search

facebookresearch / nle

The NetHack Learning Environment
Other
939 stars 113 forks source link

done is False in the dataset, even though the agent is dead #355

Open Howuhh opened 1 year ago

Howuhh commented 1 year ago

🐛 Bug

There are many trajectories in the NLD-AA dataset which, according to the metadata, ended in the agent's death. However, the done for them is False. I calculated the difference for the taster dataset, there are 1933 games, while the trajectories in which done is True only 1421.

To Reproduce

Steps to reproduce the behavior:

Load taster dataset, get full trajectory for id 1, then in metadata for 'death' key will be 'killed by a hobbit while frozen by a monster's gaze'. However, done will be false for this trajectory.

Environment

NLE version: 0.9.0 PyTorch version: 2.0.0 Is debug build: No CUDA used to build PyTorch: None

OS: Mac OSX 13.2.1 GCC version: Could not collect CMake version: version 3.25.2

Python version: 3.9 Is CUDA available: No CUDA runtime version: No CUDA GPU models and configuration: No CUDA Nvidia driver version: No CUDA cuDNN version: No CUDA

Versions of relevant libraries: [pip3] functorch==2.0.0 [pip3] haste-pytorch-cpu==0.5.5 [pip3] numpy==1.23.4 [pip3] torch==2.0.0 [pip3] torch-ac==1.4.0 [pip3] torchvision==0.14.1 [conda] functorch 2.0.0 pypi_0 pypi [conda] haste-pytorch-cpu 0.5.5 pypi_0 pypi [conda] torch 2.0.0 pypi_0 pypi [conda] torch-ac 1.4.0 pypi_0 pypi [conda] torchvision 0.14.1 pypi_0 pypi

BartekCupial commented 7 months ago

Usually in gym you get done at the first episode of the next trajectory. Maybe this is the reason you don't see them at the last step here.

Howuhh commented 7 months ago

@BartekCupial then way done is True on last step on other trajectories? 😃