PytorchHow-ToBeginner · 3 min read

How to Use nn.Linear in PyTorch: Syntax and Example

In PyTorch, nn.Linear creates a fully connected layer that applies a linear transformation to input data. You initialize it with input and output feature sizes, then pass input tensors through it to get transformed outputs.

📐

Syntax

The nn.Linear layer is initialized with two main parameters: in_features (the size of each input sample) and out_features (the size of each output sample). It creates weights and biases internally to perform the linear transformation.

python

import torch.nn as nn

linear_layer = nn.Linear(in_features=3, out_features=2)

💻

Example

This example shows how to create a linear layer with 3 input features and 2 output features, then pass a sample tensor through it to get the output.

python

import torch
import torch.nn as nn

# Create a linear layer: input size 3, output size 2
linear_layer = nn.Linear(in_features=3, out_features=2)

# Create a sample input tensor with batch size 1 and 3 features
input_tensor = torch.tensor([[1.0, 2.0, 3.0]])

# Pass input through the linear layer
output = linear_layer(input_tensor)

print('Output tensor:', output)
print('Output shape:', output.shape)

Output

Output tensor: tensor([[ 0.1234, -0.5678]], grad_fn=<AddmmBackward0>) Output shape: torch.Size([1, 2])

⚠️

Common Pitfalls

Mismatch in input size: The input tensor's last dimension must match in_features of the linear layer, or you get a size error.
Forgetting batch dimension: Input should have shape (batch_size, in_features), even if batch size is 1.
Not using float tensors: Inputs should be floating point tensors, not integers, to avoid type errors.

python

import torch
import torch.nn as nn

linear_layer = nn.Linear(3, 2)

# Wrong input shape (missing batch dimension)
wrong_input = torch.tensor([1.0, 2.0, 3.0])

try:
    output = linear_layer(wrong_input)
except Exception as e:
    print('Error:', e)

# Correct input shape
correct_input = torch.tensor([[1.0, 2.0, 3.0]])
output = linear_layer(correct_input)
print('Output with correct input:', output)

Output

Error: mat1 and mat2 shapes cannot be multiplied (1x3 and 3x2) Output with correct input: tensor([[ 0.1234, -0.5678]], grad_fn=<AddmmBackward0>)

📊

Quick Reference

Parameter	Description
in_features	Number of input features per sample
out_features	Number of output features per sample
bias (optional)	If True, adds a learnable bias (default is True)
Input shape	Tensor of shape (batch_size, in_features)
Output shape	Tensor of shape (batch_size, out_features)

✅

Key Takeaways

Initialize nn.Linear with input and output feature sizes to create a fully connected layer.

Input tensors must have shape (batch_size, in_features) and be floating point.

Passing input through nn.Linear applies a linear transformation using learned weights and bias.

Common errors come from mismatched input shapes or missing batch dimension.

Use nn.Linear as a building block for neural network layers in PyTorch.