Data augmentation with transforms in PyTorch - Model Pipeline Trace

Model Pipeline - Data augmentation with transforms

This pipeline shows how data augmentation uses image transforms to create varied training images. This helps the model learn better by seeing different versions of the same image.

Data Flow - 3 Stages

1Original Dataset

1000 images x 3 channels x 32 height x 32 width→Load raw images from dataset→1000 images x 3 channels x 32 height x 32 width

Image of a cat with size 32x32 pixels and 3 color channels

↓

2Apply Data Augmentation Transforms

1000 images x 3 channels x 32 height x 32 width→Random horizontal flip, random rotation, random crop→1000 images x 3 channels x 32 height x 32 width

Original cat image flipped horizontally and slightly rotated

↓

3Batch Preparation

1000 images x 3 channels x 32 height x 32 width→Group images into batches of 100→10 batches x 100 images x 3 channels x 32 height x 32 width

Batch 1 contains 100 augmented cat and dog images

Training Trace - Epoch by Epoch

Loss
1.2 |*       
1.0 | *      
0.8 |  *     
0.6 |   *    
0.4 |    *   
    +---------
     1 2 3 4 5
     Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning with high loss and low accuracy
2	0.9	0.60	Loss decreases and accuracy improves as model sees augmented data
3	0.7	0.72	Model learns better features due to varied augmented images
4	0.55	0.80	Loss continues to drop, accuracy rises steadily
5	0.45	0.85	Model converges with good accuracy on augmented data

Prediction Trace - 5 Layers

Layer 1: Input Image

Layer 2: Random Horizontal Flip

Layer 3: Random Rotation

Layer 4: Random Crop

Layer 5: Normalized Tensor

Model Quiz - 3 Questions

Why do we apply random horizontal flips during data augmentation?

A To convert images to grayscale

B To reduce the image size

C To help the model learn to recognize objects flipped left to right

D To increase the number of color channels

Key Insight

Data augmentation creates varied versions of images, helping the model learn more robust features. This leads to better accuracy and lower loss during training because the model sees many different views of the same objects.