search

facebookresearch / habitat-lab

A modular high-level library to train embodied AI agents across a variety of tasks and environments.
https://aihabitat.org/
MIT License
2.02k stars 495 forks source link

Issue that does not change the score of "SUCCESS", "SPL" in objectnav task #1563

Open etri-jeon0123 opened 1 year ago

etri-jeon0123 commented 1 year ago

Habitat-Lab: v0.2.5 Habitat-Sim: v0.2.5

We are training to get the desired scor[success, spl] by executing code ddppo_objectnav_hm3d.yaml.

train command: ddppo_objectnav_hm3d.yaml python -u -m habitat_baselines.run \ --config-name=objectnav/ddppo_objectnav.yaml \ habitat_baselines.trainer_name=ver

The composition of the data is, ./data/datasets/objectnav/hm3d/v2 ./datat/scene_datasets/hm3d_v0.2/train

However, there is no change in the score at all, as shown in the figure below. image

When the captured image is checked, the scores do not change even if the training step increases.

What is wrong with training?

However, when performing code ppo_pointnav_example.yaml, it was confirmed that it was properly trained.

train command: ppo_pointnav_example.yaml python -u -m habitat_baselines.run \ --config-name=pointnav/ppo_pointnav_example.yaml \ habitat_baselines.trainer_name=ver

Jay-Vim-Lv commented 1 month ago

i met the same problem, objectnav task on hm3d training by ddppo_objectnav_hm3d.yaml get results thart spl and suceesss always remain 0.000. i have trained more than 50000 updates. have you found the reason? expecting your reply!