[pytorch][grid_sample]: 知识记录

torch.nn.functional.grid_sample

torch中实现双线性插值或者双线性采样，参考：https://fesian.blog.csdn.net/article/details/108732249

output = F.grid_sample(inputv, grid, align_corners=True)  # 将inputv通过采样或插值缩放到output。

grid shape为NHoutWout2，表示输出来源于输入哪个位置x,y，xy归一化到[-1,+1]之间了，x = -1, y = -1 is the left-top pixel input, and values x = 1, y = 1 is the right-bottom pixel of input

import torch.nn.functional as F
import torch
inputv = torch.arange(4*4).view(1, 1, 4, 4).float()
print(inputv)
'''
输出尺寸为(1,1,4,4)
输出为：tensor([[[[ 0.,  1.,  2.,  3.],
          [ 4.,  5.,  6.,  7.],
          [ 8.,  9., 10., 11.],
          [12., 13., 14., 15.]]]])
'''
# 生成grid，这个grid大小为(1,8,8,2)，空间尺寸而言是原输入图片的两倍。
d = torch.linspace(-1,1, 8)
meshx, meshy = torch.meshgrid((d, d))
grid = torch.stack((meshy, meshx), 2)
grid = grid.unsqueeze(0) # add batch dim

# 进行双线性采样，其中指定align_corners=True保证了输出的整个图片的角边像素与原输入的一致性。
output = F.grid_sample(inputv, grid, align_corners=True)
print(output)
'''
tensor([[[[ 0.0000,  0.4286,  0.8571,  1.2857,  1.7143,  2.1429,  2.5714,
            3.0000],
          [ 1.7143,  2.1429,  2.5714,  3.0000,  3.4286,  3.8571,  4.2857,
            4.7143],
          [ 3.4286,  3.8571,  4.2857,  4.7143,  5.1429,  5.5714,  6.0000,
            6.4286],
          [ 5.1429,  5.5714,  6.0000,  6.4286,  6.8571,  7.2857,  7.7143,
            8.1429],
          [ 6.8571,  7.2857,  7.7143,  8.1429,  8.5714,  9.0000,  9.4286,
            9.8571],
          [ 8.5714,  9.0000,  9.4286,  9.8571, 10.2857, 10.7143, 11.1429,
           11.5714],
          [10.2857, 10.7143, 11.1429, 11.5714, 12.0000, 12.4286, 12.8571,
           13.2857],
          [12.0000, 12.4286, 12.8571, 13.2857, 13.7143, 14.1429, 14.5714,
           15.0000]]]])
'''