淺談pytorch中torch.max和F.softmax函數(shù)的維度解釋
在利用torch.max函數(shù)和F.Ssoftmax函數(shù)時(shí),對(duì)應(yīng)該設(shè)置什么維度,總是有點(diǎn)懵,遂總結(jié)一下:
首先看看二維tensor的函數(shù)的例子:
import torch import torch.nn.functional as F input = torch.randn(3,4) print(input) tensor([[-0.5526, -0.0194, 2.1469, -0.2567], [-0.3337, -0.9229, 0.0376, -0.0801], [ 1.4721, 0.1181, -2.6214, 1.7721]]) b = F.softmax(input,dim=0) # 按列SoftMax,列和為1 print(b) tensor([[0.1018, 0.3918, 0.8851, 0.1021], [0.1268, 0.1587, 0.1074, 0.1218], [0.7714, 0.4495, 0.0075, 0.7762]]) c = F.softmax(input,dim=1) # 按行SoftMax,行和為1 print(c) tensor([[0.0529, 0.0901, 0.7860, 0.0710], [0.2329, 0.1292, 0.3377, 0.3002], [0.3810, 0.0984, 0.0064, 0.5143]]) d = torch.max(input,dim=0) # 按列取max, print(d) torch.return_types.max( values=tensor([1.4721, 0.1181, 2.1469, 1.7721]), indices=tensor([2, 2, 0, 2])) e = torch.max(input,dim=1) # 按行取max, print(e) torch.return_types.max( values=tensor([2.1469, 0.0376, 1.7721]), indices=tensor([2, 2, 3]))
下面看看三維tensor解釋例子:
函數(shù)softmax輸出的是所給矩陣的概率分布;
b輸出的是在dim=0維上的概率分布,b[0][5][6]+b[1][5][6]+b[2][5][6]=1
a=torch.rand(3,16,20) b=F.softmax(a,dim=0) c=F.softmax(a,dim=1) d=F.softmax(a,dim=2) In [1]: import torch as t In [2]: import torch.nn.functional as F In [4]: a=t.Tensor(3,4,5) In [5]: b=F.softmax(a,dim=0) In [6]: c=F.softmax(a,dim=1) In [7]: d=F.softmax(a,dim=2) In [8]: a Out[8]: tensor([[[-0.1581, 0.0000, 0.0000, 0.0000, -0.0344], [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000], [-0.0344, 0.0000, -0.0344, 0.0000, -0.0344], [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000]], [[-0.0344, 0.0000, -0.0344, 0.0000, -0.0344], [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000], [-0.0344, 0.0000, -0.0344, 0.0000, -0.0344], [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000]], [[-0.0344, 0.0000, -0.0344, 0.0000, -0.0344], [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000], [-0.0344, 0.0000, -0.0344, 0.0000, -0.0344], [ 0.0000, -0.0344, 0.0000, -0.0344, 0.0000]]]) In [9]: b Out[9]: tensor([[[0.3064, 0.3333, 0.3410, 0.3333, 0.3333], [0.3333, 0.3333, 0.3333, 0.3333, 0.3333], [0.3333, 0.3333, 0.3333, 0.3333, 0.3333], [0.3333, 0.3333, 0.3333, 0.3333, 0.3333]], [[0.3468, 0.3333, 0.3295, 0.3333, 0.3333], [0.3333, 0.3333, 0.3333, 0.3333, 0.3333], [0.3333, 0.3333, 0.3333, 0.3333, 0.3333], [0.3333, 0.3333, 0.3333, 