How to Use Compose Transforms in PyTorch for Image Preprocessing

In PyTorch, use transforms.Compose to chain multiple image transformations into one pipeline. This lets you apply several preprocessing steps like resizing, converting to tensor, and normalizing in sequence with a single call.

Syntax

The transforms.Compose function takes a list of transformation functions and applies them in order to an input image or data. Each transform is called one after another.

• transforms.Compose([transform1, transform2, ...]): Combines multiple transforms.
• transform1, transform2, ...: Individual transform functions like Resize, ToTensor, Normalize.

python

from torchvision import transforms

transform_pipeline = transforms.Compose([
    transforms.Resize((128, 128)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
])

Example

This example shows how to create a Compose transform pipeline to resize an image, convert it to a tensor, and normalize it. Then it applies the pipeline to a sample image loaded with PIL.

python

from PIL import Image
from torchvision import transforms
import torch

# Define the transform pipeline
transform_pipeline = transforms.Compose([
    transforms.Resize((128, 128)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
])

# Load an example image (replace 'example.jpg' with a valid image path)
image = Image.new('RGB', (256, 256), color='red')  # Using a red image for demo

# Apply the transform pipeline
transformed_image = transform_pipeline(image)

# Print the tensor shape and type
print(f"Transformed image tensor shape: {transformed_image.shape}")
print(f"Tensor type: {type(transformed_image)}")

Output

Transformed image tensor shape: torch.Size([3, 128, 128])
Tensor type: <class 'torch.Tensor'>

Common Pitfalls

Common mistakes when using Compose include:

• Not converting images to tensor before normalization, causing errors.
• Applying transforms in the wrong order, e.g., normalizing before converting to tensor.
• Using incompatible transforms for the data type (e.g., applying image transforms to non-image data).

Always ensure ToTensor() comes before Normalize() in the pipeline.

python

from torchvision import transforms

# Wrong order: Normalize before ToTensor (will cause error)
wrong_pipeline = transforms.Compose([
    transforms.Normalize(mean=[0.5], std=[0.5]),
    transforms.ToTensor()
])

# Correct order
correct_pipeline = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5], std=[0.5])
])

Quick Reference

Transform	Purpose	Notes
Resize(size)	Change image size	Size can be int or tuple (height, width)
ToTensor()	Convert PIL image or numpy array to PyTorch tensor	Must be before Normalize
Normalize(mean, std)	Normalize tensor image with mean and std	Input tensor expected in [C, H, W] format
RandomHorizontalFlip()	Randomly flip image horizontally	Used for data augmentation
Compose([transforms])	Chain multiple transforms	Transforms applied in order

Key Takeaways

Use transforms.Compose to chain multiple image transformations in PyTorch.

Always convert images to tensor with ToTensor() before applying Normalize().

Order of transforms matters; Compose applies them sequentially.

Compose helps keep preprocessing code clean and reusable.

Test your transform pipeline on sample data to avoid common errors.