About the ycbcr and quantization
JuZiSYJ opened this issue · comments
Hi, it is a good job to incorporate the Re-parameterization into SR model. I still have some questions about the code and paper.
- In Section 2, Computation Reduction, the paper says
However, quantization of SR models often can hardly maintain the SR quality because of the precision requirement of pixel-wise prediction
The original RepVGG model does not perform well in quantization. How about the error of quantization in low-level Re-parameterization quantization? It seems that there is no further explanation or results for the quantization error in the paper.
Hi, it is a good job to incorporate the Re-parameterization into SR model. I still have some questions about the code and paper.
- In Section 2, Computation Reduction, the paper says
However, quantization of SR models often can hardly maintain the SR quality because of the precision requirement of pixel-wise prediction
The original RepVGG model does not perform well in quantization. How about the error of quantization in low-level Re-parameterization quantization? It seems that there is no further explanation or results for the quantization error in the paper.
Hi, thanks for your interest to this work.
(1) The quantization for SR task is a big topic, as it is related to the topology of network, selection of operations and quantization methods(PTQ series, QAT series...), etc. If your want to performs int8-quantization of Re-parameterization for engineering purpose, it is highly suggested that you fold the rep-networks back into a plain-net before performing PTQ. It could introduce affordable errors from our experience, more details can be refered to "Real-Time Quantized Image Super-Resolution on Mobile NPUs, Mobile AI 2021 Challenge: Report". One more thing to be added, our quantization-tools (PTQ) for this work is coming soon.
(2) Common SR models are tested on Y channel out of YCbCr, and we follow this manner, more infos can be refer to the papers from FSRCNN/ESPCN etc. There are two implementation in this repo, one is on the legacy folder which is developed and tested based on the EDSR-Pytorch, the other is on the root folder of this project which is a more lighten one. Both of them behave almost the same.
Thanks a lot!