Open mbgniuxya opened 3 years ago
@mbgniuxya, thank you for reporting the issue. What is PyTorch version and installation you're using? Is that issue persistent. Can you try instructions from our challenge setup here for training.
Thanks for your reply. I think the issue is here: https://github.com/facebookresearch/habitat-lab/blob/7a1605cd9d5b21a5c23dd91f2b9eecc6e760d4c4/habitat_baselines/rl/ddppo/policy/resnet_policy.py#L332 it seems the input size is calculated using width width rather than height width from config file. I manually changed it to fit the actual input size and it works now.
I am having the same issue -- using resnet50, visual_encoder.output_shape
is currently set to (42, 7, 7)
(as it should be). But actually, the model produces features of size (42, 8, 10)
. So there must be some error in the ResNet implementation. Reverting to the "challenge-2021" or "v0.1.7" branches resolves this for me.
Hi all, I'm having trouble running habitat object navigation baseline on matterport 3d dataset on ubuntu 20.04. I have nvidia driver 465.19.01 with cuda version 11.3, rtx 3090 gpu. The command I used to run the baseline: python habitat_baselines/run.py --exp-config habitat_baselines/config/objectnav/ddppo_objectnav.yaml --run-type
train
The output I got: