PyTorch Experiment Tracking (Learning Purpose)

📌 Overview

This is a learning project to understand how to track multiple experiments efficiently in PyTorch.
It is based on the course PyTorch for Deep Learning (Udemy).

The project uses EfficientNet-B0 and EfficientNet-B2 on the Pizza, Steak, and Sushi dataset,
which is a subset of the original Food101 dataset.

It demonstrates:

• Experiment reproducibility
• TensorBoard tracking
• Automatic dataset preparation

✨ Features

• ✅ Dataset Download & Preprocessing (10% & 20% splits)
• ✅ EfficientNet-B0 and B2 (ImageNet-pretrained)
• ✅ Transfer Learning with Frozen Feature Extractor
• ✅ TensorBoard Logging (loss, accuracy, graph)
• ✅ Custom Classifier Heads for 3-class Output
• ✅ Reproducibility with Random Seeds
• ✅ Modular Codebase and CLI Support

🗂️ Project Structure

.
├── artifacts            # Saved models
│   └── models
├── data                 # Downloaded datasets
│   ├── pizza_steak_sushi_10
│   └── pizza_steak_sushi_20
├── LICENSE
├── README.md
├── runs                  # TensorBoard logs
│   └── 2025-Aug-14
└── src
    ├── data_module.py    # DataLoader creation
    ├── data_setup.py     # Dataset download and extraction
    ├── engine.py         # Train/test steps
    ├── __init__.py
    ├── model_builder.py  # EfficientNet model creation
    ├── predictions.py    # Single image prediction and visualization
    ├── train.py          # Main training orchestration (CLI + logging + experiment loop)
    └── utils.py          # Utility functions (saving, logging, plotting)

⚙️ Installation

1. Clone the Repository

git clone https://github.com/Niloy009/pytorch_experiment_tracking.git
cd pytorch_experiment_tracking

2. Install Dependencies

pip install torch torchvision torchmetrics matplotlib requests tqdm tensorboard

🚀 Usage

Train Models (Default Settings)

python -m src.train

Custom Experiment Example

python -m src.train \
    --models effnetb0 effnetb2 \
    --data_pct 10 20 \
    --epochs 5 10 \
    --batch-size 32 \
    --lr 0.001 \
    --seed 42 \
    --data-root ./data \
    --artifacts-dir ./artifacts \
    -vv

📊 TensorBoard Logging

Start TensorBoard to monitor metrics:

tensorboard --logdir runs

Then navigate to http://localhost:6006

🔍 Predict on Single Image

from predictions import pred_and_plot_image
from model_builder import create_effnetb0
import torch

model = create_effnetb0()
model.load_state_dict(torch.load("path_to_model.pth"))
class_names = ["pizza", "steak", "sushi"]
pred_and_plot_image(model, class_names, "path_to_image.jpg")

📈 Tracked Metrics

• Train/Test Loss per Epoch
• Train/Test Accuracy per Epoch
• Model Graph Summary

🧠 Implementation Details

• Optimizer: Adam
• Loss: CrossEntropyLoss
• Accuracy: torchmetrics.Accuracy
• Transforms: Model-specific (from torchvision weights)
• Device: Automatic CUDA/CPU selection
• Seeded Execution: torch.manual_seed()

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.

👤 Author

Niloy Saha Roy
Machine Learning & Deep Learning Engineer
Email: niloysaha.887@gmail.com