Tiny-VGG Convolutional Neural Network on Fashion-MNIST

A compact convolutional neural network inspired by VGG-like architectures, trained on the Fashion-MNIST dataset for image classification tasks.

---

📚 Overview

The Tiny-VGG is a simplified version of the classical VGG-16 network, designed to be lightweight and easy to train on small datasets such as Fashion-MNIST.

It aims to:

• Demonstrate the power of convolutional feature extraction.
• Provide an educational implementation for deep learning beginners.
• Achieve high accuracy with minimal parameters.

---

⚙️ Architecture

Layer Type	Configuration	Output Shape
Input	1 × 28 × 28 grayscale image	1 × 28 × 28
Conv2D + ReLU	32 filters, 3×3 kernel, padding=1	32 × 28 × 28
Conv2D + ReLU	32 filters, 3×3 kernel, padding=1	32 × 28 × 28
MaxPool2D	2×2	32 × 14 × 14
Conv2D + ReLU	64 filters, 3×3 kernel, padding=1	64 × 14 × 14
Conv2D + ReLU	64 filters, 3×3 kernel, padding=1	64 × 14 × 14
MaxPool2D	2×2	64 × 7 × 7
Flatten	—	3136
Linear + ReLU	3136 → 128	128
Dropout(0.3)	—	128
Output (Linear)	128 → 10	10 classes

---

🧩 Implementation (PyTorch)

from torch import nn

class TinyVGG(nn.Module):
    def __init__(self, in_channels: int, hidden_units: int, out_channels: int):
        super().__init__()
        self.conv_block_1 = nn.Sequential(
            nn.Conv2d(in_channels, hidden_units, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(hidden_units, hidden_units, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2)
        )
        self.conv_block_2 = nn.Sequential(
            nn.Conv2d(hidden_units, hidden_units * 2, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.Conv2d(hidden_units * 2, hidden_units * 2, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2)
        )
        self.classifier = nn.Sequential(
            nn.Flatten(),
            nn.Linear(hidden_units * 2 * 7 * 7, 128),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(128, out_channels)
        )

    def forward(self, x):
        x = self.conv_block_1(x)
        x = self.conv_block_2(x)
        x = self.classifier(x)
        return x