0.3333, 0.3333]], [[0.3468, 0.3333, 0.3295, 0.3333, 0.3333], [0.3333, 0.3333, 0.3333, 0.3333, 0.3333], [0.3333, 0.3333, 0.3333, 0.3333, 0.3333], [0.3333, 0.3333, 0.3333, 0.3333, 0.3333]]]) In [10]: b.sum() Out[10]: tensor(20.0000) In [11]: b[0][0][0]+b[1][0][0]+b[2][0][0] Out[11]: tensor(1.0000) In [12]: c.sum() Out[12]: tensor(15.) In [13]: c Out[13]: tensor([[[0.2235, 0.2543, 0.2521, 0.2543, 0.2457], [0.2618, 0.2457, 0.2521, 0.2457, 0.2543], [0.2529, 0.2543, 0.2436, 0.2543, 0.2457], [0.2618, 0.2457, 0.2521, 0.2457, 0.2543]], [[0.2457, 0.2543, 0.2457, 0.2543, 0.2457], [0.2543, 0.2457, 0.2543, 0.2457, 0.2543], [0.2457, 0.2543, 0.2457, 0.2543, 0.2457], [0.2543, 0.2457, 0.2543, 0.2457, 0.2543]], [[0.2457, 0.2543, 0.2457, 0.2543, 0.2457], [0.2543, 0.2457, 0.2543, 0.2457, 0.2543], [0.2457, 0.2543, 0.2457, 0.2543, 0.2457], [0.2543, 0.2457, 0.2543, 0.2457, 0.2543]]]) In [14]: n=t.rand(3,4) In [15]: n Out[15]: tensor([[0.2769, 0.3475, 0.8914, 0.6845], [0.9251, 0.3976, 0.8690, 0.4510], [0.8249, 0.1157, 0.3075, 0.3799]]) In [16]: m=t.argmax(n,dim=0) In [17]: m Out[17]: tensor([1, 1, 0, 0]) In [18]: p=t.argmax(n,dim=1) In [19]: p Out[19]: tensor([2, 0, 0]) In [20]: d.sum() Out[20]: tensor(12.0000) In [22]: d Out[22]: tensor([[[0.1771, 0.2075, 0.2075, 0.2075, 0.2005], [0.2027, 0.1959, 0.2027, 0.1959, 0.2027], [0.1972, 0.2041, 0.1972, 0.2041, 0.1972], [0.2027, 0.1959, 0.2027, 0.1959, 0.2027]], [[0.1972, 0.2041, 0.1972, 0.2041, 0.1972], [0.2027, 0.1959, 0.2027, 0.1959, 0.2027], [0.1972, 0.2041, 0.1972, 0.2041, 0.1972], [0.2027, 0.1959, 0.2027, 0.1959, 0.2027]], [[0.1972, 0.2041, 0.1972, 0.2041, 0.1972], [0.2027, 0.1959, 0.2027, 0.1959, 0.2027], [0.1972, 0.2041, 0.1972, 0.2041, 0.1972], [0.2027, 0.1959, 0.2027, 0.1959, 0.2027]]]) In [23]: d[0][0].sum() Out[23]: tensor(1.)
補(bǔ)充知識(shí):多分類問(wèn)題torch.nn.Softmax的使用
為什么談?wù)撨@個(gè)問(wèn)題呢?是因?yàn)槲以诠ぷ鞯倪^(guò)程中遇到了語(yǔ)義分割預(yù)測(cè)輸出特征圖個(gè)數(shù)為16,也就是所謂的16分類問(wèn)題。
因?yàn)槊總€(gè)通道的像素的值的大小代表了像素屬于該通道的類的大小,為了在一張圖上用不同的顏色顯示出來(lái),我不得不學(xué)習(xí)了torch.nn.Softmax的使用。
首先看一個(gè)簡(jiǎn)答的例子,倘若輸出為(3, 4, 4),也就是3張4x4的特征圖。
import torch img = torch.rand((3,4,4)) print(img)
輸出為:
