Skip to content

UEFI-code/MSRA_thePracticeSpaceProject_PyTorchCUDA

Repository files navigation

Simple But Super PyTorch CUDA Extension

You need to build before try Research*.py

Go to:

myLinear_CUDA_backend
myKakuritsu_Linear_backend
myConv1D_backend

To see the detailed information and instruction.

Research*.py is to Debug and Research the Code!

Research

Go to The Wiki to see Research Status.

Quick Link: My Project Report😎

Innovation

One Dim of CUDA blockID and threadID, Simple the MultiThread Idea.

Introduced the Kakuritsu Activation Method in Linear to Instead Dropout (Still need more ANN Experiment).

Good news!

You can try to build the CUDA code even you have NO GPU with my HACK!

My Cell Level Forward Available!

My Kaso Cell Backward Available!

My Convolution 1D Available!

CPU Inference Available!

Demo Code can run now!

Almost done, need a little understand of 'grad_weights' and Experiment Report.

Todo

Add beautiful figures to explaining the theory.

Add Omoshiroii CUDA operator Features.

Finish the Experiment Report.

About

My repo to attend MSRA the Practice Space Project 2022, CUDA Implement and Optimize

Topics

Resources

Stars

Watchers

Forks

Contributors 2

  •  
  •