allenai / allenai/allenact

Cannot reproduce the performance of EmbCLIP on the ObjectNav task of the RoboTHOR

Open
#377 2 comments 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
383
Forks
59
PR merge metrics
No merged PRs in 30d

Description

## Problem / Question

Official EmbCLIP is on the leaderboard with a TestSPL score of 0.2.
But according to the [tutorial](https://github.com/allenai/embodied-clip/blob/main/readme_files/baselines_robothor_objectnav.md), I got a score of 0.17.
The command I ran was:
```
python allenact/main.py -o storage/objectnav-robothor-rgb-clip-rn50 -b projects/objectnav_baselines/experiments/robothor/clip objectnav_robothor_rgb_clipresnet50gru_ddppo -c pretrained_model_ckpts/objectnav-robothor-clip-rn50.130M.pt
```
And I got the file `storage/objectnav-robothor-rgb-clip-rn50/metrics/ObjectNav-RoboTHOR-RGB-ClipResNet50GRU-DDPPO/2024-02-15_03-32-04/metrics__test_2024-02-15_03-32-04.json`
According to the method in the [documentation](https://allenact.org/projects/objectnav_baselines/), I got the metric of the validation.
According to the scripts form [robothor-challenge](https://github.com/allenai/robothor-challenge/blob/main/robothor_challenge/scripts/convert_allenact_metrics.py), I run the command:
```
python convert_allenact_metrics.py -v storage/objectnav-robothor-rgb-clip-rn50/metrics/ObjectNav-RoboTHOR-RGB-ClipResNet50GRU-DDPPO/2024-02-15_02-36-50/metrics__val_2024-02-15_02-36-50.json -t storage/objectnav-robothor-rgb-clip-rn50/metrics/ObjectNav-RoboTHOR-RGB-ClipResNet50GRU-DDPPO/2024-02-15_03-32-04/metrics__test_2024-02-15_03-32-04.json -o submission_metrics.json.gz
```
Then submit the `submission_metrics.json.gz` file,

![image](https://github.com/allenai/allenact/assets/56984759/f7619f49-2a32-42be-a2db-cf1fa21243f5)

What should I do?

## Additional context

_(Optional)_ - To provide support it's helpful to have as many details as possible, add additional context here.

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.