Image Transformations and Augmentation Basics

Build the mental model

A transform in torchvision is simply a function (or a small callable object) that takes an image tensor and returns a modified one, and the transforms.v2 API is the current, actively maintained way to write them — it works uniformly whether you're transforming a plain image tensor, a batch, or, when needed, paired data like bounding boxes and segmentation masks. Some transforms are shape-preserving: RandomHorizontalFlip reorders pixels left-to-right but the tensor still comes out (C, H, W) with the same H and W it went in with; Normalize and color-jitter operations rescale pixel values without touching spatial dimensions at all. Other transforms are shape-changing: Resize explicitly produces a different H and W, and RandomCrop/CenterCrop cut the tensor down to a smaller region. Compose chains several transforms into one callable, applying them in order — exactly the pipeline a DataLoader runs on every sample it yields, so understanding which stage changes shape and which doesn't matters for debugging batch-shape mismatches later.

The reason augmentation transforms like flips, crops, and color jitter exist at all is that they synthetically expand a training set's diversity without collecting a single new photo: randomly flipping a training image left-to-right doesn't change what object is in it, so training on both versions teaches the network that orientation shouldn't matter for the label — a form of invariance you're deliberately injecting rather than hoping the network discovers on its own from a fixed dataset. Crucially, these random transforms are applied stochastically and only during training; at evaluation or inference time you want a deterministic, repeatable transform (fixed resize, no random flip) so that the same input always produces the same prediction and your reported metrics don't jitter run to run just because of random augmentation.

Connect it to a real scenario

The Tutorial Platform generates three thumbnail sizes — a grid-card thumbnail, a hero banner, and a mobile-view crop — from a single instructor-uploaded course image by running it through a Resize-based pipeline at page-render time, and separately maintains an internal 'course-image quality' classifier trained with RandomHorizontalFlip and small random crops added to its training set, so the model doesn't accidentally learn that only left-facing or perfectly centered banners count as good thumbnails.

• transforms.v2 transforms are plain callables you chain with Compose, applied in order to an image tensor.
• Some transforms preserve tensor shape (flip, normalize, color jitter); others deliberately change it (resize, crop, pad).
• Random augmentations should run only during training — at eval/inference you want deterministic transforms so predictions are reproducible.

Try the working example

import torch
from torchvision.transforms import v2

torch.manual_seed(0)

# Synthetic batch: a single fake RGB image, 3 channels, 64x64.
image = torch.rand(3, 64, 64)

flip = v2.RandomHorizontalFlip(p=1.0)  # p=1.0 makes this deterministic for the demo
flipped = flip(image)
print("original shape:", image.shape)
print("flipped shape:", flipped.shape)
print("flip changed pixel order:", not torch.equal(image, flipped))

resize = v2.Resize((32, 32))
resized = resize(image)
print("resized shape:", resized.shape)

# Compose chains several transforms into one callable pipeline --
# this is exactly the kind of pipeline a DataLoader applies per sample.
pipeline = v2.Compose([
    v2.Resize((32, 32)),
    v2.RandomHorizontalFlip(p=0.5),
])
augmented = pipeline(image)
print("pipeline output shape:", augmented.shape)

Prints original shape: torch.Size([3, 64, 64]), flipped shape: torch.Size([3, 64, 64]) (unchanged — flip only reorders pixels), flip changed pixel order: True (guaranteed since p=1.0 forces the flip and a random continuous image is essentially never left-right symmetric), resized shape: torch.Size([3, 32, 32]), and pipeline output shape: torch.Size([3, 32, 32]) (the resize inside the pipeline always applies, so the output shape is the same regardless of whether the 50%-probability flip happened to trigger).

PyTorch Vision Docs — Transforming images, videos, boxes and more — Computer Vision