LeNet-PyTorch with DLProf Profiling

This repository implements the classic LeNet-5 architecture using PyTorch for digit classification on the MNIST dataset. Additionally, it leverages NVIDIA DLProf for performance profiling and optimization, providing insights into GPU utilization and performance bottlenecks.

---

📂 Project Structure

• lenet.py: Main script for training, validation, and testing the LeNet-5 model.
• reference.txt: Instructions for using Docker and DLProf.
• results/: Contains profiling results and analysis (e.g., dlprof_dldb.sqlite).
• examples/: Jupyter notebooks for testing or visualizing results.
• assets/: Images and media for documentation (e.g., graphs, architecture diagrams, and output examples).

---

🔑 Key Features

• LeNet-5 Implementation: A PyTorch-based implementation of the classic architecture.
• Data Augmentation: Techniques like random rotations and crops to improve model generalization.
• DLProf Integration: Performance profiling using NVIDIA DLProf for detailed insights.
• GPU Optimization: Accelerated training using apex for mixed precision.
• Dockerized Workflow: Easily reproducible and portable setup with Docker.

---

🖼️ Sample Outputs

Here are some sample outputs from the model:

Original Image	Model Prediction

Additional examples of predictions:

Example 1	Example 2	Example 3
Example 4	Example 5	Example 6

---

🚀 Getting Started

Prerequisites

• Python >= 3.6
• PyTorch >= 1.10
• NVIDIA GPU with CUDA support
• Docker (for profiling)

Installation

1. Clone this repository:

   git clone https://github.com/sarkarsaswata/LeNet-PyTorch-DLProfiler.git
   cd LeNet-PyTorch-DLProfiler

2. Clone the Ultra-Light-Fast-Generic-Face-Detector-1MB repository into the directory:

   git clone https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB

3. Build the Docker image:

   docker build . -f Dockerfile --no-cache -t lenet_profiler

4. Run the profiling script using the provided run.sh:

   sh run.sh

---

📊 Profiling with DLProf

Generating Profiling Data

Run the following command to generate profiling data using DLProf:

dlprof --mode=pytorch -f true --iter_start 20 --iter_stop 80 python lenet.py

Viewing Profiling Results

Use dlprofviewer to visualize the profiling data:

dlprofviewer dlprof_dldb.sqlite

---

🛠️ Customization

Feel free to modify:

• Model: Experiment with different architectures or hyperparameters in lenet.py.
• Data: Replace MNIST with other datasets by adapting the train_transforms and test_transforms.
• Profiling: Adjust DLProf parameters for specific profiling scenarios.

---

📜 License

This project is licensed under the MIT License.

---