tensor([[[0.0413, 0.8728, 0.8926, 0.0693], [0.4072, 0.0302, 0.9248, 0.6676], [0.4699, 0.9197, 0.3333, 0.4809], [0.3877, 0.7673, 0.6132, 0.5203]], [[0.4940, 0.7996, 0.5513, 0.8016], [0.1157, 0.8323, 0.9944, 0.2127], [0.3055, 0.4343, 0.8123, 0.3184], [0.8246, 0.6731, 0.3229, 0.1730]], [[0.0661, 0.1905, 0.4490, 0.7484], [0.4013, 0.1468, 0.2145, 0.8838], [0.0083, 0.5029, 0.0141, 0.8998], [0.8673, 0.2308, 0.8808, 0.0532]]])
我們可以看到共三張?zhí)卣鲌D,每張?zhí)卣鲌D上對(duì)應(yīng)的值越大,說(shuō)明屬于該特征圖對(duì)應(yīng)類的概率越大。
import torch.nn as nn sogtmax = nn.Softmax(dim=0) img = sogtmax(img) print(img)
輸出為:
tensor([[[0.2780, 0.4107, 0.4251, 0.1979], [0.3648, 0.2297, 0.3901, 0.3477], [0.4035, 0.4396, 0.2993, 0.2967], [0.2402, 0.4008, 0.3273, 0.4285]], [[0.4371, 0.3817, 0.3022, 0.4117], [0.2726, 0.5122, 0.4182, 0.2206], [0.3423, 0.2706, 0.4832, 0.2522], [0.3718, 0.3648, 0.2449, 0.3028]], [[0.2849, 0.2076, 0.2728, 0.3904], [0.3627, 0.2581, 0.1917, 0.4317], [0.2543, 0.2898, 0.2175, 0.4511], [0.3880, 0.2344, 0.4278, 0.2686]]])
可以看到,上面的代碼對(duì)每張?zhí)卣鲌D對(duì)應(yīng)位置的像素值進(jìn)行Softmax函數(shù)處理, 圖中標(biāo)紅位置加和=1,同理,標(biāo)藍(lán)位置加和=1。
我們看到Softmax函數(shù)會(huì)對(duì)原特征圖每個(gè)像素的值在對(duì)應(yīng)維度(這里dim=0,也就是第一維)上進(jìn)行計(jì)算,將其處理到0~1之間,并且大小固定不變。
print(torch.max(img,0))
輸出為:
torch.return_types.max( values=tensor([[0.4371, 0.4107, 0.4251, 0.4117], [0.3648, 0.5122, 0.4182, 0.4317], [0.4035, 0.4396, 0.4832, 0.4511], [0.3880, 0.4008, 0.4278, 0.4285]]), indices=tensor([[1, 0, 0, 1], [0, 1, 1, 2], [0, 0, 1, 2], [2, 0, 2, 0]]))
可以看到這里3x4x4變成了1x4x4,而且對(duì)應(yīng)位置上的值為像素對(duì)應(yīng)每個(gè)通道上的最大值,并且indices是對(duì)應(yīng)的分類。
清楚理解了上面的流程,那么我們就容易處理了。
看具體案例,這里輸出output的大小為:16x416x416.
output = torch.tensor(output) sm = nn.Softmax(dim=0) output = sm(output) mask = torch.max(output,0).indices.numpy() # 因?yàn)橐D(zhuǎn)化為RGB彩色圖,所以增加一維 rgb_img = np.zeros((output.shape[1], output.shape[2], 3)) for i in range(len(mask)): for j in range(len(mask[0])): if mask[i][j] == 0: rgb_img[i][j][0] = 255 rgb_img[i][j][1] = 255 rgb_img[i][j][2] = 255 if mask[i][j] == 1: rgb_img[i][j][0] = 255 rgb_img[i][j][1] = 180 rgb_img[i][j][2] = 0 if mask[i][j] == 2: rgb_img[i][j][0] = 255 rgb_img[i][j][1] = 180 rgb_img[i][j][2] = 180 if mask[i][j] == 3: rgb_img[i][j][0] = 255 rgb_img[i][j][1] = 180 rgb_img[i][j][2] = 255 if mask[i][j] == 4: rgb_img[i][j][0] = 255 rgb_img[i][j][1] = 255 rgb_img[i][j][2] = 180 if mask[i][j] == 5: rgb_img[i][j][0] = 255 rgb_img[i][j][1] = 255 rgb_img[i][j][2] = 0 if mask[i][j] == 6: rgb_img[i][j][0] = 255 rgb_img[i][j][1] = 0 rgb_img[i][j][2] = 