Bug report: cannot load checkpoints for EgoTracks

[vineetparikh]

Is there any demo code to load the provided checkpoints to a STARK Tracker? I'd like to try the EgoSTARK setup on a different egocentric dataset to see how well it would perform.

[vineetparikh]

Followup from here, it looks like actually loading the provided checkpoints is bugged due to an import error

import torch
from argparse import Namespace
from tracking.utils.defaults import setup
import tracking.models.stark_tracker.stark_tracker as stark

stark_args = Namespace(model_type="STARK", config_file="egotracks_models/stark_st_base.yaml", opts=[])
stark_cfg = setup(stark_args)
tracker = stark.STARKTracker(stark_cfg)

works fine (albeit with some extra config parameters, I'm getting stuff like the checkpoint contains additional key for a lot of layer4 keys along with the two fc keys) using the provided base config, but as soon as I try to do

with open("/path/to/STARKST_ep0001.pth.tar", "rb") as f:
    tracker.model.load_state_dict(torch.load(f)["state_dict"], strict=True)

I get the error

... 
Traceback (most recent call last):
  File "<stdin>", line 2, in <module>
  File "/home/vap43/.conda/envs/hos/lib/python3.9/site-packages/torch/serialization.py", line 607, in load
    return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args)
  File "/home/vap43/.conda/envs/hos/lib/python3.9/site-packages/torch/serialization.py", line 882, in _load
    result = unpickler.load()
  File "/home/vap43/.conda/envs/hos/lib/python3.9/site-packages/torch/serialization.py", line 875, in find_class
    return super().find_class(mod_name, name)
  File "/home/vap43/.conda/envs/hos/lib/python3.9/site-packages/egotracks-1.0-py3.9.egg/tracking/tools/trainers/base_trainer.py", line 16, in <module>
    from tracking.dataset.build import build_dataloaders
  File "/home/vap43/.conda/envs/hos/lib/python3.9/site-packages/egotracks-1.0-py3.9.egg/tracking/dataset/build.py", line 12, in <module>
    from tracking.dataset.train_datasets.ego4d_vq import Ego4DVQ
  File "/home/vap43/.conda/envs/hos/lib/python3.9/site-packages/egotracks-1.0-py3.9.egg/tracking/dataset/train_datasets/ego4d_vq.py", line 8, in <module>
    from tracking.dataset.ego4d_tracking import Ego4DTracking
ModuleNotFoundError: No module named 'tracking.dataset.ego4d_tracking'

which seems to indicate that there's another ego4d_tracking data file that isn't public, and/or that one file was renamed incorrectly. Any insight on what the right file should be here?

[shengyuhao]

I also have this error and hope the authors provide more details or explanations. Thanks!

[relh]

Looks like the only other file that has a similar function signature and also has a response_tracks variable is this one: https://github.com/EGO4D/episodic-memory/blob/5275b9570a3d77884e8685bad474309f7778db1f/EgoTracks/tracking/dataset/eval_datasets/ego4d_lt_tracking_dataset.py#L45

[SharkyK]

I also have this problem. Could you please tell me how to do if you solve this error? Thanks!