search

yeezhu / SPN.pytorch

PyTorch implementation of "Soft Proposal Networks for Weakly Supervised Object Localization", ICCV 2017.
http://yzhu.work/spn.html
MIT License
212 stars 37 forks source link

Cuda code unnecessary? #20

Open chenyuntc opened 6 years ago

chenyuntc commented 6 years ago

Hi, Nice Work. But I think there is actually no need for CUDA code -- they could be implemented vectorized, which is more intuitive. For example:

__global__ void InitDistanceMetricKernel could be implemented as 

import torch as t
I = t.range(0, H).view(-1,1,1,1)
J = t.range(0,W).view(1,-1,1,1)
P = t.range(0,H).view(1,1,-1,1)
Q = t.range(0,W).view(1,1,1,-1)

distanceMetric_data = t.exp(   ((I-P)**2 + (J-Q)**2))  / (-2*e) ).view(H*w,-1)

Maybe incorrect, but you know what I mean.