about pre-trained models

[matlab2017]

about pre-trained models

[linjing7]

Hi, the encoders are from ViTPose, which are pretrained on the COCO, AIC, MPII, and CrowdPose datasets.

[matlab2017]

The model is downloaded directly in this project (https://github.com/ViTAE-Transformer/ViTPose)? Did you do any extra training?
thanks very much

[linjing7]

Hi, I have tried to finetune the models on the COCO-WholeBody datasets, but it does not work. Thus, I use the original pretrained model from ViTPose.

[matlab2017]

How many A100s did you used? How long will it take? I want to try training OSX_b, but one 4090 can be used.

[linjing7]

Hi, we train OSX_l on one A100-80G for 2 days. If you do not have enough GPU, you can run a lighter model by adding --decoder_setting wo_decoder or --decoder_setting wo_face_decoder.