Pytorch?nn.Unfold()與nn.Fold()怎么使用

本篇內容主要講解“Pytorch nn.Unfold()與nn.Fold()怎么使用”，感興趣的朋友不妨來看看。本文介紹的方法操作簡單快捷，實用性強。下面就讓小編來帶大家學習“Pytorch nn.Unfold()與nn.Fold()怎么使用”吧!

Unfold()與Fold()的用途

Unfold()與Fold()一般成對出現。常用用途有：

• 代替卷積計算，Unfold()與Fold()不互逆（參數不一樣）（卷積本來就不可逆）

• 圖片patch化，Unfold()與Fold()互逆（參數一樣，且滑動窗不重疊）

nn.Unfold()

Extracts sliding local blocks from a batched input tensor.
在各滑動窗中按行展開（行向量化），然后轉置成列向量, im2col 的批量形式

input : (N, C, ∗)
output : (N, C × ∏(kernel_size), L)

# 滑動窗口有重疊
unfold = nn.Unfold(kernel_size=(2, 3))
input = torch.randn(2, 5, 3, 4)
print("input: \n", input)
output = unfold(input)
# each patch contains 30 values (2x3=6 vectors, each of 5 channels)
# 4 blocks (2x3 kernels) in total in the 3x4 input
# output.size()  # torch.Size([2, 30, 4])
print("output: \n", output)
fold = nn.Fold((3,4),(2,3))
fold_output = fold(output)
print("fold_output: \n", fold_output)

輸出為：

input:
tensor([[[[ 0.4198, 1.0535, 0.1152, 0.3510],
[ 1.1664, 0.3376, 1.2207, 0.3575],
[-0.2174, -1.2490, 0.3432, 0.3388]],

[[-1.4956, 0.9746, -0.5145, 0.1722],
[ 1.7041, 0.9645, -0.6937, -1.9037],
[-0.1961, -0.3345, 0.3565, -1.2329]],

[[-0.9843, -0.8089, 1.8712, -0.2860],
[ 0.0960, -1.7501, -0.1226, 0.9383],
[-0.1675, 1.1498, -0.4958, -1.2953]],

[[-1.2368, 0.5667, 1.4166, -2.2567],
[ 0.9414, 0.8189, 1.5604, 0.1422],
[-1.6414, -1.5594, 0.6718, 1.2319]],

[[-0.4093, 0.6691, 1.4003, 0.7444],
[-0.2858, -0.4375, -1.1301, 0.7377],
[-0.0956, -0.1844, 0.7697, -0.3077]]],

[[[ 0.4264, -0.0700, -1.5600, -0.0491],
[ 1.5027, 3.1625, 0.6080, -1.8794],
[-0.3148, 0.6377, -0.7242, 0.1692]],

[[ 0.2757, -0.5403, 0.7748, -1.1795],
[ 0.1504, -0.4671, 0.9355, 1.3050],
[-0.4920, -0.8581, 0.0559, -0.0446]],

[[ 2.1627, 0.6758, -0.0968, 1.3401],
[-0.1105, 0.8299, -0.3827, -1.0687],
[-0.2234, -1.0423, 1.2436, -0.6514]],

[[ 0.8085, -0.4159, 0.2022, 0.5747],
[-0.1265, 0.2828, -1.3530, 0.2831],
[-0.1571, 0.9005, 0.4556, -1.4360]],

[[-1.2417, 0.1829, 0.3825, -0.8555],
[-2.0170, 0.7537, 2.3406, 0.5866],
[-1.1704, -1.8986, -0.7958, 0.2652]]]])
output:
tensor([[[ 0.4198, 1.0535, 1.1664, 0.3376],
[ 1.0535, 0.1152, 0.3376, 1.2207],
[ 0.1152, 0.3510, 1.2207, 0.3575],
[ 1.1664, 0.3376, -0.2174, -1.2490],
[ 0.3376, 1.2207, -1.2490, 0.3432],
[ 1.2207, 0.3575, 0.3432, 0.3388],
[-1.4956, 0.9746, 1.7041, 0.9645],
[ 0.9746, -0.5145, 0.9645, -0.6937],
[-0.5145, 0.1722, -0.6937, -1.9037],
[ 1.7041, 0.9645, -0.1961, -0.3345],
[ 0.9645, -0.6937, -0.3345, 0.3565],
[-0.6937, -1.9037, 0.3565, -1.2329],
[-0.9843, -0.8089, 0.0960, -1.7501],
[-0.8089, 1.8712, -1.7501, -0.1226],
[ 1.8712, -0.2860, -0.1226, 0.9383],
[ 0.0960, -1.7501, -0.1675, 1.1498],
[-1.7501, -0.1226, 1.1498, -0.4958],
[-0.1226, 0.9383, -0.4958, -1.2953],
[-1.2368, 0.5667, 0.9414, 0.8189],
[ 0.5667, 1.4166, 0.8189, 1.5604],
[ 1.4166, -2.2567, 1.5604, 0.1422],
[ 0.9414, 0.8189, -1.6414, -1.5594],
[ 0.8189, 1.5604, -1.5594, 0.6718],
[ 1.5604, 0.1422, 0.6718, 1.2319],
[-0.4093, 0.6691, -0.2858, -0.4375],
[ 0.6691, 1.4003, -0.4375, -1.1301],
[ 1.4003, 0.7444, -1.1301, 0.7377],
[-0.2858, -0.4375, -0.0956, -0.1844],
[-0.4375, -1.1301, -0.1844, 0.7697],
[-1.1301, 0.7377, 0.7697, -0.3077]],

