Introduction

torch.nn.Conv2d applies a 2D convolution to image-like tensors, and its parameters describe both the input shape assumptions and how the convolution should move across the input. The most important parameters are in_channels, out_channels, kernel_size, stride, padding, dilation, groups, and bias.

A Simple Example

1import torch
2import torch.nn as nn
3
4conv = nn.Conv2d(
5    in_channels=3,
6    out_channels=16,
7    kernel_size=3,
8    stride=1,
9    padding=1,
10)
11
12x = torch.randn(8, 3, 32, 32)
13y = conv(x)
14print(y.shape)

Output shape:

torch.Size([8, 16, 32, 32])

This one example already explains several parameters at once.

What Each Parameter Means

in_channels

The number of channels in the input tensor. For a normal RGB image, this is usually 3. If the input is grayscale, it is often 1.

out_channels

The number of filters the layer learns. It also becomes the number of channels in the output tensor.

kernel_size

The height and width of the convolution window. A value of 3 means a 3 x 3 kernel.

stride

How far the kernel moves each step. A stride of 1 checks every nearby position. A stride of 2 skips positions and reduces spatial resolution faster.

padding

How many pixels are added around the input border before convolution. Padding is often used to preserve spatial dimensions.

dilation

How spread out the kernel elements are. Dilation enlarges the receptive field without increasing the kernel size directly.

groups

Controls grouped convolution. At groups=1, every filter sees all input channels. Higher group values split channel connections. Depthwise convolution is a special case where groups equals in_channels.

bias

Whether each output channel has an additive bias term. It is often left enabled unless a following normalization layer makes it unnecessary.

Shape Intuition

PyTorch expects the input shape in this order:

batch, channels, height, width

That is why the example uses torch.randn(8, 3, 32, 32): eight images, three channels, and a 32 x 32 spatial size.

If you accidentally pass channels last, you will get a shape error or a wrong interpretation of the data.

Why Padding and Stride Matter So Much

A 3 x 3 kernel with stride=1 and padding=1 is very common because it preserves height and width. Without padding, the output shrinks.

That means these three values work together, not independently. Many architecture bugs come from thinking about each parameter in isolation.

A Second Example With Stride

conv = nn.Conv2d(3, 16, kernel_size=3, stride=2, padding=1)
x = torch.randn(1, 3, 32, 32)
print(conv(x).shape)

With stride 2, the spatial dimensions shrink because the filter moves in larger steps.

Common Pitfalls

Bias and Padding Mode

Two smaller parameters are easy to forget. bias=True adds a learnable offset to each output channel, and padding_mode controls how padded values are filled when non-default behavior is needed. Most beginner examples leave these at their defaults, which is usually fine until a model architecture needs something more specific.

The most common mistake is mixing up in_channels with batch size. Batch size is the first dimension, not the channel count.

Another mistake is passing image data in channels-last order when Conv2d expects channels first.

Developers also often change stride, padding, or kernel_size without checking how the output shape changes through the rest of the network.

Summary

• 'in_channels describes the number of input feature channels.'
• 'out_channels is the number of learned filters and output channels.'
• 'kernel_size, stride, and padding control how the convolution moves and what output size it produces.'
• 'groups changes channel connectivity and enables grouped or depthwise convolution.'
• 'Conv2d expects tensors in batch, channels, height, width order.'