An extension of the HLS library in RFNOC-HLS-NeuralNet to support 2-D convolutions. Also includes demo project(s).
Use Ristretto-Caffe to train and fine-tine a quantized deep network on the MNIST dataset.
Extract the weights using the provided scripts, and convert into a suitable format for HLS/SDK.
Use the provided scripts to generate an HLS project. Co-sim and verify your network.
The reference designs for Vivado IP Integrator, and sample code for Xilinx SDK are also provided.
The code for the Conv2D layer uses a linebuffer approach, inspired from here.