[[ 0.4264, -0.0700, 1.5027, 3.1625],
[-0.0700, -1.5600, 3.1625, 0.6080],
[-1.5600, -0.0491, 0.6080, -1.8794],
[ 1.5027, 3.1625, -0.3148, 0.6377],
[ 3.1625, 0.6080, 0.6377, -0.7242],
[ 0.6080, -1.8794, -0.7242, 0.1692],
[ 0.2757, -0.5403, 0.1504, -0.4671],
[-0.5403, 0.7748, -0.4671, 0.9355],
[ 0.7748, -1.1795, 0.9355, 1.3050],
[ 0.1504, -0.4671, -0.4920, -0.8581],
[-0.4671, 0.9355, -0.8581, 0.0559],
[ 0.9355, 1.3050, 0.0559, -0.0446],
[ 2.1627, 0.6758, -0.1105, 0.8299],
[ 0.6758, -0.0968, 0.8299, -0.3827],
[-0.0968, 1.3401, -0.3827, -1.0687],
[-0.1105, 0.8299, -0.2234, -1.0423],
[ 0.8299, -0.3827, -1.0423, 1.2436],
[-0.3827, -1.0687, 1.2436, -0.6514],
[ 0.8085, -0.4159, -0.1265, 0.2828],
[-0.4159, 0.2022, 0.2828, -1.3530],
[ 0.2022, 0.5747, -1.3530, 0.2831],
[-0.1265, 0.2828, -0.1571, 0.9005],
[ 0.2828, -1.3530, 0.9005, 0.4556],
[-1.3530, 0.2831, 0.4556, -1.4360],
[-1.2417, 0.1829, -2.0170, 0.7537],
[ 0.1829, 0.3825, 0.7537, 2.3406],
[ 0.3825, -0.8555, 2.3406, 0.5866],
[-2.0170, 0.7537, -1.1704, -1.8986],
[ 0.7537, 2.3406, -1.8986, -0.7958],
[ 2.3406, 0.5866, -0.7958, 0.2652]]])
fold_output:
tensor([[[[ 0.4198, 2.1070, 0.2304, 0.3510],
[ 2.3328, 1.3502, 4.8828, 0.7150],
[-0.2174, -2.4979, 0.6865, 0.3388]],

[[-1.4956, 1.9493, -1.0290, 0.1722],
[ 3.4083, 3.8582, -2.7747, -3.8074],
[-0.1961, -0.6690, 0.7129, -1.2329]],

[[-0.9843, -1.6178, 3.7424, -0.2860],
[ 0.1920, -7.0003, -0.4904, 1.8766],
[-0.1675, 2.2995, -0.9917, -1.2953]],

[[-1.2368, 1.1334, 2.8331, -2.2567],
[ 1.8829, 3.2758, 6.2418, 0.2843],
[-1.6414, -3.1189, 1.3435, 1.2319]],

[[-0.4093, 1.3382, 2.8006, 0.7444],
[-0.5717, -1.7500, -4.5204, 1.4754],
[-0.0956, -0.3688, 1.5395, -0.3077]]],

[[[ 0.4264, -0.1399, -3.1201, -0.0491],
[ 3.0053, 12.6500, 2.4319, -3.7589],
[-0.3148, 1.2753, -1.4484, 0.1692]],

[[ 0.2757, -1.0806, 1.5497, -1.1795],
[ 0.3007, -1.8684, 3.7421, 2.6100],
[-0.4920, -1.7162, 0.1119, -0.0446]],

