search

hughperkins / DeepCL

OpenCL library to train deep convolutional neural networks
Mozilla Public License 2.0
869 stars 200 forks source link

Implementation of DNN on FPGA #154

Open anajar2198 opened 3 years ago

anajar2198 commented 3 years ago

Hi I have a trained DNN which weight is extracted on a .h5 file. What's the road map in order to implement this DNN on a FPGA?

karanmathur commented 3 years ago

Hi did you find how to go about implementing the DNN on an FPGA? If you did can you please share, it would be a great help.

anajar2198 commented 3 years ago

Hi did you find how to go about implementing the DNN on an FPGA? If you did can you please share, it would be a great help.

Hi There are several ways to implement a DNN on FPGA

  1. Using OpenCL and providing kernels of layers and a host code. finally you can implement on trasic SOC boards for more information about this method you can refer to bellow link https://github.com/Er1cZ/Deploying_CNN_on_FPGA_using_OpenCL/blob/master/GettingStartedTutorial.md

  2. Using VitisAI and freezing your DNN, then Quantize it and then convert it to FPGA codes using VitisAI. the disadvantages of this method is that FPGAs that VitisAI supports are expensive. if you don't have money problem, you can consider this method too.

  3. the next solution is using board raspberry pi. you can use linux in there and install python packages such as tensorflow, pythorch, numpy and other requirements. Finally using connecting integrated camera you can design a real-time deep learning based engine.

  4. and the final solution that i know. using another board that supports linux. Using jetson nano and using a little graphic card that can be integrated to it. Using this tools and also TensorRT, you can deploy your DNN on jetson nano. TensorRT accelerates your DNN and converts your weights to FP32, FP16 and INT8.

these was the solutions that I found before. There are also other and maybe better solutions if you continue searching. I hope I could help you.