Data Augmentation in Practice

Build the mental model

The core idea behind data augmentation is showing the model a slightly-randomly-transformed version (flip, rotate, color shift, etc.) of each training image rather than the raw pixels every time. Because the model sees a different version of the same image on nearly every epoch, it becomes much harder for it to memorize exact pixel patterns, and it's pushed instead toward learning the more stable visual structure — shape, texture arrangement — that actually determines the class. This achieves a similar regularizing effect to dropout but through a different mechanism: dropout perturbs the network's internals, while augmentation effectively expands the training distribution itself.

For image data specifically, augmentation works so well because it mimics real-world variation that shouldn't change an object's identity — the angle a photo was taken from, the lighting, whether the subject faces left or right. A flipped cat, or a cat photographed under different contrast, is still a cat. Exposing the model to all of these variations during training — using transforms that are applied only at training time, never at inference — makes it substantially more robust to variation it hasn't literally seen before in the test set.

Connect it to a real scenario

Suppose the Tutorial Platform only has a small handful of user-submitted lesson diagrams (flowcharts, architecture sketches) labeled for training an internal 'low-quality diagram' classifier. With so few samples, applying this kind of augmentation pipeline (flip, slight rotation, color jitter) to each training image effectively expands the dataset and keeps the classifier from simply memorizing the handful of diagrams it has, instead of learning what actually makes a diagram look low-quality.

• torchvision.transforms.v2.Compose chains multiple random transforms into one pipeline.
• Calling the same pipeline on the same input repeatedly yields different outputs each time — direct proof of its randomness.
• Augmentation effectively expands the training distribution, discouraging the model from memorizing exact pixel patterns.

Try the working example

import torch
import torchvision.transforms.v2 as T

torch.manual_seed(0)

transform = T.Compose([
    T.RandomHorizontalFlip(p=0.5),
    T.RandomRotation(degrees=15),
    T.ColorJitter(brightness=0.3, contrast=0.3),
])

# A single fake RGB image, 3x64x64, values in [0, 1]
image = torch.rand(3, 64, 64)

for i in range(3):
    augmented = transform(image)
    print(f"call {i}: shape={tuple(augmented.shape)}, mean={augmented.mean().item():.4f}")

Prints 3 lines, each showing shape=(3, 64, 64) — the spatial size never changes. The mean value, however, differs across all three calls, because each call independently samples new random flip/rotation/color-jitter parameters (the original input tensor itself is left unchanged since these transforms return a new tensor).

Wikipedia — Data augmentation — Computer Vision