[[ 2.1627, 1.3515, -0.1936, 1.3401],
[-0.2210, 3.3198, -1.5307, -2.1373],
[-0.2234, -2.0846, 2.4872, -0.6514]],

[[ 0.8085, -0.8318, 0.4044, 0.5747],
[-0.2529, 1.1310, -5.4121, 0.5662],
[-0.1571, 1.8010, 0.9112, -1.4360]],

[[-1.2417, 0.3659, 0.7650, -0.8555],
[-4.0339, 3.0146, 9.3626, 1.1733],
[-1.1704, -3.7972, -1.5917, 0.2652]]]])

Unfold()與Fold() 變化模式圖解

以上面代碼輸出為例，其實是以如下的格式對原數據進行組織排列的：
在各滑動窗中按行展開（行向量化），然后轉置成列向量, 是im2col的批量形式。

然后，對Unfold()的結果以相同參數運用Fold()后（Fold()的講解在下面，這里先給出結果），結果如下：

nn.Fold()

nn.Fold() 是 nn.Unfold() 函數的逆操作。 （參數相同、滑動窗口沒有重疊的情況下，可以完全恢復【真互逆】。滑動窗口有重疊情況下不能恢復到Unfold的輸入）

需要注意的是，如果滑動窗口有重疊，那么重疊部分相加【倍數關系】。同時，如果原來的圖像不夠劃分的話就會舍去。在恢復時就會以 0 填充。

單通道 滑動窗口無重疊

# 單通道  滑動窗口無重疊
import torch.nn as nn
import torch
batches_img = torch.rand(1,1,6,6)
print("batches_img: ",batches_img)
unfold = nn.Unfold(kernel_size=(3,3),stride=3)
patche_img = unfold(batches_img)
print("patche_img.shape: ",patche_img.shape)
print(patche_img)
fold = torch.nn.Fold(output_size=(6, 6), kernel_size=(3, 3), stride=3)
inputs_restore = fold(patche_img)
print("inputs_restore:", inputs_restore)

輸出：

batches_img: tensor([[[[0.0174, 0.3919, 0.0073, 0.4660, 0.6537, 0.0584],
[0.9763, 0.9982, 0.6250, 0.1332, 0.2123, 0.9500],
[0.5482, 0.4291, 0.9430, 0.6837, 0.6975, 0.1992],
[0.5275, 0.6800, 0.0490, 0.0350, 0.8571, 0.2449],
[0.3719, 0.7484, 0.7677, 0.4164, 0.2151, 0.8875],
[0.0784, 0.3839, 0.7567, 0.4217, 0.3208, 0.3025]]]])
patche_img.shape: torch.Size([1, 9, 4])
tensor([[[0.0174, 0.4660, 0.5275, 0.0350],
[0.3919, 0.6537, 0.6800, 0.8571],
[0.0073, 0.0584, 0.0490, 0.2449],
[0.9763, 0.1332, 0.3719, 0.4164],
[0.9982, 0.2123, 0.7484, 0.2151],
[0.6250, 0.9500, 0.7677, 0.8875],
[0.5482, 0.6837, 0.0784, 0.4217],
[0.4291, 0.6975, 0.3839, 0.3208],
[0.9430, 0.1992, 0.7567, 0.3025]]])
inputs_restore: tensor([[[[0.0174, 0.3919, 0.0073, 0.4660, 0.6537, 0.0584],
[0.9763, 0.9982, 0.6250, 0.1332, 0.2123, 0.9500],
[0.5482, 0.4291, 0.9430, 0.6837, 0.6975, 0.1992],
[0.5275, 0.6800, 0.0490, 0.0350, 0.8571, 0.2449],
[0.3719, 0.7484, 0.7677, 0.4164, 0.2151, 0.8875],
[0.0784, 0.3839, 0.7567, 0.4217, 0.3208, 0.3025]]]])

模擬圖片數據（b,3,9,9），通道數 C 為3，滑動窗口無重疊。

相較于上面的代碼，變化僅此

# 模擬圖片數據（b,3,9,9），通道數 C 為3，滑動窗口無重疊。 相較于上面的代碼，變化僅此
import torch.nn as nn
import torch
batches_img = torch.rand(1,3,6,6)
print("batches_img: ",batches_img)
unfold = nn.Unfold(kernel_size=(3,3),stride=3)
patche_img = unfold(batches_img)
print("patche_img.shape: ",patche_img.shape)
print(patche_img)
fold = torch.nn.Fold(output_size=(6, 6), kernel_size=(3, 3), stride=3)
inputs_restore = fold(patche_img)
print("inputs_restore:", inputs_restore)

