search

nickgkan / 3d_diffuser_actor

Code for the paper "3D Diffuser Actor: Policy Diffusion with 3D Scene Representations"
https://3d-diffuser-actor.github.io/
MIT License
227 stars 28 forks source link

Training on a single task #31

Closed pawanw17 closed 6 months ago

pawanw17 commented 6 months ago

Hi, Thanks for the great work!

I am trying to train 3d_diffuser_actor on a single task "turn tap", the graphs (attached below) show that the model is learning well and converging but when using the inference script. I am not seeing the desired results, can you please help identify what could be going wrong?

In the training script I changed "ngpus" to 1 and passed only "turn_tap" to "tasks"

image image

I am training on a single A4000 GPU

twke18 commented 6 months ago

Hi,

It is very likely to a problem of normalization. You can set this argument to 1, so that the location bound is selected to be the same as training. If you only specify one task in the training script, this function will select the bound of that task, instead of aggregating among all tasks.

pawanw17 commented 6 months ago

Hi,

That was it! changing the parameter single_task_gripper_loc_bounds to 1 worked. Thank you so much! I wish I had asked this sooner...