Optimizing MLP Kernels

This project involves implementing various versions of a Multilayer Perceptron (MLP) kernel to improve execution time while maintaining accuracy.

Project Structure

• resources/
• scripts/
  • infer.sh
  • profile.sh
  • pytorch_train.sh
• src/
  • application_training/
    • train.py
  • kernels/
    • __init__.py
    • improved_kernel.py
    • naive_kernel.py
    • tensor_kernel.py
  • model_weights/
    • model.pt
  • results/
    • batch_times_mem.png
    • batch_times_proc.png
    • report_1024.ncu-rep
    • report.ncu-rep
    • time_taken_hidden_proc.png
    • time_taken_hidden.png
  • analysis.py
  • config.yaml
  • inference.py
  • model.py
  • profiler.py
  • utils.py
• .gitignore
• ReadME.md

Description

The goal is to optimize different implementations of MLP kernels to achieve better execution times without sacrificing accuracy.

Directories and Files

• scripts/: Contains shell scripts for training, inference, and profiling.
  • pytorch_train.sh: Script to train the pytorch MLP model to test performance.
  • infer.sh: Script to perform inference using the trained model.
  • profile.sh: Script to profile the MLP kernels.
• src/: Source code and related files.
  • application_training/: Training application scripts.
    • train.py: Training script for the pytorch MLP model to test performance.
  • kernels/: Different MLP kernel implementations.
    • naive_kernel.py: Basic MLP kernel implementation.
    • improved_kernel.py: Optimized MLP kernel.
    • tensor_kernel.py: MLP kernel using tensor operations.
  • model_weights/: Contains saved model weights.
    • model.pt: Trained model file.
  • results/: Profiling reports and result images.
    • batch_times_mem.png, batch_times_proc.png: Batch time analysis plots.
    • report.ncu-rep, report_1024.ncu-rep: Profiling reports.
    • time_taken_hidden.png, time_taken_hidden_proc.png: Hidden layer time analysis plots.
  • analysis.py: Script for analyzing results and plotting graphs.
  • config.yaml: Configuration file for the project.
  • inference.py: Script to run inference.
  • model.py: Defines the MLP model architecture.
  • profiler.py: Profiling utilities.
  • utils.py: Helper functions.
• tensor_kernel_correctness_check: Verifies that the output of tensor_kernel matches actual numpy implementation.

Getting Started

Prerequisites

• Python 3.x
• PyTorch
• CUDA Toolkit (if using GPU acceleration)

Installation

Clone the repository and navigate to the project directory:

git clone https://github.com/yourusername/e4750-2024fall-project-nerf.git
cd e4750-2024fall-project-nerf

Install the required Python packages:

pip install -r requirements.txt

Running the Project

• Training:

  bash scripts/pytorch_train.sh

• Inference:

  bash scripts/infer.sh

• Profiling:

  bash scripts/profile.sh

Results

Results and analysis can be found in the src/results/ directory.

License

This project is part of the EECSE4750 course.