pytorch中的自定义反向传播,求导实例

时间:2022-09-12 12:29:16

pytorch中自定义backward()函数。在图像处理过程中,我们有时候会使用自己定义的算法处理图像,这些算法多是基于numpy或者scipy等包。

那么如何将自定义算法的梯度加入到pytorch的计算图中,能使用Loss.backward()操作自动求导并优化呢。下面的代码展示了这个功能`

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
import torch
import numpy as np
from PIL import Image
from torch.autograd import gradcheck
class Bicubic(torch.autograd.Function):
def basis_function(self, x, a=-1):
  x_abs = np.abs(x)
  if x_abs < 1 and x_abs >= 0:
    y = (a + 2) * np.power(x_abs, 3) - (a + 3) * np.power(x_abs, 2) + 1
  elif x_abs > 1 and x_abs < 2:
    y = a * np.power(x_abs, 3) - 5 * a * np.power(x_abs, 2) + 8 * a * x_abs - 4 * a
  else:
    y = 0
  return y
def bicubic_interpolate(self,data_in, scale=1 / 4, mode='edge'):
  # data_in = data_in.detach().numpy()
  self.grad = np.zeros(data_in.shape,dtype=np.float32)
  obj_shape = (int(data_in.shape[0] * scale), int(data_in.shape[1] * scale), data_in.shape[2])
  data_tmp = data_in.copy()
  data_obj = np.zeros(shape=obj_shape, dtype=np.float32)
  data_in = np.pad(data_in, pad_width=((2, 2), (2, 2), (0, 0)), mode=mode)
  print(data_tmp.shape)
  for axis0 in range(obj_shape[0]):
    f_0 = float(axis0) / scale - np.floor(axis0 / scale)
    int_0 = int(axis0 / scale) + 2
    axis0_weight = np.array(
      [[self.basis_function(1 + f_0), self.basis_function(f_0), self.basis_function(1 - f_0), self.basis_function(2 - f_0)]])
    for axis1 in range(obj_shape[1]):
      f_1 = float(axis1) / scale - np.floor(axis1 / scale)
      int_1 = int(axis1 / scale) + 2
      axis1_weight = np.array(
        [[self.basis_function(1 + f_1), self.basis_function(f_1), self.basis_function(1 - f_1), self.basis_function(2 - f_1)]])
      nbr_pixel = np.zeros(shape=(obj_shape[2], 4, 4), dtype=np.float32)
      grad_point = np.matmul(np.transpose(axis0_weight, (1, 0)), axis1_weight)
      for i in range(4):
        for j in range(4):
          nbr_pixel[:, i, j] = data_in[int_0 + i - 1, int_1 + j - 1, :]
          for ii in range(data_in.shape[2]):
            self.grad[int_0 - 2 + i - 1, int_1 - 2 + j - 1, ii] = grad_point[i,j]
      tmp = np.matmul(axis0_weight, nbr_pixel)
      data_obj[axis0, axis1, :] = np.matmul(tmp, np.transpose(axis1_weight, (1, 0)))[:, 0, 0]
      # img = np.transpose(img[0, :, :, :], [1, 2, 0])
  return data_obj
 
def forward(self,input):
  print(type(input))
  input_ = input.detach().numpy()
  output = self.bicubic_interpolate(input_)
  # return input.new(output)
  return torch.Tensor(output)
 
def backward(self,grad_output):
  print(self.grad.shape,grad_output.shape)
  grad_output.detach().numpy()
  grad_output_tmp = np.zeros(self.grad.shape,dtype=np.float32)
  for i in range(self.grad.shape[0]):
    for j in range(self.grad.shape[1]):
      grad_output_tmp[i,j,:] = grad_output[int(i/4),int(j/4),:]
  grad_input = grad_output_tmp*self.grad
  print(type(grad_input))
  # return grad_output.new(grad_input)
  return torch.Tensor(grad_input)
 
def bicubic(input):
return Bicubic()(input)
 
def main():
    hr = Image.open('./baboon/baboon_hr.png').convert('L')
    hr = torch.Tensor(np.expand_dims(np.array(hr), axis=2))
    hr.requires_grad = True
    lr = bicubic(hr)
    print(lr.is_leaf)
    loss=torch.mean(lr)
    loss.backward()
if __name__ =='__main__':
    main()

要想实现自动求导,必须同时实现forward(),backward()两个函数。

1、从代码中可以看出来,forward()函数是针对numpy数据操作,返回值再重新指定为torch.Tensor类型。因此就有这个问题出现了:forward输入input被转换为numpy类型,输出转换为tensor类型,那么输出output的grad_fn参数是如何指定的呢。调试发现,当main()中hr的requires_grad被指定为True,即hr被指定为需要求导的叶子节点。只要Bicubic类继承自torch.autograd.Function,那么output也就是代码中的lr的grad_fn就会被指定为<main.Bicubic object at 0x000001DD5A280D68>,即Bicubic这个类。

2、backward()为求导的函数,gard_output是链式求导法则的上一级的梯度,grad_input即为我们想要得到的梯度。只需要在输入指定grad_output,在调用loss.backward()过程中的某一步会执行到Bicubic的backwward()函数

以上这篇pytorch中的自定义反向传播,求导实例就是小编分享给大家的全部内容了,希望能给大家一个参考,也希望大家多多支持服务器之家。

原文链接:https://blog.csdn.net/xuxiaoyuxuxiaoyu/article/details/86737492