"Duplicate coordinates found" error when initializing "SparseTensor"

[jingyi-zhang]

Thanks for your good code！ but I got an error occasionally when use function 'SparseTensor' I dont know what caused this error ca

[jingyi-zhang]

This is a part code in 'getitem' function of DataSet

        coords = lidar_camera[:,0:2]
        #print(coords.shape[0])
        features = lidar_camera[:,2:4]
        labels = object_list
        coords /= self.quantization_size
        #print(coords.shape)
        mapping = ME.utils.sparse_quantize(coords=coords,
            return_index=True)
        #print(len(mapping))
        return coords[mapping], features[mapping], labels`

this my test code

train_set = SparseKittiDataset(classes=FLAGS.classes, split='training', augment=False, cfg=cfg)
train_loader = DataLoader(dataset=train_set, batch_size=FLAGS.batch_size, num_workers=FLAGS.workers,\
    shuffle=True, drop_last=True,collate_fn=ME.utils.batch_sparse_collate)
for batch_idx, data_label in enumerate(train_loader):
    coords, features, label = data_label
    #print(features.shape)
    #print(coords.shape)
    input = ME.SparseTensor(features, coords, force_creation=True)
    print(input)
    break

[chrischoy]

It happens when you have duplicate coordinates.

The standard way to fix this error would be sampling one point per voxel. This process is known as the voxel down-sampling. In addition, a sparse tensor requires a list of integers to represent a coordinate, so you need to floor a coordinate to a quantized indices.

The engine supports such down-sampling and quantization. Please refer to sparse_quantize.

In most cases, a training pipeline consists of

1. quantization https://github.com/StanfordVL/MinkowskiEngine/blob/415098d4e457a1373e0a0499414cee0cbfbd6144/examples/training.py#L139
2. collation and batched coordinate generation https://github.com/StanfordVL/MinkowskiEngine/blob/415098d4e457a1373e0a0499414cee0cbfbd6144/examples/training.py#L154
3. SparseTensor generation https://github.com/StanfordVL/MinkowskiEngine/blob/415098d4e457a1373e0a0499414cee0cbfbd6144/examples/training.py#L199

However, for classification like tasks where you can safely ignore duplicate coordinates, you can use force_creation=True. This should not be used for segmentation like tasks.

Please let me know if you have further questions.

[chrischoy]

In your example, you missed batched_coordinates which was in the training collation function.

This results in 2D sparse tensor. Coordinates must have size N x 4, not N x 3 to create a 3D sparse tensor.

[jingyi-zhang]

Thanks for your patience! You are so nice~

Took your advice, I used 'batched_coordinates' instead of 'batch_sparse_collate' to load data, then, the error has been solved.

But actually, I am still confused about the difference between 'batch_sparse_collate', 'SparseCollation' and 'batched_coordinates', I checked their definition in 'collation.py', seems like they have same function.