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

Checkpoints size #5

Closed rpartsey closed 4 years ago

rpartsey commented 4 years ago

@srama2512 Could you, please, also give some insights on why the size of the checkpoints of the same model type vary a lot

For example, let's consider OccAnt(depth):

README.md Model name Training data Val SPL Checkpoint URL Config URL Checkpoint size
Pretrained models OccAnt(depth) Gibson 4+ 0.912 ckpt.11.pth ppo_exploration.yaml 125 MB
Habitat Challenge OccAnt(depth) Gibson 2+ 0.463 ckpt.13.pth ppo_exploration.yaml 94.6 MB
ECCV OccAnt(depth) Gibson 4+ 0.784 ckpt.6.pth ppo_navigation_evaluate.yaml 6.6 MB

Especially, I'm curious about the last row with the checkpoint size of 6.6 MB.

srama2512 commented 4 years ago

Haha, yes. The models have changed over time. The ECCV models were pretty lightweight with the unet_nsf value being 16 and the pose estimator being smaller. Since then, I have changed the pose estimator capacity and tuned unet_nsf further.

The OccAnt(depth) model in the 1st row is large since unet_nsf is 64. If unet_nsf is 16 instead, the model would be just 30MB. For the habitat challenge model, while unet_nsf is 16, the pose estimator capacity is higher. I hope that explains it. I remember trying a larger capacity UNet for the challenge. It didn't give too many benefits and it slowed down the evaluation. So I just stuck to unet_nsf = 16.

rpartsey commented 4 years ago

@srama2512 Thanks for the explanation :)

Closing the issue as resolved.