Closed damor-rbz closed 1 year ago
If we do post training quantization we are able to get a model that can be compiled, but only if we first do some model passes with a dataset with near cero values, then the training with real data. This has been tested on version 2.0.
The description here is not very clear, could you please try it on Vitis-AI docker version 3.0 to see if the problem still exists?
This issue has been inactive for 4 months, so we will close it for now. If you have any further concerns or need assistance, please feel free to reopen it at any time.
Dear,
We have a yolov3 tiny model that can run on the DPU. Quantization is ok but when compiling we get the following error:
[UNILOG][FATAL][XCOM_UNSUPPORT_QUANTIZATION][The fix info is error or unsupported.] elew0's quantization info's variation is too large.
We are using Vitis-AI docker version 2.0 We are able to reproduce the error using version 2.5.
We tried quantization using random data and real images, in both case we got the same result. If we use a vector with all 0zeros as input data then the model compiles.
If we do post training quantization we are able to get a model that can be compiled, but only if we first do some model passes with a dataset with near cero values, then the training with real data. This has been tested on version 2.0.
Thanks,