[QNN] Support for tflite quantized model with per-channel scales in weights.

apache / tvm

Open deep learning compiler stack for cpu, gpu and specialized accelerators

https://tvm.apache.org/

Apache License 2.0

11.67k stars 3.45k forks source link

[QNN] Support for tflite quantized model with per-channel scales in weights. #3999

Closed mingwayzhang closed 5 years ago

mingwayzhang commented 5 years ago

Hi TVM developers,

I noticed that recent TFLite quantized models generated by post-training quantization have multiple scales per weight tensor. I am not entirely sure whether qnn operators support per-channel based requantization or not. If not, do you guys have any plan on the roadmap?

Thanks. -Mingwei

tqchen commented 5 years ago

Thanks for bringing up the discuss thread, please open a new thread on https://discuss.tvm.ai/ as that is our main medium for technical discussions within the community.