search

jurgisp / pydreamer

PyTorch implementation of DreamerV2 model-based RL algorithm
MIT License
205 stars 47 forks source link

Minigrid environments not working #4

Closed stephmilani closed 2 years ago

stephmilani commented 2 years ago

Thanks for this Pytorch implementation of Dreamer-v2! I'm trying to get the code working with the Minigrid environments. However, I'm encountering the following error:

Traceback (most recent call last): File "/home/steph/projects/pydreamer/train.py", line 612, in run(conf) File "/home/steph/projects/pydreamer/train.py", line 227, in run model.training_step(obs, File "/home/steph/projects/pydreamer/pydreamer/models/dreamer.py", line 122, in training_step loss_probe, metrics_probe, tensors_probe = self.probe_model.training_step(features.detach(), obs) File "/home/steph/projects/pydreamer/pydreamer/models/probes.py", line 38, in training_step map_coord = insert_dim(obs['map_coord'], 2, I) KeyError: 'map_coord'

Here's the command I ran: xvfb-run -a -s "-screen 0 1400x900x24" python train.py --config defaults minigrid debug --env_id MiniGrid-Empty-8x8-v0 --device cuda

jurgisp commented 2 years ago

Thanks for feedback! It seems that my minigrid config is using some custom features of the MiniGrid environment, which is not in the default implementation (and not required for RL training).

In particular, it is caused by this line in config which specifies probe_model: map.

Please try commenting it out, or adding --probe_model none command line parameter.

stephmilani commented 2 years ago

It works now! Thanks so much for your help.