Data loading with torchvision in Computer Vision - Model Pipeline Trace

Model Pipeline - Data loading with torchvision

This pipeline shows how image data is loaded and prepared using torchvision. It starts with raw image files, applies transformations, and prepares batches for training a model.

Data Flow - 3 Stages

1Raw Image Dataset

1000 images (varied sizes)→Load images from disk folder→1000 images (varied sizes)

Image file: 'dog_001.jpg', size 500x400 pixels

↓

2Apply Transformations

1000 images (varied sizes)→Resize to 224x224, convert to tensor, normalize pixel values→1000 images x 3 channels x 224 x 224

Image tensor shape: (3, 224, 224), pixel values normalized with mean and std

↓

3Create DataLoader

1000 images x 3 x 224 x 224→Batch images into groups of 32, shuffle data→32 images x 3 x 224 x 224 per batch, 31 batches total

Batch 1 shape: (32, 3, 224, 224)

Training Trace - Epoch by Epoch

Loss
1.2 |****
0.9 |***
0.7 |**
0.5 |*
0.4 |

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning, loss high, accuracy low
2	0.9	0.60	Loss decreases, accuracy improves
3	0.7	0.72	Training progressing well
4	0.5	0.80	Model learning features effectively
5	0.4	0.85	Good convergence, ready for evaluation

Prediction Trace - 4 Layers

Layer 1: Input Image Tensor

Layer 2: Model Forward Pass

Layer 3: Softmax Activation

Layer 4: Prediction

Model Quiz - 3 Questions

What is the shape of the image tensor after transformations?

A 3 x 224 x 224

B 224 x 224 x 3

C 32 x 3 x 224 x 224

D 1000 x 3 x 224 x 224

Key Insight

Using torchvision for data loading simplifies preparing images for training. Transformations like resizing and normalization ensure consistent input for the model, which helps training converge faster and more reliably.