Parsonaut allows configuring multiple possibly nested classes. Moreover, you can dynamically select which classes to use via enums.

To explore more advanced features, please see the following tutorials:

• zero - simple, but complete experiment
• hero - dynamic and nested configurations