双线性插值实现（bilinear）

双线性插值实现（bilinear）

import torch.nn.functional as F
import torch


def linear(x, x1, x2, v1, v2):
    return (x-x1)/(x2-x1)*(v2-v1)+v1


def bilinear(x, y, src):
    x1 = int(x)
    y1 = int(y)
    x2 = x1 + 1
    y2 = y1 + 1
    h, w = src.shape
    if x == h-1 or y == w-1:
        if x == h-1 and y < w-1:
            value = linear(y, y1, y2, src[x1, y1], src[x1, y2])
        elif x < h-1 and y == w-1:
            value = linear(x, x1, x2, src[x1, y1], src[x2, y1])
        else:
            value = src[x1, y1]
    else:
        v11 = src[x1, y1]
        v21 = src[x2, y1]
        v12 = src[x1, y2]
        v22 = src[x2, y2]
        v_y1 = linear(x, x1, x2, v11, v21)
        v_y2 = linear(x, x1, x2, v12, v22)
        value = linear(y, y1, y2, v_y1, v_y2)
    return value


def get_src_coordinate(x, y, src, des):
    src_h, src_w = src.shape
    des_h, des_w = des.shape
    src_x = src_h/des_h*(x+0.5)-0.5
    src_y = src_w/des_w*(y+0.5)-0.5
    src_x = get_new_coordinate(src_x, src_h-1)
    src_y = get_new_coordinate(src_y, src_w-1)
    return src_x, src_y


def get_new_coordinate(value, boundary):
    value = max(value, 0)
    value = min(value, boundary)
    return value


def get_des(src, des_shape):
    des = torch.zeros(des_shape)
    h, w = des_shape
    for x in range(h):
        for y in range(w):
            src_x, src_y = get_src_coordinate(x, y, src, des)
            des[x][y] = bilinear(src_x, src_y, src)
    return des


a = torch.arange(9, dtype=torch.float32).reshape(1, 1, 3, 3)
b = F.interpolate(a, size=(5, 5), mode='bilinear')
a = a.squeeze()
b = b.squeeze()
des = get_des(a, b.shape)
print(torch.sum(b - des))

参考链接：https://blog.csdn.net/Big_Huang/article/details/106209992