focal-ce-loss 0.0.2

A loss function for

Focal Cross Entropy Loss

A PyTorch-compatible implementation of the Focal Cross Entropy loss function, designed to address class imbalance by down-weighting easy examples and focusing training on hard negatives.

Install

pip install focal_ce_loss_funcs_pytorch

Quick start

from focal_ce_loss import FocalCELoss

criterion = FocalCrossEntropyLoss()
loss = criterion(inputs, targets)

FocalCELoss

A PyTorch nn.Module implementing the Focal Cross Entropy loss.

from focal_ce_loss import FocalCELoss

criterion = FocalCELoss(alpha=0.25, gamma=2.0, reduction="sum")

# Compute loss from logits and targets
loss = criterion(inputs, targets)
loss.backward()

Parameters

• gamma: focusing parameter; higher values down-weight easy examples more aggressively (default: 2.0)
• alpha: optional class weight tensor of shape (C,) for additional class balancing (default: 0.25)
• reduction: specifies the reduction to apply — "mean", "sum", or "none" (default: "sum")
• ignore_index: target value to ignore during loss computation (default: -100)

Methods

• forward(input, target) — accepts a torch.Tensor of shape (N, C) and integer targets of shape (N,); returns a scalar loss tensor (or per-sample tensor if reduction="none")

Inputs

• input: torch.Tensor of shape (N, C)
• target: torch.Tensor of shape (N,) ` Example

import torch
from focal_ce_loss import FocalCELoss

inputs  = torch.randn(8, 10)          # batch of 8, 10 classes
targets = torch.randint(0, 10, (8,))  # ground-truth labels

criterion = FocalCrossEntropyLoss(gamma=2.0)
loss = criterion(inputs, targets)
loss.backward()

BFCELoss

A PyTorch nn.Module implementing the Binary Focal Cross Entropy loss for binary classification tasks.

from focal_ce_loss import BFCELoss

criterion = BFCELoss(alpha=0.25, gamma=2.0, reduction="sum")

# Compute loss from logits and binary targets
loss = criterion(inputs, targets)
loss.backward()

Parameters

• gamma: focusing parameter; higher values down-weight easy examples more aggressively (default: 2.0)
• alpha: optional scalar weight for the positive class (default: 0.25)
• reduction: specifies the reduction to apply — "mean", "sum", or "none" (default: "mean")
• ignore_index: target value to ignore during loss computation (default: -100)

Methods

• forward(input, target) — accepts a torch.Tensor of shape (N,) or (N, 1) and binary targets of shape (N,); returns a scalar loss tensor (or per-sample tensor if reduction="none")

Inputs

• input: torch.Tensor of shape (N,) or (N, 1) of predicted probabilities with values in [0, 1].
• target: torch.Tensor of shape (N,) containing binary labels {0, 1}

Example

import torch
from focal_ce_loss import BFCELoss

inputs  = torch.randn(8)           # batch of 8 logits
targets = torch.randint(0, 2, (8,)).float()  # binary ground-truth labels

criterion = BFCELoss(gamma=2.0, alpha=0.25)
loss = criterion(inputs, targets)
loss.backward()