redefine index

[aivision2020]

PointCNN.Pytorch/utils/util_funcs.py

Line 57 in 6ec6c29

n, d = ref.size()

It's unhealthy to define a variable "n" inside a loop that uses "n" as it's enumerator

(how does this even work?)

[aabbcco]

PointCNN.Pytorch/utils/util_funcs.py

Line 57 in 6ec6c29

n, d = ref.size()

It's unhealthy to define a variable "n" inside a loop that uses "n" as it's enumerator

(how does this even work?)

seems the function is absolutely wrong,giving an incorrect index for K-Nearst Neighbors
knn is much easier in torch after torch.topk function released,just follow the codes below

def knn_indices_func_gpuuu(seed:cuda.FloatTensor,   #(B,C,npoint)
                            pts:cuda.FloatTensor,   #(B,C,N)
                            k:int                   
                            ):     #(N,npoint,K)
    _,_,N = seed.shape
    _,_,M = pts.shape
    mseed = seed.unsqueeze(-2).expand(-1,-1,M,-1)
    mpts  = pts.unsqueeze(-1).expand(-1,-1,-1,N)
    mdist = torch.sum((mpts-mseed)**2,dim=1)
    _,idx = torch.topk(mdist,k=k+1,largest=False)

    return idx[:,:,1:]

alone with the test code

import torch
from torch import cuda, FloatTensor, LongTensor
from shapenet import shapenet
from torch.utils.data import DataLoader

# shapedata = shapenet('../shapenet_part_seg_hdf5_data',split='val')
# shapeloader = DataLoader(shapedata,batch_size=5)
#data,label = iter(shapeloader).next()
#data=data.to('cuda')
#src = data[:,:,:10]
data = torch.tensor([[1,2,3,4,5,6,7,8,9],[1,2,3,4,5,6,7,8,9]]).to('cuda').unsqueeze(0).repeat(2,1,1)
src = torch.tensor([[0,1,2],[0,1,2]]).unsqueeze(0).to('cuda').repeat(2,1,1)
print(src.shape)
print(data.shape)

idx = knn_indices_func_gpuuu(data,src,2)
print(idx.shape)
print(idx)

and result

torch.Size([2, 2, 3])
torch.Size([2, 2, 9])
torch.Size([2, 3, 2])
tensor([[[1, 2],
         [1, 2],
         [2, 0]],

        [[1, 2],
         [1, 2],
         [2, 0]]], device='cuda:0')

hope it will help you even after two and half years,:D