search

EricGuo5513 / momask-codes

Official implementation of "MoMask: Generative Masked Modeling of 3D Human Motions (CVPR2024)"
https://ericguo5513.github.io/momask/
MIT License
689 stars 56 forks source link

How to train other datasets #9

Open aooliver opened 5 months ago

aooliver commented 5 months ago

Hi, thank you for your excellent work. I would like to know how to train on other datasets, such as the Motion-X dataset, and what kind of preprocessing operations are required. Thank you for your answer.

Murrol commented 5 months ago

Hi, thanks for your interest.

In GET YOUR READ - 3.Get Data, there are instructions on how to process data. Motion-X also includes the AMASS dataset used in HumanML3D. If you only want to use Motion-X for main-body motion (excluding hands and face), it would be almost the same, except for the difference between SMPL-H and SMPL-X.

For full-body motion, you need to modify the processing scripts and the dataloader.

LinghaoChan commented 5 months ago

Hi, thank you for your excellent work. I would like to know how to train on other datasets, such as the Motion-X dataset, and what kind of preprocessing operations are required. Thank you for your answer.

@aooliver Motion-X supports the tomato motion format for model training. Please refer to this.

hdjsjyl commented 5 months ago

Hi @LinghaoChan , Are you successful of training momask on motion-x data? I meet the issue "loss nan". Any suggestion from you? Thanks

LinghaoChan commented 5 months ago

Hi @LinghaoChan , Are you successful of training momask on motion-x data? I meet the issue "loss nan". Any suggestion from you? Thanks

@hdjsjyl

Not yet. However, we have trained HumanTOMATO, MLD, T2M-GPT, MotionDiffuse, and MDM on Motion-X yet. And the TMR was also trained on Motion-X. We plan to release the TMR training/inference on Motion-X soon. You can try the demo here. To make the code easy to use, we are working hard to comment out every potentially questionable code block and encapsulate it as PyPI. Please be patient. Regarding your question, I am not clear about your detailed background now. Sorry about this.

Ling-Hao CHEN

hdjsjyl commented 5 months ago

Hi @LinghaoChan , really appreciate your detailed reply. Would you share your training list, val list and testing list? In total, I get 81k for whole-body generation, is it correct? Thank you

LinghaoChan commented 5 months ago

Hi @LinghaoChan , really appreciate your detailed reply. Would you share your training list, val list and testing list? In total, I get 81k for whole-body generation, is it correct? Thank you

I remember the training list length is about 64k. Seems good.

hdjsjyl commented 5 months ago

hi @LinghaoChan , thanks for your reply. If I want to replace HumanML3D with Motion-X to train the models, such as T2M, MLD. Do you have some suggestions about the training? Should I keep the same training parameters such as, kinematic chain, learning rate, number of epochs, batch size? Any other concerns I should consider? Thank you

LinghaoChan commented 5 months ago

hi @LinghaoChan , thanks for your reply. If I want to replace HumanML3D with Motion-X to train the models, such as T2M, MLD. Do you have some suggestions about the training? Should I keep the same training parameters such as, kinematic chain, learning rate, number of epochs, batch size? Any other concerns I should consider? Thank you

Seems you are right. The kinematic chain should follow the SMPL-H/-X with tomato format.

hdjsjyl commented 5 months ago

hi @LinghaoChan , thank you.