search

deepinsight / insightface

State-of-the-art 2D and 3D Face Analysis Project
https://insightface.ai
23.02k stars 5.37k forks source link

Subcenter ArcFace drop.py #2087

Open abhatta1234 opened 2 years ago

abhatta1234 commented 2 years ago

@nttstar I checked the pre-trained model that was made available and there is no fc7_*_weight layers - referred in Line 100 of drop.py (Using r100's pretrained models)I think these layers comes from parallal_local_module_v1? Is there any new push to be done to this repo? Trying to run drop.py -- any suggestions?

nttstar commented 2 years ago

Hi, you have to train from scratch using ArcFace and also save the 'fc7_*' weights which is required by drop functions.

abhatta1234 commented 2 years ago

@nttstar Thank you for your suggestion! I did the training part but couldn't save "fc7*" weights. Any suggestions? I guess I am confused about saving "fc7*" for label_name in model/while calling parallel module. I can't find any documentation to be able to save fc7 weights. Thank you for your help!!

nttstar commented 2 years ago

set config.ckpt_embedding = False may work.

abhatta1234 commented 2 years ago

@nttstar Did that too, but it seems to be not working. When I get parallel module returned, it seems like it's not returning fc7_ layers. Seems like something is missing with either self.params_initialized or set_params(). Seems like setparams() adds fc7 to parallel modules. Any recommendations??

PR451 commented 1 year ago

@abhatta1234 I am facing same issue, were you able to solve this?