search

Xilinx / Vitis-AI-Tutorials

MIT License
389 stars 146 forks source link

quantize output nodes may have a problem #8

Open xieyi4650 opened 4 years ago

xieyi4650 commented 4 years ago

When I quantize yolo v3 model flow this tutorial, I found that the output nodes in this code block should be conv2d_59/BiasAdd,conv2d_67/BiasAdd,conv2d_75/BiasAdd while not conv2d_59/convolution,conv2d_67/convolution,conv2d_75/convolution.

!vai_q_tensorflow quantize \
  --input_frozen_graph model_data/yolov3_voc.pb \
  --input_nodes input_1 \
  --input_shapes ?,416,416,3 \
  --output_nodes conv2d_59/convolution,conv2d_67/convolution,conv2d_75/convolution \
  --input_fn input_fn.calib_input \
  --method 1 \
  --gpu 0 \
  --calib_iter 100
jimheaton commented 4 years ago

Yes, you are correct. The output node names should match what's output by the inspect command. I will fix this in the next release of the tutorial.