How to Use Transforms in PyTorch for Image Preprocessing

In PyTorch, you use transforms from torchvision.transforms to preprocess images by chaining operations like resizing, cropping, and normalization. You create a transform pipeline using transforms.Compose and apply it to images before feeding them into a model.

Syntax

The basic syntax to use transforms in PyTorch involves importing transforms from torchvision, creating a pipeline with transforms.Compose, and applying it to images.

• transforms.Compose([list_of_transforms]): Chains multiple transforms.
• Common transforms include Resize, CenterCrop, ToTensor, and Normalize.
• Apply the composed transform to an image by calling it like a function.

python

from torchvision import transforms

transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

Example

This example shows how to load an image, apply a transform pipeline to resize, crop, convert to tensor, and normalize it for model input.

python

from PIL import Image
from torchvision import transforms
import torch

# Define the transform pipeline
transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

# Load an image
img = Image.new('RGB', (300, 300), color='red')  # Create a red image for demo

# Apply the transform
img_t = transform(img)

# Check the tensor shape and type
print(f"Tensor shape: {img_t.shape}")
print(f"Tensor type: {type(img_t)}")
print(f"Tensor min/max: {img_t.min().item():.4f}/{img_t.max().item():.4f}")

Output

Tensor shape: torch.Size([3, 224, 224])
Tensor type: <class 'torch.Tensor'>
Tensor min/max: -2.1179/2.6400

Common Pitfalls

Common mistakes when using transforms include:

• Not converting images to tensor with ToTensor() before normalization.
• Applying normalization before converting to tensor causes errors.
• Using wrong mean and std values for normalization leads to poor model performance.
• For grayscale images, mean and std should have one value each.

Always apply ToTensor() before Normalize().

python

from torchvision import transforms

# Wrong order (will cause error or wrong results)
wrong_transform = transforms.Compose([
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    transforms.ToTensor()
])

# Correct order
correct_transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

Quick Reference

Transform	Purpose	Example Usage
Resize	Change image size	transforms.Resize(256)
CenterCrop	Crop center region	transforms.CenterCrop(224)
ToTensor	Convert image to tensor	transforms.ToTensor()
Normalize	Scale tensor values	transforms.Normalize(mean, std)
RandomHorizontalFlip	Flip image randomly	transforms.RandomHorizontalFlip(p=0.5)
RandomRotation	Rotate image randomly	transforms.RandomRotation(30)

Key Takeaways

Use torchvision.transforms.Compose to chain multiple image preprocessing steps.

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

Common transforms include Resize, CenterCrop, ToTensor, and Normalize.

Apply transforms to images before feeding them into your PyTorch model.

Check transform order carefully to avoid errors or poor model input.