pytorch中自定義backward()函數(shù)。在圖像處理過程中，我們有時候會使用自己定義的算法處理圖像，這些算法多是基于numpy或者scipy等包。
那么如何將自定義算法的梯度加入到pytorch的計算圖中，能使用Loss.backward()操作自動求導(dǎo)并優(yōu)化呢。下面的代碼展示了這個功能`
|
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 torchimport numpy as npfrom PIL import Imagefrom torch.autograd import gradcheckclass 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 ydef 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_objdef 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() |
要想實現(xiàn)自動求導(dǎo)，必須同時實現(xiàn)forward(),backward()兩個函數(shù)。
1、從代碼中可以看出來，forward()函數(shù)是針對numpy數(shù)據(jù)操作，返回值再重新指定為torch.Tensor類型。因此就有這個問題出現(xiàn)了：forward輸入input被轉(zhuǎn)換為numpy類型，輸出轉(zhuǎn)換為tensor類型，那么輸出output的grad_fn參數(shù)是如何指定的呢。調(diào)試發(fā)現(xiàn)，當(dāng)main()中hr的requires_grad被指定為True，即hr被指定為需要求導(dǎo)的葉子節(jié)點。只要Bicubic類繼承自torch.autograd.Function，那么output也就是代碼中的lr的grad_fn就會被指定為<main.Bicubic object at 0x000001DD5A280D68>，即Bicubic這個類。
2、backward()為求導(dǎo)的函數(shù)，gard_output是鏈?zhǔn)角髮?dǎo)法則的上一級的梯度，grad_input即為我們想要得到的梯度。只需要在輸入指定grad_output,在調(diào)用loss.backward()過程中的某一步會執(zhí)行到Bicubic的backwward()函數(shù)
以上這篇pytorch中的自定義反向傳播,求導(dǎo)實例就是小編分享給大家的全部內(nèi)容了，希望能給大家一個參考，也希望大家多多支持服務(wù)器之家。
原文鏈接：https://blog.csdn.net/xuxiaoyuxuxiaoyu/article/details/86737492
