search

DefTruth / torchlm

💎A high level pipeline for face landmarks detection, it supports training, evaluating, exporting, inference(Python/C++) and 100+ data augmentations, can easily install via pip.
https://github.com/DefTruth/torchlm
MIT License
244 stars 25 forks source link

Missing mobilenet_v2 model file #53

Closed gordinmitya closed 2 years ago

gordinmitya commented 2 years ago

It's mentioned in readme but page with releases contain only pipnet with resnet18 and resnet101 backbone. @DefTruth would be great if you can share mobilenet_v2 backboned pipnet! Because even resnet18 is too heavy for mobile application (44mb). Thank you for great work, very useful!

John1231983 commented 2 years ago

Me too, I also want to know performance of mob with augmentation in the repo. It seems that the performance in table of README copy from paper report

DefTruth commented 2 years ago

Me too, I also want to know performance of mob with augmentation in the repo. It seems that the performance in table of README copy from paper report

hi~, torchlm is not aims to build it's own SOTA algorithm but reproduce some SOTA models report by papers and other open source codes, so, i have re-implentment PIPNet carefully and make sure the users can use the SOTA's pretrained weights directly. I like it's Heatmap+Regression designs. The performance of PIPNet is source from their paper. It's very easy to fine tune PIPNet with the high-level APIs in torchlm, some transforms, such as MixUp, may boost the performance. Also, the inference codes with Pytorch/ONNXRuntime(Python), ONNXRuntime/MNN/NCNN/TNN(C++) have already release at torchlm. For the commercial reason, the models trained with my custom datasets will not be release.

DefTruth commented 2 years ago

It's mentioned in readme but page with releases contain only pipnet with resnet18 and resnet101 backbone. @DefTruth would be great if you can share mobilenet_v2 backboned pipnet! Because even resnet18 is too heavy for mobile application (44mb). Thank you for great work, very useful!

i have re-implentment PIPNet carefully and make sure the users can use the SOTA's pretrained weights directly. They did not release the pretrained weight with mobilenet_v2 backbone, but it's very easy to train your own PINet with mobilenet_v2 backbone through the high level APIs in torchlm. For the commercial reason, the models trained with my custom datasets will not be release.

John1231983 commented 2 years ago

Have you train model from scratch and what is performance from scratch?

DefTruth commented 2 years ago

Have you train model from scratch and what is performance from scratch?

It's seems can not get good performance with dense landmarks (>=300) but suitable for sparse landmarks (<=100)

DefTruth commented 2 years ago

C++ re-implement of some realtime face landmarks models can be found at my another project lite.ai.toolkit

wwdok commented 2 years ago

If someone trained pipnet with mobilent backbone, please share it, i don't mind its accurrcy