Fenn: Friendly Environment for Neural Networks

Stop writing boilerplate. Start training.

Friendly Environment for Neural Networks (fenn) is a simple framework that automates ML/DL workflows by providing prebuilt trainers, templates, logging, configuration management, and much more. With fenn, you can focus on your model and data while it takes care of the rest.

Support fenn

If fenn is useful for your work or research, consider supporting its development.

You can support the project by starring the repository on GitHub. It improves visibility and helps others discover fenn.

Sponsorship also helps fund maintenance, improvements, and new features.

Support the project: https://github.com/sponsors/blkdmr

Why fenn?

• Auto-Configuration: YAML files are automatically parsed and injected into your entrypoint with CLI override support. No more hardcoded hyperparameters or scattered config logic.

• Unified Logging: All logs, print statements, and experiment metadata are automatically captured to local files and remote tracking backends simultaneously with no manual setup required.

• Backend Monitoring: Native integration with industry-standard trackers like Weights & Biases (W&B) for centralized experiment tracking and TensorBoard for real-time metric visualization

• Instant Notifications: Get real-time alerts on Discord and Telegram when experiments start, complete, or fail—no polling or manual checks.

• Trainers: Built-in support for training loops, validation, and testing with minimal boilerplate. Just define your model and data, and let fenn handle the rest.

• Template Ready: Built-in support for reproducible, shareable experiment templates.

Quickstart

Install the fenn library using

pip install fenn

or

uv pip install fenn

Initialize a Project

Use the CLI to discover and download a project template.

1. List available templates

fenn list

This fetches the directory listing from pyfenn/templates and prints the templates you can use.

2. Pull a template

fenn pull <template> [path]

Examples:

fenn pull empty            # pull into the current directory
fenn pull empty ./my-proj  # pull into ./my-proj (created if missing)

Each template ships at least a main.py entrypoint and a fenn.yaml configuration file in the target directory. Most templates also include a README.md, a requirements.txt, and a modules/ directory with example model and dataset code.

To avoid clobbering work, fenn pull refuses to write into a non-empty target directory. Pass --force to overwrite existing files:

fenn pull empty --force

Hidden entries (those starting with ., such as .git) do not count as "non-empty".

3. Customize and run

Open the generated fenn.yaml and adjust hyperparameters, paths, logging, and integrations for your project (see Configuration below). Then run the entrypoint:

python main.py

Common issues

• Template <name> not found — The template name doesn't match a directory in pyfenn/templates. Run fenn list to see valid names.
• Refusing to pull into non-empty directory — Either pull into an empty directory, point path at a fresh one, or pass --force to overwrite.
• Network error / Failed to check template existence — Check connectivity. The CLI uses the unauthenticated GitHub API to look up and download templates, which is subject to GitHub's rate limit.

Configuration

fenn relies on a simple YAML structure to define hyperparameters, paths, logging options, and integrations. You can configure the fenn.yaml file with the hyperparameters and options for your project.

The structure of the fenn.yaml file is:

# ---------------------------------------
# Fenn Configuration (Modify Carefully)
# ---------------------------------------

project: empty

# ---------------------------
# Logging & Tracking
# ---------------------------

logger:
  dir: logger

export:
  dir: exports

# ---------------------------------------
# Example of User Section
# ---------------------------------------

train:
    lr: 0.001

Write Your Code

Use the @app.entrypoint decorator. Your configuration variables are automatically passed via args.

from fenn import Fenn

app = Fenn()

@app.entrypoint
def main(args):
    # 'args' contains your fenn.yaml configurations
    print(f"Training with learning rate: {args['train']['lr']}")

    # Your logic here...

if __name__ == "__main__":
    app.run()

By default, fenn will look for a configuration file named fenn.yaml in the current directory. If you would like to use a different name, a different location, or have multiple configuration files for different configurations, you can call set_config_file() and update the path or the name of your configuration file. You must assign the filename before calling run().

The optional export.dir setting centralizes where artifacts are written. Components that export files can use this shared directory instead of requiring an output path to be passed through every call.

app = Fenn()
app.set_config_file("my_file.yaml")

Run It

You can run your code as usual

python main.py

and fenn will take care of the rest for you.

Training Models

Use built-in trainers to handle your training loops with minimal boilerplate.

import torch.nn as nn
import torch.optim as optim

from fenn.nn.trainers import ClassificationTrainer
from fenn.nn.utils import Checkpoint

@app.entrypoint
def main(args):
        
    # Define your data
    train_loader = DataLoader(train_dataset, batch_size=args["train"]["batch"], shuffle=True)
    val_loader = DataLoader(val_dataset, batch_size=args["test"]["batch"], shuffle=False)
    test_loader = DataLoader(test_dataset, batch_size=args["test"]["batch"], shuffle=False)
    
    # Define your model
    model = nn.Sequential( ... )     
    loss = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(),
                            lr=float(args["train"]["lr"]))

    # Initialize a ClassificationTrainer
    trainer = ClassificationTrainer(
        model=model,
        loss_fn=loss,
        optim=optimizer,
        num_classes=4
    )

    # Train and predict your model
    trainer.fit(train_loader, epochs=10, val_loader=val_loader)
    preds = trainer.predict(test_loader)

Cite fenn

If you use fenn in your work or research, please cite the project as:

@software{fenn,
  author       = {Alessio Russo},
  title        = {pyfenn/fenn: Release v0.2.0},
  month        = may,
  year         = 2026,
  publisher    = {Zenodo},
  version      = {v0.2.0},
  doi          = {10.5281/zenodo.20178660},
  url          = {https://doi.org/10.5281/zenodo.20178660},
}

Contributing

Contributions are welcome!

Interested in contributing? Join the community on Discord.

We can then discuss a possible contribution together, answer any questions, and help you get started!

Please consult our CONTRIBUTING.md and CODE_OF_CONDUCT.md before opening a pull request.

Maintainers

The development and long-term direction of fenn is guided by the following maintainers:

Maintainer	Role
@blkdmr	Creator & Project Administrator
@giuliaOddi	Project Administrator
@franciscolima05	Core Maintainer

Maintainers oversee the project roadmap, review pull requests, coordinate releases, and ensure the long-term stability and quality of the framework.

Thank you for supporting the project.