How to Use BCEWithLogitsLoss in PyTorch: Simple Guide

Use torch.nn.BCEWithLogitsLoss in PyTorch to combine a sigmoid layer and binary cross-entropy loss in one step. It expects raw model outputs (logits) and target labels as inputs and returns the loss value for training binary classifiers.

Syntax

The BCEWithLogitsLoss class is initialized without required arguments but supports optional parameters like weight for per-sample weighting, pos_weight to emphasize positive examples, and reduction to control how losses are combined.

During training, call the loss instance with two arguments: input (raw model outputs, called logits) and target (ground truth labels as floats 0 or 1).

python

loss_fn = torch.nn.BCEWithLogitsLoss(weight=None, pos_weight=None, reduction='mean')
loss = loss_fn(input, target)

Example

This example shows how to use BCEWithLogitsLoss with a simple model output and target labels. It computes the loss value from raw logits and prints it.

python

import torch

# Create example logits (raw outputs from model) and target labels
logits = torch.tensor([0.2, -1.5, 3.0, 0.7])
targets = torch.tensor([1., 0., 1., 0.])

# Initialize BCEWithLogitsLoss
loss_fn = torch.nn.BCEWithLogitsLoss()

# Calculate loss
loss = loss_fn(logits, targets)

print(f'Loss: {loss.item():.4f}')

Output

Loss: 0.4170

Common Pitfalls

• Passing probabilities instead of logits: BCEWithLogitsLoss expects raw logits, not sigmoid probabilities. Applying sigmoid before this loss causes incorrect results.
• Incorrect target types: Targets must be floats (0.0 or 1.0), not integers or other types.
• Ignoring reduction: By default, losses are averaged. Use reduction='none' to get per-sample losses.

python

import torch

# Wrong: passing probabilities (sigmoid applied) to BCEWithLogitsLoss
probs = torch.sigmoid(torch.tensor([0.2, -1.5, 3.0, 0.7]))
targets = torch.tensor([1., 0., 1., 0.])
loss_fn = torch.nn.BCEWithLogitsLoss()

try:
    loss_wrong = loss_fn(probs, targets)
except Exception as e:
    print(f'Error: {e}')

# Right: pass raw logits
logits = torch.tensor([0.2, -1.5, 3.0, 0.7])
loss_right = loss_fn(logits, targets)
print(f'Correct loss: {loss_right.item():.4f}')

Output

Correct loss: 0.4170

Quick Reference

BCEWithLogitsLoss combines sigmoid activation and binary cross-entropy loss in one step for numerical stability.

• input: raw model outputs (logits), shape (N, *)
• target: float labels 0 or 1, same shape as input
• weight: optional manual rescaling weight per sample
• pos_weight: optional weight for positive examples
• reduction: 'mean' (default), 'sum', or 'none'

Use it for binary classification tasks where model outputs logits.

Key Takeaways

Use BCEWithLogitsLoss with raw logits, not probabilities.

Targets must be floats with values 0.0 or 1.0.

It combines sigmoid and binary cross-entropy for better stability.

Set reduction to 'none' to get individual losses per sample.

Use pos_weight to handle class imbalance by weighting positives.