Pretrained checkpoint hardcodes data path, causing NotImplementedError when directory structure is different

[CiSong10]

Issue Description:

I'm trying to use the provided pretrained model (PointGroup-PAPER.pt checkpoint) to run prediction on my own point cloud dataset, but encountered a NotImplementedError during execution. This appears to be caused by a hardcoded dataroot embedded in the checkpoint config, which assumes the dataset is located at data_set1_5classes.

Since I reorganized the data directory to a different structure (e.g., ./data/data_set1_5classes/), the dataset loader failed to find the data and fell back to download(), which is not implemented — resulting in a crash.

Error Trace

[2025-06-11 14:35:17,201][torch_points3d.trainer][INFO] - DEVICE : cuda
[2025-06-11 14:35:18,132][torch_points3d.metrics.model_checkpoint][INFO] - Loading checkpoint from /home/cisong/ForAINet/outputs/pretrained/PointGroup-PAPER.pt
Traceback (most recent call last):
  File "PointCloudSegmentation/predict.py", line 13, in main
    trainer = Trainer(cfg)
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/trainer.py", line 48, in __init__
    self._initialize_trainer()
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/trainer.py", line 90, in _initialize_trainer
    self._dataset: BaseDataset = instantiate_dataset(self._checkpoint.data_config)
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/dataset_factory.py", line 46, in instantiate_dataset
    dataset = dataset_cls(dataset_config)
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/panoptic/treeins_set1.py", line 633, in __init__
    self.test_dataset = dataset_cls(
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/segmentation/treeins_set1.py", line 599, in __init__
    super().__init__(root, grid_size, *args, **kwargs)
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/segmentation/treeins_set1.py", line 239, in __init__
    super(TreeinsOriginalFused, self).__init__(root, transform, pre_transform, pre_filter)

...

  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/segmentation/treeins_set1.py", line 625, in download
    super().download()
  File "/usr/local/lib/python3.8/dist-packages/torch_geometric/data/in_memory_dataset.py", line 50, in download
    raise NotImplementedError
NotImplementedError

My Analysis

I tried to use VS Code to debug it, after loading the checkpoint, near line 90 of trainer.py, self._checkpoint.data_config.dataroot shows data_set1_classes. And a new ./data_set1_5classes/treeinsfused/raw directory was created. So I assume this fixed dataset configuration (data_config) prevents the model from being easily portable to other data directory structures.

This is also probably the cause of issue #1 (comment) , and why the workaround to change file path proposed in prs-eth/PanopticSegForLargeScalePointCloud#10 (comment) works.

Possible Improvements

If my analysis above is correct -- excuse me if not -- Ideally, checkpoints should store only model weights and architecture-related configs, not absolute or relative paths to training datasets.

Current Workaround

For now I put the /data_set1_5classes/treeinsfused/raw dataset folder back to where it is and solved this issue. I don't quite understand, if I am deploying the pretrained model weights to my own dataset, why does it needs to see the treeinsfused dataset. Or am I doing something wrong in configuration?

More issue

After the workaround, the NotImplementedError solved, another error aroused:

