EN605617-FinalProject

Yujiang's final project of the course introduction to GPU programming at JHU

Project Description

In this project, I explore how to take advantage of CUDA for data preprocessing and Neural Network training.

For the data preprocessing, I

1. wrote a kernel function in CUDA (grayscaleimagenormalization_kernel.cu)
2. wrote the host side caller (grayscaleimagenormalization_kernel.cu,launch_grayscaleimagenormalization.h)
3. wrote an python,C++ interfacing file (grayscaleimagenormalization_ops.cpp)
4. compiled our custom cuda_module using JIT option
5. normalized the pixel of raw gray scale image,which is between [0,255], to a number between [-1,1] by calling the custom cuda kernel from pytorch

I also compared timing among pixel normalizer function implemented in

1. pytorch called custom CUDA kernel,
2. pure pytorch tensor running on CPU,
3. pure pytorch sensor running on GPU.

Then, I trained a Neural network using pytorch to do dignit recognition based on the data preprocessed by the custom pixel normalizer defined in CUDA kernel.

For Neural Network Trainig and Testing, I

1. wrote a neural network in C++ targeting CPU operation (Largely borrowed the work from BobMcDear, See reference 1)
2. wrote a neural network in C++/CUDA targeting GPU operation (Largely borrowed the work from BobMcDear, See reference 1)
3. compared time used on CPU and GPU to train NN for MNIST digit recognition

additional notes:

• To run the project, create a new conda environment, activate the conda environment, install pip under conda environment by conda install pip, then type pip install -r requirements.txt to install the required python modules
• The data t10k-images-idx3-ubyte,t10k-labels-idx1-ubyte,train-images-idx3-ubyte,train-labels-idx1-ubyte can be downloaded from http://yann.lecun.com/exdb/mnist/. A local copy is avaliable in the data folder
• The csv datasets used by ./CPU/min_minist.cpp and ./GPU/min_minist.cu are too large to be uploaded onto github. They can be generated by running the ./pytorch/onehotencodelabel.ipynb
• For CUDA kernel for data preprocessing and Neural Network Training for MNIST Digit Recognition using PyTorch, just run ./pytorch/digitrecognition.ipynb.
• For CUDA/C++ based Neural Network for MNIST digit recognision,
  • run the ./pytorch/onehotencodelabel.ipynb first to generate the required csv files
  • then use the makefile in ./CPU to compile executable cpukernel
  • and use makefile in ./GPU to compile executable gpukernel

Future Work

References:

• Neural network in CUDA/C++
• CUSTOM CUDA Example
• CUSTOM C++ AND CUDA EXTENSIONS
• MNIST Handwritten Digit Recognition in PyTorch
• pybind11 — Seamless operability between C++11 and Python
• PyTorch C++ API