PyTorch Image Classification Practice: Learning the Art of Balancing Regularization and Generalization with Small Datasets

This article introduces ishaandindwar's PyTorch image classification project on GitHub. Using a self-built small dataset of 104 images across four categories (bottles, headphones, Spider-Man dolls, watches), it constructs a CNN model and experimentally explores the practical effects of overfitting, regularization techniques, and hyperparameter tuning. The core is to understand the art of balancing regularization and generalization with small datasets.

Project Source: Original author/maintainer ishaandindwar, platform GitHub, original title image-classifier-neural-network, link https://github.com/ishaandindwar/image-classifier-neural-network, release date June 13, 2026.

Learning Motivation: Understand how neural network training behavior changes with parameters and regularization techniques; the small dataset setup makes it easy to observe overfitting.

Dataset Features: 26 images per category, 4 categories total (104 images), taken under different angles and lighting conditions; advantages of self-built dataset: controllable quality, fast iteration, full understanding of data.

Network Structure: Classic CNN including convolutional layers (extract spatial features), batch normalization (accelerate convergence + regularization), pooling layers (dimensionality reduction + translation invariance), Dropout layers (prevent overfitting), fully connected layers (classification decisions).

Training Configuration: Optimizer Adam, learning rate 0.001, batch size 16, training epochs 15, Dropout rate 0.3, loss function cross-entropy loss; uses training-validation split, backpropagation to update weights.

Experiment 1: Initial training-validation accuracy is about 71%, with high training accuracy but low validation accuracy—typical overfitting.

Experiment 2: After applying L2 regularization (weight decay 1e-4), validation accuracy drops to about 43% due to underfitting caused by the overly small dataset, revealing that regularization is not always beneficial.

Experiment 3: After comprehensive tuning (adjusting Dropout, learning rate, epochs, batch size), validation accuracy reaches about 76% and the loss gap narrows; optimal training duration is 12-14 epochs, as further training leads to overfitting.

Core Takeaways: Lowering training loss does not equal improving validation accuracy; overfitting and generalization need to be balanced (empirical risk/structural risk minimization, bias-variance tradeoff); hyperparameters (learning rate, batch size, etc.) have significant impacts; regularization is a double-edged sword (easy to underfit with small datasets).

Project Value: Educational value (moderate scale, clear problem, complete experiments, detailed records); practical insights (start with small datasets, monitor training dynamics, use regularization cautiously, apply early stopping, record experiments).

Expandable directions:

1. Data augmentation (rotation, flipping, etc.)
2. Transfer learning (pre-trained models like ResNet)
3. Complex architectures (ResNet residual connections, Inception modules)
4. Learning rate scheduling
5. Cross-validation
6. Expand more categories

Although small in scale, this project has rich learning value, showing overfitting issues in small datasets and attempted solutions; the author not only learned to build a CNN with PyTorch but also understood the difference between overfitting and generalization, limitations of regularization, impact of hyperparameters, and importance of experimental records; hands-on experiments plus observation and reflection are far better than theoretical reading; the core truth is that models need to balance learning and regularization to achieve good generalization.