search

Xilinx / Vitis-AI

Vitis AI is Xilinx’s development stack for AI inference on Xilinx hardware platforms, including both edge devices and Alveo cards.
https://www.xilinx.com/ai
Apache License 2.0
1.49k stars 630 forks source link

YOLOX QAT file size and xmodel file size #1091

Closed aiqiangbrave closed 1 year ago

aiqiangbrave commented 1 year ago

Could you help me analyze the size of the model? I remember that the xmodel model after yolov4 QAT is about 7M, but the model after yolox QAT is only twice shrunk (about 100M), is this normal? Why is the quantified model still so large？ size_problem image

Thank you so much~~~

now,I ignored the previous question for a moment. I used the run_train.sh trained float model to generate the xmodel file, and although it was successful, I compiled it with the following error: Instruction format: vai_c_xir -x /PATH/TO/quantized.xmodel -a /PATH/TO/arch.json -o /OUTPUTPATH -n netname error: [UNILOG][FATAL][XCOM_PM_FAIL][The compiler occurs an error when generating instructions, please contact us.]

huisunCompiler commented 1 year ago

Hi @aiqiangbrave, Here is the answer of the first question. The const data stored in the provided model "YOLOX_0_int.xmodel" are still in floating-point format. The data will be converted to XINT8, and the output size will be approximately 25% of the original size after compiling. I have tried compiling the model using zcu104, the output xmodel size is 27MB.

For the second question, please check the model architecture to see if there are any modifications comparing with the official version. If your quantized.xmodel is different from the official version, there may exist some unencountered issues. I have used the official version and compiling it on zcu102, zcu104, and kv260 and doesn't reproduce your problem.

Could you please provide your quantized.xmodel, your architecture information and your vitis-ai version, so that we can reproduce your scenario.

qianglin-xlnx commented 1 year ago

Closing since no activity for more than 1 month, please reopen if you still have any questions, thanks