GoogleNet Implementation

This repository demonstrates the training of a GoogLeNet model on the MNIST dataset using PyTorch. The dataset is resized and normalized to suit the GoogLeNet architecture, and the implementation achieves high accuracy.

---

GoogLeNet Architecture

GoogLeNet, also known as Inception v1, is a convolutional neural network architecture that uses inception modules to capture multi-scale features efficiently. This implementation uses the standard GoogLeNet variant.

---

Project Steps

1. Import Required Libraries

Set up the environment using the following commands:

conda create --name pytorch_env python=3.11.9 --file requirements.txt
conda activate pytorch_env

2. Load the MNIST Dataset

• The dataset is loaded using torchvision.datasets.MNIST.
• Images are resized to 32x32 (to fit GoogLeNet requirements).
• Images are normalized and converted to tensors.

3. Split the Dataset

The dataset is split as follows:

• Training Set: 42,000 samples
• Validation Set: 14,000 samples
• Test Set: 14,000 samples

4. Load Pre-trained GoogLeNet

• GoogLeNet is used with modifications:
  • Adjusted for single-channel (grayscale) input.
  • Output layer configured for 10 classes.
• The model is moved to GPU if available.

5. Define Loss Function and Optimizer

• Loss Function: CrossEntropyLoss
• Optimizer: Adam with a learning rate of 0.001

6. Train the Model

The model is trained for 10 epochs using the following key steps:

• The training set is passed through the model to compute predictions.
• The loss is calculated using CrossEntropyLoss.
• Gradients are computed and weights are updated using the Adam optimizer.

---

Results

Test Accuracy

• The model achieves a test accuracy of 99.04%.

Training and Validation Losses

Plots showing the training and validation loss over epochs:

Confusion Matrix

A heatmap of the confusion matrix shows the model's performance across all classes:

---

Requirements

• Python 3.11.9
• PyTorch
• torchvision
• numpy
• matplotlib
• PIL (Pillow)

For a detailed list of dependencies, refer to requirements.txt.

---

Acknowledgments

• PyTorch
• MNIST Dataset
• GoogLeNet Paper