FeedforwardNet

FeedforwardNet is a toolkit that simplifies the development and deployment of deep learning systems for FPGA based devices, either embedded or resource rich. The toolkit enables fast mapping pre-trained network models or applications on FPGA platforms and accelerate with pure logical hardware.

This toolkit contains a bottom up designed inference network library together with the pre-trained CAFFE model converter and network construction flow. This version of code is still under active development. Future version of well integrated toolkit with a web based GUI will be released soon.

File structure

Feedforward | |--example // folder contains demoe example designs and original CAFFE demos. |--fpga_cnn // main folder contains the cnn layer designs and CAFFE converter. |--stb_image // stb based image processing lib. |--README.md // This README file.

Status

FeedforwardNet is under active development and is currently un-finished.

Help and Support

If you have any suggestions and questions, please contact yao.chen@adsc.com.sg.

Getting Started

1. cd to FeedforwardNet/fpga_cnn/caffe_converter
2. run ./run.sh to generate the executable converter file. Pre-installed caffe and openCV is required in order to compile caffe converter.
3. cp ../../example/caffe_demos/lenet/lenet_2conv2max/lenet-1.prototxt ../../example/caffe_demos/lenet/lenet_2conv2max/lenet-1_1000_itr.caffemodel ./
4. run ./caffe_converter lenet-1.prototxt lenet-1_iter_10000.caffemodel to generate the weights.txt file.
5. Recently the automation of network construction is still under progress, so we have to construct the network model in fpga_cnn/construct_net.h as the provided example in it.
6. cp the weights.txt to example folder and rename it as weights_lenet.txt.
7. cp the input.txt(or example picture) from example/demo/lenet to example and rename as input_3.txt, it is the input image file.
8. change the compilation mode in config.h, HLS_MODE enables HLS process. for C compilation, comment HLS_MODE. If test with BATCH data, then comment KERNEL_MODE, otherwise the default is processing single picture instead of dataset.
9. cd to example folder and run ./run.sh to generate the executable file of the C++ design to test the correctness of the network model.
10. run ./fftest to predict the input file (or run the test dataset).
11. change the setting of HLS_MODE in config.h file to enable hardware synthesis.
12. cd to hls_impl folder and run ./syn.sh to generate and start the HLS project.

Bibtex:

@misc{feedforward2016,

author = {Yao Chen, Yang Yu, Chunrong Zhong},

title = {{FeedforwardNet: Enabling Efficient Convolutional Neural Network Application Design on Embedded FPGAs, Version 0.1}},

year = {2016},

url = {https://github.com/microideax/FeedforwardNet} }