torchlosses

A lightweight library of advanced PyTorch loss functions — ready to plug into your training loop.
Designed for deep learning practitioners who want more than just MSE and CrossEntropy.

Available on PyPI: torchlosses

---

Features

• Focal Loss — handle class imbalance in classification.
• Dice Loss — segmentation-friendly overlap metric.
• Contrastive Loss — learn pairwise similarity (Siamese nets).
• Triplet Loss — enforce anchor-positive vs negative separation.
• Cosine Embedding Loss — similarity learning with cosine distance.
• Huber Loss — robust regression, less sensitive to outliers.
• KL Divergence Loss — probability distribution alignment.

---

Created By

Naga Adithya Kaushik (GenAIDevTOProd) https://medium.com/@GenAIDevTOProd/from-loss-functions-to-training-utilities-building-pytorch-packages-from-scratch-91e884d14001

Installation

pip install torchlosses

## Usage

import torch
from torchlosses import FocalLoss, DiceLoss

# Focal Loss for classification
inputs = torch.randn(4, 5, requires_grad=True)  # logits for 5 classes
targets = torch.randint(0, 5, (4,))
criterion = FocalLoss()
loss = criterion(inputs, targets)
print("Focal Loss:", loss.item())

# Dice Loss for segmentation
inputs = torch.randn(4, 1, 8, 8)
targets = torch.randint(0, 2, (4, 1, 8, 8))
criterion = DiceLoss()
loss = criterion(inputs, targets)
print("Dice Loss:", loss.item())