search

TNTwise / rife-ncnn-vulkan

RIFE, Real-Time Intermediate Flow Estimation for Video Frame Interpolation implemented with ncnn library
MIT License
34 stars 4 forks source link

尝试对预训练好的模型进行int8量化失败 #9

Open kinier88 opened 2 months ago

kinier88 commented 2 months ago

$ ncnn2table flownet-opt.param flownet-opt.bin imagelist.txt flownet.table 'mean=[104.0,117.0,123.0]' 'norm=[1.0,1.0,1.0]' 'shape=[227,227,3]' pixel=BGR method=kl

layer rife.Warp not exists or registered zsh: segmentation fault ncnn2table flownet-opt.param flownet-opt.bin imagelist.txt flownet.table

TNTwise commented 2 months ago

rife.Warp is a custom layer specific to this repo, you can't convert the models to anything else.

rife.Warp 是此 repo 特有的自定义层，您不能将模型转换为其他任何东西。

kinier88 commented 2 months ago

请问是不是需要对将POW算子替换成rife.Warp前的模型进行int8量化后，再替换rife.Warp算子才行？建议仓库提供int8量化后模型

TNTwise commented 2 months ago

sorry, i dont know how to do that, as i am not the original author of this repo, i just convert new models

抱歉，我不知道该怎么做，因为我不是这个 repo 的原作者，我只是转换新模型