cannot use GST_predictor_non_rand

[hungwlor]

when i change the --model-dir and --test_model arguments, it shows this bug:
Traceback (most recent call last):
File "/home/sora/colcon_ws/install/crowdnav_base/lib/crowdnav_base/test", line 33, in
sys.exit(load_entry_point('crowdnav-base', 'console_scripts', 'test')())
File "/home/sora/colcon_ws/build/crowdnav_base/crowdnav_base/test.py", line 173, in main
actor_critic.load_state_dict(torch.load(load_path, map_location=device))
File "/home/sora/.local/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1604, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for Policy:
size mismatch for base.spatial_attn.embedding_layer.0.weight: copying a param with shape torch.Size([128, 12]) from checkpoint, the shape in current model is torch.Size([128, 2]).
What should i do to resolve it?

[arrakakki]

you can just change the evaluation environment configuration in trained_models/GST_predictor_rand/arguments.py, set the
environment as below:

    # No prediction: for orca, sf, old_srnn, selfAttn_srnn_noPred ablation: 'CrowdSimVarNum-v0',
    # for constant velocity Pred, ground truth Pred: 'CrowdSimPred-v0'
    # gst pred: 'CrowdSimPredRealGST-v0'
    parser.add_argument(
        '--env-name',
        default='CrowdSimPred-v0',
        help='name of the environment')

the reason why the origin configure doesn't work is that the model will initialize the policy network architecture via the environment you are using. For example, if env not in ['CrowdSimPred-v0', 'CrowdSimPredRealGST-v0'], the policy doesn‘t need to setup attention form st-graph built via 5-step trajectory prediction. you can verify this from the policy's attention module class in rl/networks/selfAttn_srnn_temp_node.py:

I set this navigation policy which was pretrained in none random human environment working right for me through this configuration change, hope this works for you too !