輸出為：

batches_img: tensor([[[[0.6072, 0.9496, 0.4149, 0.1085, 0.6808, 0.3949],
[0.9770, 0.4831, 0.3964, 0.6597, 0.1749, 0.7326],
[0.4379, 0.0159, 0.2946, 0.4129, 0.1445, 0.5479],
[0.1664, 0.6725, 0.5104, 0.4171, 0.6656, 0.3146],
[0.5126, 0.2331, 0.8167, 0.2695, 0.6420, 0.8591],
[0.2282, 0.6300, 0.9205, 0.6741, 0.6085, 0.7866]],

[[0.7943, 0.8348, 0.5379, 0.1951, 0.2629, 0.7281],
[0.5726, 0.4912, 0.5636, 0.7816, 0.9746, 0.3764],
[0.5440, 0.3434, 0.5914, 0.5925, 0.9556, 0.0455],
[0.0810, 0.0730, 0.2580, 0.0785, 0.2483, 0.3810],
[0.4182, 0.7024, 0.4904, 0.6935, 0.1789, 0.1015],
[0.2571, 0.9138, 0.1987, 0.6266, 0.0760, 0.4618]],

[[0.3554, 0.2476, 0.3415, 0.5014, 0.1018, 0.3563],
[0.2180, 0.5690, 0.9975, 0.8152, 0.5812, 0.2704],
[0.5717, 0.9419, 0.4398, 0.5708, 0.2666, 0.3507],
[0.3868, 0.6889, 0.0326, 0.7873, 0.7444, 0.8057],
[0.1440, 0.9667, 0.2522, 0.9718, 0.6078, 0.2911],
[0.1442, 0.3061, 0.4116, 0.4190, 0.2343, 0.2608]]]])
patche_img.shape: torch.Size([1, 27, 4])
tensor([[[0.6072, 0.1085, 0.1664, 0.4171],
[0.9496, 0.6808, 0.6725, 0.6656],
[0.4149, 0.3949, 0.5104, 0.3146],
[0.9770, 0.6597, 0.5126, 0.2695],
[0.4831, 0.1749, 0.2331, 0.6420],
[0.3964, 0.7326, 0.8167, 0.8591],
[0.4379, 0.4129, 0.2282, 0.6741],
[0.0159, 0.1445, 0.6300, 0.6085],
[0.2946, 0.5479, 0.9205, 0.7866],
[0.7943, 0.1951, 0.0810, 0.0785],
[0.8348, 0.2629, 0.0730, 0.2483],
[0.5379, 0.7281, 0.2580, 0.3810],
[0.5726, 0.7816, 0.4182, 0.6935],
[0.4912, 0.9746, 0.7024, 0.1789],
[0.5636, 0.3764, 0.4904, 0.1015],
[0.5440, 0.5925, 0.2571, 0.6266],
[0.3434, 0.9556, 0.9138, 0.0760],
[0.5914, 0.0455, 0.1987, 0.4618],
[0.3554, 0.5014, 0.3868, 0.7873],
[0.2476, 0.1018, 0.6889, 0.7444],
[0.3415, 0.3563, 0.0326, 0.8057],
[0.2180, 0.8152, 0.1440, 0.9718],
[0.5690, 0.5812, 0.9667, 0.6078],
[0.9975, 0.2704, 0.2522, 0.2911],
[0.5717, 0.5708, 0.1442, 0.4190],
[0.9419, 0.2666, 0.3061, 0.2343],
[0.4398, 0.3507, 0.4116, 0.2608]]])
inputs_restore: tensor([[[[0.6072, 0.9496, 0.4149, 0.1085, 0.6808, 0.3949],
[0.9770, 0.4831, 0.3964, 0.6597, 0.1749, 0.7326],
[0.4379, 0.0159, 0.2946, 0.4129, 0.1445, 0.5479],
[0.1664, 0.6725, 0.5104, 0.4171, 0.6656, 0.3146],
[0.5126, 0.2331, 0.8167, 0.2695, 0.6420, 0.8591],
[0.2282, 0.6300, 0.9205, 0.6741, 0.6085, 0.7866]],

