TorchCriterion

torchcriterion is a modular, extensible library of PyTorch-compatible loss functions ("criteria") for classification, regression, segmentation, and metric learning tasks. It offers a curated collection of both standard and custom loss functions, built with flexibility and composability in mind.

🚀 Features

• 🧱 Modular architecture for clean API and extension
• 🧪 Ready-to-use implementations of popular losses
• 🧩 Supports multi-loss composition and custom scheduling
• ⚡ Fully compatible with PyTorch’s autograd and GPU acceleration

📦 Installation

pip install torchcriterion

🧰 Supported Losses

Classification

• CrossEntropyLoss
• FocalLoss

Regression

• MSELoss
• HuberLoss

Segmentation

• DiceLoss
• TverskyLoss

Metric Learning

• TripletLoss
• ContrastiveLoss

NLP Metrics (new)

Under torchcriterion.nlp we provide common NLP evaluation metrics.

Example usage:

from torchcriterion.nlp.bleu import bleu_score
from torchcriterion.nlp.rouge import rouge_l_batch
from torchcriterion.nlp.perplexity import perplexity

preds = ["the cat sat on mat"]
refs  = ["the cat is on the mat"]

print("BLEU:", bleu_score(preds, refs))
print("ROUGE-L:", rouge_l_batch(preds, refs))

# LM perplexity example
logits = model_out  # (batch, seq_len, vocab)
targets = labels    # (batch, seq_len)
print("Perplexity:", perplexity(logits, targets))

🧪 Example Usage

from torchcriterion import FocalLoss

criterion = FocalLoss(gamma=2.0, alpha=0.25)
loss = criterion(predictions, targets)

📁 Project Structure

torchcriterion/
├── classification/
│   ├── cross_entropy.py
│   ├── focal.py
├── regression/
│   ├── mse.py
│   ├── huber.py
├── segmentation/
│   ├── dice.py
│   ├── tversky.py
├── metric_learning/
│   ├── triplet.py
│   ├── contrastive.py
├── nlp/
│   ├── bleu.py
│   ├── perplexity.py
│   ├── rouge.py
├── base.py
├── __init__.py

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.

🙌 Contributing

Pull requests, ideas, and issues are welcome! Feel free to open a PR or start a discussion.

👤 Author

Developed by Liodon AI

⭐️ Star the Repo

If you find this library useful, please consider starring it to show your support!

🔗 Related Projects

• torchmetrics — for evaluation metrics
• timm — for models with built-in loss support

📚 Citation

If you use TorchCriterion in your research or project, please consider citing it:

@misc{torchcriterion2025,
  title        = {TorchCriterion: Advanced Loss Functions for PyTorch},
  year         = {2025},
  publisher    = {GitHub},
  howpublished = {\url{https://github.com/Liodon-AI/TorchCriterion}},
  note         = {GitHub repository},
}

Made with ❤️ and PyTorch