This is an example of a Quantized Tiny YOLO v2 on FPGA using the Xilinx FINN framework, specifically BNN-PYNQ.
The board targeted is ZC706 (Zynq 7z045).
You need an understanding of that repo to run this example successfully.
(Tiny yolo with depthwise separable convolution has been introduced, inorder to have fast inference )
We use W1A4.
We use DoReFa-Net like training, except that we use channel-wise mean for the weights as in XNOR-Net.
Please refer to the QNN-MO-PYNQ Tiny YOLO example for the topology.
We use a custom PyTorch implementation of Darknet called Lightnet.
Follow the instructions at https://eavise.gitlab.io/lightnet/notes/02-examples.html#pascal-voc to prepare the PASCAL VOC dataset for use with Lightnet.
Launch training from the Training/Lightnet/examples/yolo-voc folder with
python3 bin/train.py -c -n cfg/tinyyolo.py /path/to/pretrained/weights/if/present.pt
To train on your custom datasets, please refer to the Lightnet documentation.
To draw boxes on a test image, use the Training/Lightnet/examples/basic/detect.py.
For evaluating the mAP score on the test dataset,
python3 bin/test.py -c -n cfg/tinyyolo.py /path/to/trained/weights.pt -f
We have included weights having 51.5% mAP score for the PASCAL VOC dataset, at Training/Weights-Packing/pretrained.pt.
The .pt to .npz converter and Finnthesizer are also present in the same directory. Call gen-weights.py from there to create the .bin files.
Such files for the pretrained weights are provided at BNN-PYNQ/bnn/params/tinyyolo/cnvW1A1.
We have modified the original BNN-PYNQ to support DoReFa-Net trained networks instead of their vanilla BNN WnAn like networks.
The library may be hard-coded for W1A4.
The hardware description is found in ./BNN-PYNQ/bnn/src/network/cnvW1A1.
Copy the bitstream and the parameters to the board. A shared library object for the network must also be compiled as per the new configuration for correct weight loading.
Run the notebook 'Inference/TinyYOLO_LOOP.ipynb'
Please note that we use the shared object and its python interface to only load the accelerator weights. Rest of the inference steps are done using Python pynq package.
Moreover, we do not use a custom-written post-processor for YOLO, not the one in Darknet.
As linked above.