fdq 0.0.22

ML runtime (https://pypi.org/project/fdq/)

FDQ | Fonduecaquelon

If you’d rather enjoy a delicious fondue than waste time on repetitive PyTorch boilerplate, this project is for you. FDQ streamlines your deep learning workflow so you can focus on what matters—your experiments, not your setup.

https://github.com/mstadelmann/fonduecaquelon

https://pypi.org/project/fdq/

SETUP

pip install fdq

or

git clone https://github.com/mstadelmann/fonduecaquelon.git
cd fonduecaquelon
pip install -e .

USAGE

Local

All experiment parameters must be stored in a config file. Note that config files can be based on a parent file.

fdq <path_to_config_file.json>

Slum Cluster

If you want to run your experiment on a Slurm cluster, you have to add the slurm_cluster section, check here for an example.

python <path_to>/fdq_submit.py <path_to_config_file.json>

Configuration

To run an experiment with FDQ, you need to define your experiment loop, a data-loader and, optionally, a test loop.

Model(s)

The model can either be a pre-installed one—such as Chuchichaestli — or a custom model that you define and import yourself. Models, losses, and data loaders are always defined as dictionaries. For example, the following configuration:

"models": {
    "ccUNET": {
        "class_name": "chuchichaestli.models.unet.unet.UNet"
    }
}

allows you to access the model in your training loop via experiment.models["ccUNET"]. The same dictionary-based structure applies to losses and data loaders as well. This setup enables you to define and manage as many models, losses, and data loaders as needed for your experiment.

Data-Loader(s)

The Data-Loader class must provide a function create_datasets(experiment, args) which is expected to return a dictionary, e.g.

return {
    "train_data_loader": train_loader,
    "val_data_loader": val_loader,
    "test_data_loader": test_loader,
    "n_train_samples": n_train,
    "n_val_samples": n_val,
    "n_test_samples": n_test,
    "n_train_batches": len(train_loader),
    "n_val_batches": len(val_loader) if val_loader is not None else 0,
    "n_test_batches": len(test_loader),
}

These values can then be accessed from your training loop.

Train

Define the path to your training loop in the train section. FDQ expects this file to provide the following function:

def fdq_train(experiment: fdqExperiment):

Within the training loop, you can access the training arguments, model, and data as follows:

nb_epochs = experiment.exp_def.train.args.epochs
data_loader = experiment.data["OXPET"].train_data_loader
model = experiment.models["ccUNET"]

See train_oxpets.py for an example.

Test

Similar to the training loop, the test loop can be defined in a custom file (this can also be the same file as the training loop). FDQ expects the specified file to provide the following function:

def fdq_test(experiment: fdqExperiment):

See oxpets_test.py for an example.

Install additional pip packages

If your experiment needs extra Python packages, you can install them on the worker by specifying them in your configuration. Just add the required package names (and optionally the version) under a additional_pip_packages section in your experiment setup file. This ensures all dependencies are installed automatically before your code runs.

Example:

    "slurm_cluster": {
        "fdq_version": "0.0.7",
        "...": "...",
        "additional_pip_packages": [
            "monai==1.4.0",
            "prettytable"
        ]
    }