This project is from Magma Hackathon by Yujun Lin, Kaidi Cao and Song Han
This design implements a one-layer binary neural network for handwritten digits recognition (0-9). Both the weights and activations are binary which is very computation-efficient. We also provided the Tensorflow code to train binary neural networks under nn_train
. However, if you are only interested in the hardware part, you can use our pre-trained model under nn_train/BNN.pkl
You may start with tutorial_digits_recognition_on_icestick.ipynb
which will guide you through the entire flow.
The architecture is output stationary. The image is resized to 16x16, and operand width in execution stage of pipeline is 16 bit. The pipeline contains five stages:
- IF: calculate IDX and CYCLE as shown in figure
- RR: read weight matrix and image block from ROM and LUTs
- EXE: using NXOR and Popcount to perform binary matrix multiplication and accumulate results
- CP: compare result to previous maximum result and save the index (prediction number) for maximum result
- FI: wait to show result until all calculation completes
There are five lights (D0, D1, D2, D3, D4) on the IceStick FPGA. The D5 LED is green which indicates the finish of calculation. Others are red and used for indicates binary representation of predited number.
nn_train
contains tutorials of training a binary neural network for digits recognition and format of saving weight matrix and images used for FPGA.
nn_train/BNN.pkl
contains the weight matrix and images we seleted. Only digit 4 is incorrectly recognized as 9 and other numbers can be perfectly recognized.
test_modules
shows the simulation of our pipeline.
tutorial_digits_recognition_on_icestick
is the tutorial to convert our code to verilog and download to icestick FPGA for experiments.
The MIT License
Copyright (c) 2017 Yujun Lin, Kaidi Cao, Song Han
Contact: songhan@mit.edu, yujunlin@mit.edu