Mismatched weights running --test

[bendoesai]

Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "C:\Python311\Lib\site-packages\tmrl\__main__.py", line 88, in <module>
    main(arguments)
  File "C:\Python311\Lib\site-packages\tmrl\__main__.py", line 26, in main
    rw = RolloutWorker(env_cls=partial(GenericGymEnv, id=cfg.RTGYM_VERSION, gym_kwargs={"config": config}),
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Python311\Lib\site-packages\tmrl\networking.py", line 505, in __init__
    self.actor = self.actor.load(self.model_path, device=self.device)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Python311\Lib\site-packages\tmrl\actor.py", line 137, in load
    self.load_state_dict(torch.load(path, map_location=self.device))
  File "C:\Python311\Lib\site-packages\torch\nn\modules\module.py", line 2041, in load_state_dict
    raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for SquashedGaussianMLPActor:
        Missing key(s) in state_dict: "net.0.weight", "net.0.bias", "net.2.weight", "net.2.bias".
        Unexpected key(s) in state_dict: "net.conv1.weight", "net.conv1.bias", "net.conv2.weight", "net.conv2.bias", "net.conv3.weight", "net.conv3.bias", "net.conv4.weight", "net.conv4.bias", "net.mlp.0.weight", "net.mlp.0.bias", "net.mlp.2.weight", "net.mlp.2.bias".

Seems that the provided weights and the weights expected are a little different. Is this an issue that can be fixed in config or do I just need a new set of weights?

[yannbouteiller]

This is strange, none of the two neural networks embedded in tmrl have changed for years as far as I know.

Perhaps something changed in how Pytorch loads weights recently?

[bendoesai]

Figured it out. my RUN_NAME and ENV settings were mismatched. I should have guessed that when unexpected keys were for conv layers. Until now I thought RUN_NAME was arbitrary and could be set to whatever.