[2025-06-12 16:04:32,709][torch_points3d.trainer][INFO] - DEVICE : cuda
[2025-06-12 16:04:33,722][torch_points3d.metrics.model_checkpoint][INFO] - Loading checkpoint from /home/cisong/ForAINet/outputs/pretrained/PointGroup-PAPER.pt
Processing...
Traceback (most recent call last):
  File "/home/cisong/ForAINet/PointCloudSegmentation/predict.py", line 13, in main
    trainer = Trainer(cfg)
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/trainer.py", line 48, in __init__
    self._initialize_trainer()
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/trainer.py", line 90, in _initialize_trainer
    self._dataset: BaseDataset = instantiate_dataset(self._checkpoint.data_config)
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/dataset_factory.py", line 46, in instantiate_dataset
    dataset = dataset_cls(dataset_config)
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/panoptic/treeins_set1.py", line 633, in __init__
    self.test_dataset = dataset_cls(
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/segmentation/treeins_set1.py", line 599, in __init__
    super().__init__(root, grid_size, *args, **kwargs)
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/segmentation/treeins_set1.py", line 239, in __init__
    super(TreeinsOriginalFused, self).__init__(root, transform, pre_transform, pre_filter)
  File "/usr/local/lib/python3.8/dist-packages/torch_geometric/data/in_memory_dataset.py", line 60, in __init__
    super().__init__(root, transform, pre_transform, pre_filter)
  File "/usr/local/lib/python3.8/dist-packages/torch_geometric/data/dataset.py", line 86, in __init__
    self._process()
  File "/usr/local/lib/python3.8/dist-packages/torch_geometric/data/dataset.py", line 165, in _process
    self.process()
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/panoptic/treeins_set1.py", line 557, in process
    super().process()
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/segmentation/treeins_set1.py", line 622, in process
    super().process_test(self.test_area)
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/segmentation/treeins_set1.py", line 492, in process_test
    xyz, semantic_labels, instance_labels = read_treeins_format(
  File "/home/cisong/ForAINet/PointCloudSegmentation/torch_points3d/datasets/segmentation/treeins_set1.py", line 73, in read_treeins_format
    semantic_labels = data['semantic_seg'].astype(np.int64)-1
ValueError: no field of name semantic_seg

Can you help me with this? Why Python is looking for ['semantic_seg'] column from the data? My main python script is predict.py and the configuration file I use is predict.yaml

[CiSong10]

@Proeliorr I was finally able to run the pre-trained model on custom data by assigning dummy zero values to the ground truth labels, as suggested in #20 (comment).
You can refer to the note I wrote in my branch for more details.

In short, here’s what I did:

• Prepare data: Assign dummy values (e.g., zeros) for both semantic_seg and treeID.
• Make sure the folder data_set1_5classes/treeinsfused/raw is located at the top level of your working directory, even if you're only using your own dataset. Otherwise, you'll encounter the NotImplementedError mentioned in this issue.
• I made changes to PointCloudSegmentation/torch_points3d/metrics/panoptic_tracker_pointgroup_treeins_partseg.py to handle errors caused by the dummy (zero) ground truth values.

@bxiang233 Dear developer, could you please take a look at my approach?

If possible, it would be incredibly helpful to update the repo so that users can directly deploy the model on custom data (even without ground truth labels), without needing to apply these workarounds. Also, resolving the hardcoded dependency on the folder path (data_set1_5classes/treeinsfused/raw at the root) would make the repo much more flexible and accessible for the community.

Thank you very much!

[CiSong10]

I was able to pinpoint the issue.

In trainer.py (line 80), the checkpoint is loaded with:

self._checkpoint: ModelCheckpoint = ModelCheckpoint(...)

This checkpoint contains a self._checkpoint.data_config.dataroot (in my case, "data_set1_5classes").

Then, on line 90:

self._dataset: BaseDataset = instantiate_dataset(self._checkpoint.data_config)

the data_config is passed to instantiate_dataset().

Inside PointCloudSegmentation/torch_points3d/datasets/dataset_factory.py, line 45 instantiate_dataset function,

dataset_cls = get_dataset_class(dataset_config)

And within get_dataset_class(), line 15:

dataset_config.dataroot = hydra.utils.to_absolute_path(dataset_config.dataroot)

This call transforms "data_set1_5classes" into an absolute path, which, according to function docstring, "if the input path is relative, it's interpreted as relative to the original working directory", not any location controlled by the config file.

So unless the folder data_set1_5classes exists directly in the working directory (e.g., ./data_set1_5classes), this causes the dataset loader to fail, eventually triggering an unimplemented fallback in download(), which raises a NotImplementedError.

Future Improvements

Personally, I still need to understand why a dataroot ("data_set1_5classes") is needed in the code. In the future, I will consider editing PointCloudSegmentation/torch_points3d/datasets/dataset_factory.py, line 15, instead of calling hydra.utils.to_absolute_path, dataset_config.data_root can be assigned as some directory address that is related to the input data location.