DataLoader basics in PyTorch - Model Pipeline Trace

Model Pipeline - DataLoader basics

This pipeline shows how raw data is loaded, prepared, and fed into a model using PyTorch's DataLoader. It helps handle data in batches, shuffle it, and make training efficient and smooth.

Data Flow - 3 Stages

1Raw Dataset

1000 samples x 28 x 28 pixels→Load images and labels from disk→1000 samples x 28 x 28 pixels

Image pixel values for handwritten digits and their labels (0-9)

↓

2Dataset Object

1000 samples x 28 x 28 pixels→Wrap raw data into a PyTorch Dataset class→1000 samples x 28 x 28 pixels

Dataset object that returns (image_tensor, label) pairs

↓

3DataLoader

1000 samples x 28 x 28 pixels→Batch data into groups of 32, shuffle samples→32 samples x 1 x 28 x 28 pixels per batch

Batch of 32 images and labels ready for training

Training Trace - Epoch by Epoch

Loss
1.2 |****
0.8 |***
0.5 |**
0.3 |*
0.25|*
    +------------
     Epochs 1-5

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning; loss is high, accuracy low
2	0.8	0.65	Loss decreases, accuracy improves as model learns
3	0.5	0.80	Training progressing well; model getting better
4	0.3	0.90	Loss low, accuracy high; model converging
5	0.25	0.92	Training stabilizes with good performance

Prediction Trace - 3 Layers

Layer 1: DataLoader batch fetch

Layer 2: Model input

Layer 3: Loss calculation

Model Quiz - 3 Questions

What does the DataLoader do with the dataset?

A Changes image sizes

B Groups data into batches and shuffles it

C Creates new labels

D Removes samples randomly

Key Insight

Using DataLoader helps efficiently feed data in batches to the model, enabling smooth training and better learning by shuffling and batching samples.