cnn-accelerator

A Convolutional Neural Network (CNN) hardware accelerator for image recognition. Tested with the CIFAR-10 dataset.

Directories:

• ./algorithm/cnn: CNN reference code in python used to classify images of the cifar-10 test_batch (for more instructions see ./algorithm/cnn/README).

• ./pre-hls: CNN pre-hls code

  • ./pre-hls/db: database folder with cifar10 images and labels.

  • ./pre-hls/vanilla: CNN implementation in C++, also tested with cifar-10 test_batch (for more instructions see ./pre-hls/vanilla/README).

  • ./pre-hls/fixed-point: CNN implementation in C++ using fixed poit precision with the library ac_fixed (for more instructions see ./pre-hls/fixed-point/README).

• ./hls: HLS code used to generate Convolution+ReLU RTL with Catapult.

Research links

• Deep Learning Book: Focus on Chap. 1, 6, 9, 11 and 12.

• Stanford Course CS231n: Convolutional Neural Networks for Visual Recognition

  • Project Report: Design and Analysis of a Hardware CNN Accelerator
  • Assignments 1, 2 and 3: experiments of someone based on course assignments.

• Hardware Acceleration of Convolutional Neural Networks

• AI GitBook - ReLU Layer

• Krizhevsky et al. - ImageNet Challenge Winner 2012

• Intuitive Explanation to ConvNets

• Convolution

Tutorials

• Python Numpy Tutorial
• C++ Template Tutorial