TinyVGG and FashionMNIST: Image Classification Practice from Linear Baseline to Convolutional Networks

Project Source Original Author/Maintainer: Siva-Sainath Source Platform: GitHub Original Project Title: tinyvgg-fashionmnist-classifier Original Link: https://github.com/Siva-Sainath/tinyvgg-fashionmnist-classifier Release Time: 2026-06-14

Core Guide This project implements the TinyVGG convolutional neural network using the PyTorch framework for the FashionMNIST fashion item classification task. It compares the performance differences between the linear baseline model and CNN, demonstrates the complete training process and visualization analysis, helping to understand the development context of deep learning from linear to convolutional models and the advantages of CNN in image tasks.

FashionMNIST Dataset Features FashionMNIST contains 70000 28x28 grayscale images covering 10 categories of fashion items (T-shirts, pants, sweaters, etc.), with 60000 for training and 10000 for testing, and a balanced category distribution. Compared to MNIST, its textures and shapes are more complex, making it difficult for linear models to achieve ideal results.

Data Preprocessing Key Points

1. Normalization: Scale pixel values from [0,255] to [0,1] or [-1,1] to improve convergence speed and numerical stability.
2. Data Augmentation: Expand training data through random rotation, translation, and flipping to enhance generalization ability.
3. Batch Processing: Use PyTorch DataLoader for efficient batch loading, supporting multi-threaded prefetching and data shuffling.

TinyVGG Network Architecture TinyVGG is a lightweight CNN inspired by VGG but with fewer parameters. Core components include:

• Convolutional Layers: 3x3 convolution kernels + BatchNorm + ReLU activation, stacked to extract features.
• Pooling Layers: 2x2 max pooling, halving the feature map size, retaining significant features and reducing computational load.
• Fully Connected Layers: Input flattened features, output 10-class probabilities, combined with Dropout to prevent overfitting. The structure follows the repeated "convolution-convolution-pooling" pattern, with channel numbers from 32→64→128, extracting visual features from low-level to high-level.

Training Optimization Strategy

• Loss Function: Cross-entropy loss (measures the difference between predicted distribution and true labels).
• Optimizer: Adam (combines momentum and adaptive learning rate, with learning rate decay).
• Training Loop: Custom training/validation loop, monitor validation loss; early stopping mechanism to avoid overfitting.

Visualization Analysis

• Loss Curves: Show training/validation loss changes with epochs to judge convergence and overfitting.
• Accuracy Curves: Directly reflect the trend of model performance improvement.
• Confusion Matrix: Identify easily confused categories (e.g., shirts and T-shirts) to provide direction for improvement.

Model Comparative Evidence The linear baseline model (input flattened images into fully connected layers) achieves an accuracy of about 80% on FashionMNIST, while TinyVGG can easily exceed 90%, clearly demonstrating the advantage of CNN in capturing spatial features.

Practice Insights

• Modular Design: Separate logic such as data preprocessing, model definition, and training loops to improve code readability and maintainability.
• Experiment Records: Save hyperparameters and performance results for easy comparative analysis and tuning.

Project Summary This project is a well-structured deep learning teaching case that demonstrates the powerful capabilities of CNN. Through comparison with linear baselines, it helps understand why CNN is superior to traditional methods, making it an ideal starting point for learning PyTorch and computer vision.

Expansion Direction Suggestions

1. Transfer Learning: Load ImageNet pre-trained weights and fine-tune on FashionMNIST to improve results.
2. Architecture Improvement: Introduce modern components such as residual connections (ResNet) and attention mechanisms (SE Block).
3. Network Adjustment: Try deeper or wider network structures to further improve classification accuracy.