search

robotgradient / grasp_diffusion

Pytorch implementation of diffusion models on Lie Groups for 6D grasp pose generation https://sites.google.com/view/se3dif/home
MIT License
234 stars 24 forks source link

How to train the model to achieve consistent results? #10

Closed caijunhao closed 10 months ago

caijunhao commented 10 months ago

Hi @TheCamusean,

Thanks for your great work!
 I'm trying to train a point cloud-conditioned model from scratch based on the parameters provided in the code. However, the evaluation results seem to have a large difference from those evaluated with the checkpoints you provided (44% vs. 92% of grasp success rates). Is there any hyperparameter I need to modify? If not, what should I do to achieve consistent results? Thank you!

robotgradient commented 10 months ago

Training on Pointclouds usually takes a long time.

Even if you observe the learning rate of the diffusion getting constant, it is good to keep training the model.

For how long did you train your model?

Best regards, -Julen

caijunhao commented 10 months ago

Thank you for your reply. I tried to increase the batch size to 16 and trained the model for about a day, and now it can achieve the same results.