180 if mask[i][j] == 7: rgb_img[i][j][0] = 255 rgb_img[i][j][1] = 0 rgb_img[i][j][2] = 255 if mask[i][j] == 8: rgb_img[i][j][0] = 255 rgb_img[i][j][1] = 0 rgb_img[i][j][2] = 0 if mask[i][j] == 9: rgb_img[i][j][0] = 180 rgb_img[i][j][1] = 0 rgb_img[i][j][2] = 0 if mask[i][j] == 10: rgb_img[i][j][0] = 180 rgb_img[i][j][1] = 255 rgb_img[i][j][2] = 255 if mask[i][j] == 11: rgb_img[i][j][0] = 180 rgb_img[i][j][1] = 0 rgb_img[i][j][2] = 180 if mask[i][j] == 12: rgb_img[i][j][0] = 180 rgb_img[i][j][1] = 0 rgb_img[i][j][2] = 255 if mask[i][j] == 13: rgb_img[i][j][0] = 180 rgb_img[i][j][1] = 255 rgb_img[i][j][2] = 180 if mask[i][j] == 14: rgb_img[i][j][0] = 0 rgb_img[i][j][1] = 180 rgb_img[i][j][2] = 255 if mask[i][j] == 15: rgb_img[i][j][0] = 0 rgb_img[i][j][1] = 0 rgb_img[i][j][2] = 0 cv2.imwrite('output.jpg', rgb_img)
最后保存得到的圖為:
以上這篇淺談pytorch中torch.max和F.softmax函數(shù)的維度解釋就是小編分享給大家的全部?jī)?nèi)容了,希望能給大家一個(gè)參考,也希望大家多多支持腳本之家。
相關(guān)文章
Python使用matplotlib簡(jiǎn)單繪圖示例
這篇文章主要介紹了Python使用matplotlib簡(jiǎn)單繪圖,結(jié)合實(shí)例形式分析了Python基于matplotlib繪制正弦與余弦曲線相關(guān)操作技巧,需要的朋友可以參考下2018-02-02Python報(bào)錯(cuò)AssertionError:can only test a c
這篇文章主要介紹了Python報(bào)錯(cuò)AssertionError:can only test a child proc問(wèn)題,具有很好的參考價(jià)值,希望對(duì)大家有所幫助,如有錯(cuò)誤或未考慮完全的地方,望不吝賜教2023-09-09Python實(shí)現(xiàn)匯率轉(zhuǎn)換操作
這篇文章主要介紹了Python實(shí)現(xiàn)匯率轉(zhuǎn)換操作,具有很好的參考價(jià)值,希望對(duì)大家有所幫助。一起跟隨小編過(guò)來(lái)看看吧2020-05-05python安裝庫(kù)的最詳細(xì)方法(以安裝pygame庫(kù)為例)
在學(xué)習(xí)了一個(gè)學(xué)期的python之后,我決定對(duì)pygame下手了,下面這篇文章主要給大家介紹了關(guān)于python安裝庫(kù)的最詳細(xì)方法,本文主要以安裝pygame庫(kù)為例,文中通過(guò)圖文介紹的非常詳細(xì),需要的朋友可以參考下2023-05-05Python?實(shí)現(xiàn)驅(qū)動(dòng)AI機(jī)器人
這篇文章主要介紹了Python?實(shí)現(xiàn)驅(qū)動(dòng)AI機(jī)器人,下文圍繞利用Python?實(shí)現(xiàn)驅(qū)動(dòng)AI機(jī)器人的相關(guān)資料展開(kāi)內(nèi)容,需要的小伙伴可以參考一下2022-02-02pandas 對(duì)每一列數(shù)據(jù)進(jìn)行標(biāo)準(zhǔn)化的方法
今天小編就為大家分享一篇pandas 對(duì)每一列數(shù)據(jù)進(jìn)行標(biāo)準(zhǔn)化的方法,具有很好的參考價(jià)值,希望對(duì)大家有所幫助。一起跟隨小編過(guò)來(lái)看看吧2018-06-06PyCharm Python Console中文輸出亂碼問(wèn)題及解決
這篇文章主要介紹了PyCharm Python Console中文輸出亂碼問(wèn)題及解決方案,具有很好的參考價(jià)值,希望對(duì)大家有所幫助。如有錯(cuò)誤或未考慮完全的地方,望不吝賜教2023-07-07Python基于PyGraphics包實(shí)現(xiàn)圖片截取功能的方法
這篇文章主要介紹了Python基于PyGraphics包實(shí)現(xiàn)圖片截取功能的方法,涉及PyGraphics包的安裝及基于PyGraphics包實(shí)現(xiàn)圖片的打開(kāi)、讀取、截取等相關(guān)操作技巧,需要的朋友可以參考下2017-12-12