CNN AI Accelerator using DARKNET-19 YOLO v3 Tiny

Our GOAL is to design HW AI accelerator that works low power, high performance.

Algorithm: DARKNET-19 AI Model: YOLO v3 Tiny Boards: PYNQ-Z2 FPGA Borad Main tool: Xilinx Vivado, Vitis AI

Developing

• Quantization DARKNET-19 parameters to INT8 type.
• Coding reference code in python. (all parameters are called in txt files.)
• Design Conv layer.
• Design Pooling layer.

Substep

Quantization DARKNET-19 parameters to INT8 type.

• My team discuss about Bacth Normalization.
  • For Implementation, We guess Bacth Norm is not essential & causes large HW resources.
  • That discussion would be great paper subjects.

Coding reference code in python. (all parameters are called in txt files.)

• That isn't too hard but we didn't get parameter txt files written by INT8 type.
• OpenCV, TensorFlow, Numpy, etc... will be used.
• https://github.com/ValentinFigue/TinyYOLOv3-PyTorch : developed the model from their work.
• https://github.com/openvinotoolkit/open_model_zoo : converted weight file from their work.

Design Conv layer.

• The components of Conv layer is "Convolution", "BatchNorm", and "Leaky ReLU".
• Convolution layer
  • almost done.
  • top module and controller are remain.
• BatchNorm
  • For HW resources problem, we determine delete BatchNorm layer.
  • all parts would be done and have problem, we will apply this layer.
• Leaky ReLU
  • ReLU vs Leaky ReLU
  • One of our issues.

Design Pooling layer.