[[0.7943, 0.8348, 0.5379, 0.1951, 0.2629, 0.7281],
[0.5726, 0.4912, 0.5636, 0.7816, 0.9746, 0.3764],
[0.5440, 0.3434, 0.5914, 0.5925, 0.9556, 0.0455],
[0.0810, 0.0730, 0.2580, 0.0785, 0.2483, 0.3810],
[0.4182, 0.7024, 0.4904, 0.6935, 0.1789, 0.1015],
[0.2571, 0.9138, 0.1987, 0.6266, 0.0760, 0.4618]],

[[0.3554, 0.2476, 0.3415, 0.5014, 0.1018, 0.3563],
[0.2180, 0.5690, 0.9975, 0.8152, 0.5812, 0.2704],
[0.5717, 0.9419, 0.4398, 0.5708, 0.2666, 0.3507],
[0.3868, 0.6889, 0.0326, 0.7873, 0.7444, 0.8057],
[0.1440, 0.9667, 0.2522, 0.9718, 0.6078, 0.2911],
[0.1442, 0.3061, 0.4116, 0.4190, 0.2343, 0.2608]]]])

單通道 滑動窗口有重疊。

kernel_size=(3,3),stride=2

# 單通道 滑動窗口有重疊。  kernel_size=(3,3),stride=2
import torch.nn as nn
import torch
batches_img = torch.rand(1,1,6,6)
print("batches_img: \n",batches_img)
unfold = nn.Unfold(kernel_size=(3,3),stride=2)
patche_img = unfold(batches_img)
print("patche_img.shape: ",patche_img.shape)
print(patche_img)
fold = torch.nn.Fold(output_size=(6, 6), kernel_size=(3, 3), stride=2)
inputs_restore = fold(patche_img)
print("inputs_restore: \n", inputs_restore)

輸出為：

batches_img:
tensor([[[[0.4171, 0.0129, 0.2183, 0.0610, 0.5242, 0.9530],
[0.7112, 0.7892, 0.2548, 0.4604, 0.7200, 0.0294],
[0.0754, 0.0451, 0.2892, 0.6765, 0.8671, 0.5574],
[0.4220, 0.4499, 0.8946, 0.0149, 0.6790, 0.0719],
[0.1529, 0.2815, 0.8502, 0.5781, 0.0339, 0.9916],
[0.6900, 0.4843, 0.3190, 0.0676, 0.8558, 0.0060]]]])
patche_img.shape: torch.Size([1, 9, 4])
tensor([[[0.4171, 0.2183, 0.0754, 0.2892],
[0.0129, 0.0610, 0.0451, 0.6765],
[0.2183, 0.5242, 0.2892, 0.8671],
[0.7112, 0.2548, 0.4220, 0.8946],
[0.7892, 0.4604, 0.4499, 0.0149],
[0.2548, 0.7200, 0.8946, 0.6790],
[0.0754, 0.2892, 0.1529, 0.8502],
[0.0451, 0.6765, 0.2815, 0.5781],
[0.2892, 0.8671, 0.8502, 0.0339]]])
inputs_restore:
tensor([[[[0.4171, 0.0129, 0.4365, 0.0610, 0.5242, 0.0000],
[0.7112, 0.7892, 0.5095, 0.4604, 0.7200, 0.0000],
[0.1507, 0.0902, 1.1567, 1.3530, 1.7342, 0.0000],
[0.4220, 0.4499, 1.7892, 0.0149, 0.6790, 0.0000],
[0.1529, 0.2815, 1.7005, 0.5781, 0.0339, 0.0000],
[0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000]]]])

# 重復累加次數最多的元素：
batches_img[0,0,2,2]*4
# 輸出：tensor(1.1567)

卷積等價于：Unfold + Matrix Multiplication + Fold (或view()到卷積輸出形狀)

注： 使用 Unfold + Matrix Multiplication + Fold 來代替卷積時，Fold 中的 kernel size 需要為 (1,1)

inp = torch.randn(1, 3, 10, 12)
w = torch.randn(2, 3, 4, 5)
inp_unf = torch.nn.functional.unfold(inp, (4, 5))
out_unf = inp_unf.transpose(1, 2).matmul(w.view(w.size(0), -1).t()).transpose(1, 2)
out = torch.nn.functional.fold(out_unf, (7, 8), (1, 1))
# or equivalently (and avoiding a copy),
# out = out_unf.view(1, 2, 7, 8)
(torch.nn.functional.conv2d(inp, w) - out).abs().max()  # tensor(1.9073e-06)

到此，相信大家對“Pytorch nn.Unfold()與nn.Fold()怎么使用”有了更深的了解，不妨來實際操作一番吧！這里是億速云網站，更多相關內容可以進入相關頻道進行查詢，關注我們，繼續學習！