search

facebookresearch / OccupancyAnticipation

This repository contains code for our publication "Occupancy Anticipation for Efficient Exploration and Navigation" in ECCV 2020.
MIT License
78 stars 26 forks source link

Problem of hyper-parameter #19

Closed Abundant-me closed 3 years ago

Abundant-me commented 3 years ago

Thanks for your great work. I have a problem with the settting of hyper-parameters. In supplementary materials of the paper, part S8 gives some implementation of hyper-parameters. I am in doubt that the setting is for which kind of experiment? Is it corresponding to the occant_rgbd exploration on Gibson? But it is different with the default setting in ‘configs/model_configs/occant_rgbd/ppo_exploration.yaml’ in the code. Which one is the setting of the result of the paper. Specifically, I don't find 'Mapper update interval' parameter in the code. Is it same with the paper? Detials of difference is shown as follows

image

Is there any mistake in above table. If there is, please tell me.

By the way, the default value of parameter 'MAX_EPISODE_STEPS' is 1001 in gibson_train.yaml. Should it be 1000?

srama2512 commented 3 years ago

As we have stated in the README, this is a well-tuned version of the model we submitted to the habitat-20 challenge.

This code implements our winning entry to the Habitat-20 PointNav challenge which is a well-tuned version with improved ground-truth generation, faster training, and better heuristics for planning.

I would recommend using the provided configs as they generally tend to work well and train fast. We currently do not support training on the settings we used in the paper. However, we have provided the models + evaluation instructions in the eccv_2020_eval branch for replicability.