May I ask how to fix this error? (When using my own trained model to infer) about lidar-mos HOT 4 CLOSED

Please provide more information such as how many cards are used for training.

This is a common general error, because adding module.* to the key prefix.

You could try like this:
https://github.com/haomo-ai/MotionSeg3D/blob/dc4c95fcdba2f0819d2bbc4a419f231c55e9c6f3/modules/user.py#L63-L69

self.model = SalsaNextWithMotionAttention(self.parser.get_n_classes(), ARCH, num_batch=self.infer_batch_size)
self.model = nn.DataParallel(self.model)
checkpoint = "SalsaNextWithMotionAttention_refine_module_valid_best"
w_dict = torch.load(f"{self.modeldir}/{checkpoint}", map_location=lambda storage, loc: storage)
# self.model.load_state_dict(w_dict['main_state_dict'], strict=True)
self.model.load_state_dict({f"module.{k}":v for k,v in w_dict['main_state_dict'].items()}, strict=True)

from lidar-mos.

Thanks a lot my friend. It seems like we just need to make a little change to the original code in user.py:
line 64: self.model.load_state_dict(w_dict['state_dict'], strict=Ture) --> self.model.load_state_dict(w_dict['state_dict'], strict=False)
line 70: self.model.load_state_dict(w_dict['state_dict'], strict=Ture) --> self.model.load_state_dict(w_dict['state_dict'], strict=False)
I successfully solved this problem by making changes as shown above, and I hope it may help some other friends who also meet this problem.

from lidar-mos.

Hi, @silence-tang, I am very worried that if you use it like this, it will not load successfully.
The logic of strict=False is to load values with equal keys. Although there is no error, it may not load any weight from the checkpoint.
As shown above, I think use this command is better, because strict=False is used directly, other key mismatches cannot be avoided, It is very likely that it did not load any weights from the checkpoint.

# self.model.load_state_dict(w_dict['main_state_dict'], strict=True)
self.model.load_state_dict({f"module.{k}":v for k,v in w_dict['main_state_dict'].items()}, strict=True)

from lidar-mos.

Thank you very much! @MaxChanger I just visualized the .label file generated by my method, but just as you said, it didn't load weight from the checkpoint and every point was labeled 9, which isn't what we expected. After applying the method you proposed, it finally worked in a proper way, thanks!

from lidar-mos.