search

mks0601 / Hand4Whole_RELEASE

Official PyTorch implementation of "Accurate 3D Hand Pose Estimation for Whole-Body 3D Human Mesh Estimation", CVPRW 2022 (Oral.)
MIT License
314 stars 31 forks source link

Questions about training. #13

Open mls1999725 opened 2 years ago

mls1999725 commented 2 years ago

Hi, I'm very interested in your excellect work! Could you please tell me the number of epochs in each of the three stages (pre-train Hand4Whole, pre-train hand-only Pose2Pose and combine pre-trained Hand4Whole and hand-only Pose2Pose and fine-tune it), and could it be that there are only 7 epochs in the first stage and only 12 epochs in the second stage as shown in the codes?

By the way, according to the paper, is the learning rate multiplied by 0.1 at the 10th epoch for all three stages?

Looking forward to your reply!

mks0601 commented 2 years ago
  1. pre-train Hand4Whole: set lr_dec_epoch in config.py to [4,6] and end_epoch in config.py to 7. The network should be trained after initializing it with 2D pose estimation network as described in README.
  2. pre-train hand-only Pose2Pose: set lr_dec_epoch in config,py to [10,12] and end_epoch in config.py to 13.
  3. combine pre-trained ones and fine-tune it: set lr_dec_epoch in config.py to [4,6] and end_epoch in config.py to 7. The learning rate should be 1e-5 as described in README.
mls1999725 commented 2 years ago

Thanks a lot !