open-mmlab / mmrazor

OpenMMLab Model Compression Toolbox and Benchmark.
https://mmrazor.readthedocs.io/en/latest/
Apache License 2.0
1.48k stars 231 forks source link

When will the quantization algorithm be released, and is it expected to be applicable to the model of transformer structure? #300

Open lb-hit opened 2 years ago

lb-hit commented 2 years ago

Checklist

Describe the question you meet

[here]

Post related information

  1. The output of pip list | grep "mmcv\|mmrazor\|^torch" [here]
  2. Your config file if you modified it or created a new one.
[here]
  1. Your train log file if you meet the problem during training. [here]
  2. Other code you modified in the mmrazor folder. [here]
humu789 commented 2 years ago

Sorry for the late reply, we will release the quantization part at the end of the year if all goes well. The transformer structure will not be supported in the first version, but it will be considered in subsequent releases.

JinqingZhengTju commented 2 years ago

Sorry for the late reply, we will release the quantization part at the end of the year if all goes well. The transformer structure will not be supported in the first version, but it will be considered in subsequent releases.

Is the quantization part in mmdeploy same as in this part in mmrazor?

humu789 commented 2 years ago

Sorry for the late reply, we will release the quantization part at the end of the year if all goes well. The transformer structure will not be supported in the first version, but it will be considered in subsequent releases.

Is the quantization part in mmdeploy same as in this part in mmrazor?

They are different, mmrazor is to implement some quantization algorithms, including QAT and PTQ, such as LSQ, Adaround and so on. mmdeploy is to provide baisc PTQ quantization functions for deploying better.

JinqingZhengTju commented 2 years ago

Sorry for the late reply, we will release the quantization part at the end of the year if all goes well. The transformer structure will not be supported in the first version, but it will be considered in subsequent releases.

Is the quantization part in mmdeploy same as in this part in mmrazor?

They are different, mmrazor is to implement some quantization algorithms, including QAT and PTQ, such as LSQ, Adaround and so on. mmdeploy is to provide baisc PTQ quantization functions for deploying better.

Thanks for your reply. I look forward to the new version mmrazor.

jeannotes commented 1 year ago

@humu789 hi, when will you release quant part which support transformer(like vit, swin-transformer etc) quantization, including ptq and qat? thanks