INT_MAX error

[L-Reichardt]

Hello,

I am encountering the following error, when using the model on larger pointclouds. I have reduced the model size and batch size, however I still get the following error.

index_0 = p2v_map.unsqueeze(-1).expand(-1, -1, k)[mask_mat] #[M, ]
RuntimeError: nonzero is not supported for tensors with more than INT_MAX elements,   file a support request

Do you know what causes this? I havent been able to find anything useful online.

[X-Lai]

It means that the number of query-key pairs is too large. You may consider setting a larger voxel_size.

[L-Reichardt]

Thank you for your answer. I thought maybe its caused so something else, because the error I get with larger grids is

 pointops_cuda.attention_step1_forward_cuda_v2(N_k, M, h, C, n_max, q, k, index0_offsets, index1, output)
RuntimeError: Caught an unknown exception!

For features I use an (N, 3) pointcloud, and also the same for Coordinates.

My model settings are :

        dataset_dict = {
            'stem_transformer': True,
            'use_xyz': True,
            'sync_bn': False,
            'rel_query': True,
            'rel_key': True,
            'rel_value': True,
            'quant_size': 1.0, # 0.01
            'downsample_scale': 8,                          # Scale for Stratified window
            'num_layers': 1, # 2
            'patch_size': 1,
            'window_size': 8,
            'depths': [1, 1, 1, 1], # [2, 2, 6, 2]
            'channels': [32, 32, 32, 64],  # [48, 96, 192, 384]
            'num_heads': [4, 8, 16, 32], # [3, 6, 12, 24]
            'up_k': 3,
            'drop_path_rate': 0.3,
            'concat_xyz': False,
            'grid_size': 4.0, # 0.04
            'max_batch_points': 200000,
            'max_num_neighbors': 34, # 34
            'ratio': 0.25,                                  # Downsample ratio
            'k': 16, # 16
            'sigma': 1.0
            }

Basically I reduced the model to be small because the pointclouds I work with are so large I rather just get it running in a "minimum" configuration, as an initial step.

[L-Reichardt]

Update : Error disappears with 'stem_transformer' = False, so I guess there is an issue there

[X-Lai]

The error RuntimeError: Caught an unknown exception! is caused because you set the channel number per head to 8. Here we only support 16 or 32.

[L-Reichardt]

@X-Lai Thank you, that solved the error. However now I get the n_max error. Do you have an overview how/what parameters to set? The parameters are a little in-transparent to me, so far I have not been able to use the model with external pointclouds or random tensors.

[X-Lai]

The parameters like voxel_size quant_size grid_size max_batch_points can be very important.

[L-Reichardt]

@X-Lai How exactly does quant_size relate to the voxels and grid sizes? What are the downsides of increasing the value?

I've been needing to use a quant size >= 0.1 to prevent the n_max error. Do I understand correctly that quant affects the samping inside the transformer windows?