Can not run pretrained weights of ObjectNav

[elicassion]

Hi,

I'm trying to run the pretrained weights from https://dl.fbaipublicfiles.com/habitat/data/baselines/v1/objectnav_v2_baseline_habitat_navigation_challenge_2023.pth but getting:

RuntimeError: Error(s) in loading state_dict for PPO:
        Missing key(s) in state_dict: "actor_critic.net.prev_action_embedding.bias", "actor_critic.action_distribution.mu_maybe_std.weight", "actor_critic.action_distribution.mu_maybe_std.bias".
        Unexpected key(s) in state_dict: "actor_critic.action_distribution.linear.weight", "actor_critic.action_distribution.linear.bias".
        size mismatch for actor_critic.net.prev_action_embedding.weight: copying a param with shape torch.Size([7, 32]) from checkpoint, the shape in current model is torch.Size([32, 6]).

The running script I am using is

python -u -m habitat_baselines.run \
    --config-name="ddppo_objectnav_v2_hm3d_stretch.yaml" \
    habitat_baselines.evaluate=True \
    habitat_baselines.eval_ckpt_path_dir="data/pretrained_checkpoints" \
    habitat.dataset.split=val \

which is specified at starter-code

[ykarmesh]

We have provided you with the discrete action agent but it looks like you aren't passing config params required to use the agent to the model. Please take a look here. These are the set of configs that we are passing to get the discrete action agent working:

    bash submission.sh \
    --model-path demo.ckpt.pth \
    --input-type rgbd \
    --task objectnav \
    --action_space discrete_waypoint_controller \
    --task-config configs/ddppo_objectnav_v2_hm3d_stretch.yaml \
    habitat_baselines.rl.policy.action_distribution_type=categorical \
    habitat_baselines.load_resume_state_config=True" \

[elicassion]

Thanks! Yes it was the action space problem and it works now. Sorry for the late reply.

[etri-jeon0123]

Thanks! Yes it was the action space problem and it works now. Sorry for the late reply.

How did you fix the error caused by the lack of "action_space_config"?