Segment Fault due to resolution

[gitouni]

I use the unet to extract features (B, N, 32) from point clouds (B, N, 3). When I changed resolution from 50 to 24, the segment fault was thrown. This is not due to too low resolution, because it also threw segment fault in other data when I set resolution to 100. My Code snippet is shown below.

UNet

import torch.nn as nn
import sparseconvnet as scn
import torch
from typing import Iterable

class SCNet(nn.Module):
    def __init__(self, dimension=3, reps=1, m=32, nPlanes=[32, 64, 96, 128, 160], spatialSize=24*8, nClassesTotal=50):
        super(SCNet, self).__init__()
        self.sparseModel = scn.Sequential().add(
           scn.InputLayer(dimension, spatialSize, mode=4)).add(
           scn.SubmanifoldConvolution(dimension, 1, m, 3, False)).add(
           scn.UNet(dimension, reps, nPlanes, residual_blocks=False, downsample=[2,2])).add(
           scn.BatchNormReLU(m)).add(
           scn.OutputLayer(dimension))
        self.linear = nn.Linear(m, nClassesTotal)

    def forward(self,x) -> torch.Tensor:
        x = self.sparseModel(x)
        x = self.linear(x)
        return x
    
    def unet_foward(self,x) -> torch.Tensor:
        return self.sparseModel(x)

DataProcessing

def scn_input_wrapper(tensor_list: Iterable[torch.Tensor], scale:float, bias:float, device=None):
    D = tensor_list[0].shape[-1]
    device = device if device is not None else tensor_list[0].device
    coord_params = dict(dtype=torch.int64, device=device)
    feature_params = dict(dtype=torch.float32, device=device)
    C = torch.empty([0,D+1], **coord_params)  # (N, D+1) last dim for batch_idx
    F = torch.empty([0,1], **feature_params)  # (N,1) 1-dim feature for each point
    for b, tensor in enumerate(tensor_list):
        coords = torch.hstack([((tensor+bias)*scale).to(**coord_params), b*torch.ones([tensor.shape[0],1],**coord_params)])
        features = torch.ones([tensor.shape[0],1],**feature_params)
        C = torch.vstack([C,coords])
        F = torch.vstack([F,features])
    return [C, F]

def scn_output_wrapper(tensor:torch.Tensor, batch_idx:torch.Tensor, batch_size:int):
    batch = torch.zeros(batch_size, tensor.shape[0] // batch_size, tensor.shape[1], dtype=tensor.dtype, device=tensor.device)
    for bi in range(batch_size):
        bi_indices = batch_idx==bi
        batch[bi, ...] = tensor[bi_indices,:]
    return batch

Please download our data and unzip debug_pc1.npy and debug_pc2.npy and run the Main program.
debug_pc.zip

Main

pc1 = np.load('debug_pc1.npy')
pc2 = np.load('debug_pc2.npy')
scnet = SCNet()
# scnet.load_state_dict(torch.load("pretrained/scn.pth",map_location='cpu')) # here weights do not affect the results
scnet.eval()
scnet.cuda()
resolution = 24 # segment fault. But no fault emerged if resolution = 50
sinput1 = scn_input_wrapper([torch.from_numpy(pc1).cuda()],resolution,0)
sinput2 = scn_input_wrapper([torch.from_numpy(pc2).cuda()],resolution,0)
print("start scn forward") # segment fault is thrown right after this line
feat1 = scnet.unet_foward(sinput1).detach().cpu().numpy()
feat2 = scnet.unet_foward(sinput2).detach().cpu().numpy()
print("Feature extraction Finished.")

I annotate the segment fault in the above code snippet.
Thanks for your reply.