search

PKU-ICST-MIPL / FinePOSE_CVPR2024

FinePOSE: Fine-Grained Prompt-Driven 3D Human Pose Estimation via Diffusion Models
MIT License
40 stars 8 forks source link

When training for more than 100 epochs the results obtained are worse than those of the first 100 epochs #2

Open JihuaPeng opened 2 months ago

JihuaPeng commented 2 months ago

This is a great work for 3D pose estimation. When I train, I find that the best result I save in the first 100 epochs is the 95th epoch, but after I train 200 epochs I get the 156th epoch as the best result. After I evaluate the 95th epoch and the 156th epoch, the result of the 156th epoch is worse than the 95th one, can the author explain the reason? QQ_1721055042012 QQ_1721055060878

rabbitchenrui commented 1 week ago

maybe model convergence around 100 epoches. Since when evaluate, it sampled from guassion noise, result will be different when different initlization. By the way, can you share me the batch_size and GPU you used when training? I use the default batch_size 1024, can not get the result like you showed in picture. @JihuaPeng