Dataset class (custom datasets) in PyTorch - Model Pipeline Trace

Model Pipeline - Dataset class (custom datasets)

This pipeline shows how a custom dataset class in PyTorch loads and prepares data for training a model. It starts with raw data, processes it into usable form, feeds it into a model, and tracks training progress.

Data Flow - 4 Stages

1Raw Data Loading

1000 rows x 2 columns→Load raw data from CSV file→1000 rows x 2 columns

[{'feature': 0.5, 'label': 1}, {'feature': 0.3, 'label': 0}, ...]

↓

2Custom Dataset Initialization

1000 rows x 2 columns→Create Dataset object with __len__ and __getitem__→Dataset object with 1000 samples

dataset[0] -> (tensor([0.5]), tensor(1))

↓

3DataLoader Batching

Dataset with 1000 samples→Batch data into groups of 32 samples→Batches of 32 samples each

batch[0] -> (tensor([[0.5], [0.3], ...]), tensor([1, 0, ...]))

↓

4Model Training

Batch of 32 samples→Feed batch to model, compute loss, update weights→Updated model weights

Model output logits for batch of 32

Training Trace - Epoch by Epoch

Loss
0.7 |****
0.6 |*** 
0.5 |**  
0.4 |*   
0.3 |    
0.2 |     
     1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.65	0.60	Model starts learning, loss is high, accuracy moderate
2	0.48	0.75	Loss decreases, accuracy improves
3	0.35	0.85	Model converging, better predictions
4	0.28	0.90	Loss low, accuracy high
5	0.22	0.93	Training stabilizes with good performance

Prediction Trace - 5 Layers

Layer 1: Dataset __getitem__

Layer 2: Batching in DataLoader

Layer 3: Model forward pass

Layer 4: Loss calculation

Layer 5: Backpropagation and optimizer step

Model Quiz - 3 Questions

What does the __getitem__ method in a custom Dataset class return?

A A single data sample and its label as tensors

B The entire dataset at once

C Only the labels of the dataset

D The batch size

Key Insight

Creating a custom Dataset class lets you control how data is loaded and accessed. This flexibility helps feed data correctly into models, enabling smooth training and better learning.