himewel / yolowell

A set of hardware architectures to build a co-design of convolutional neural networks inference at FPGA devices

yolowell is a lib to generate VHDL code of convolutional networks and becomes part of a co-design hardware/software with a simplified version of Darknet running in a NIOS II processor. To improve parameter passing and generate different versions of the architecture better then with TCL or other type of script, we implement all the code in python language with HWToolKit.

How to use

To generate the VHDL code, you will need to extract the weights, bias, and batch normalization params from your Darknet model and write him in binary format. After this, you can set a yaml config file as the following example.

weights_path: "./binary/weights.pickle"
variance_path: "./binary/variance.pickle"
mean_path: "./binary/mean.pickle"
scale_path: "./binary/scale.pickle"
biases_path: "./binary/biases.pickle"
output_path: "./generated"
channels: 3
layer_groups:
- filters: 16
  layers:
  - type: "conv_layer"
    parallelism: 4
    size: 3
    binary: False
  - type: "max_pool_layer"
- filters: 32
  layers:
  - type: "conv_layer"
    size: 3
    binary: True
  - type: "max_pool_layer"

So, you need to set some variables:

• output_path: path to the generate vhdl files;
• weights_path: file path to the float weight values in binary format;
• bn_variance_path: file path to the float variance values from batch normalization in binary format;
• bn_mean_path: file path to the float mean variance from batch normalization values in binary format;
• scale_path: file path to the float scale values in binary format;
• biases_path: file path to the float biases values in binary format;
• channels: set the input channels of the architecture;
• filters: number of filters in the current block of layers (layer_groups will a list of dicts);
• type: "conv_layer" or "max_pool_layer";
• size: size of the filters (e.g., 3x3 -> 3, 1x1 -> 1);
• binary: type of operations, false to use multipliers, true to use xor gates;

If you are still here, import NetworkParser and be happy (or not):

from components.network_parser import NetworkParser
from utils import get_std_logger, to_vhdl

get_std_logger()
net = NetworkParser("config.yaml")
layers = net.parse_network()
net.generate(layers, to_vhdl)

References

• Darknet: https://github.com/AlexeyAB/darknet;
• HWToolKit: https://github.com/Nic30/hwt;
• A High-Throughput and Power-Efficient FPGA Implementation of YOLO CNN for Object Detection: https://ieeexplore.ieee.org/document/8678682;
• XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks: https://arxiv.org/abs/1603.05279;