介紹自動(dòng)編碼器實(shí)際上是一種人工神經(jīng)網(wǎng)絡(luò),用于對(duì)以無監(jiān)督方式提供的輸入數(shù)據(jù)進(jìn)行解壓縮和壓縮。解壓縮和壓縮操作是有損的且特定于數(shù)據(jù)的。數(shù)據(jù)特定意味著自動(dòng)編碼器將只能實(shí)際壓縮已經(jīng)訓(xùn)練過的數(shù)據(jù)。例如,如果你用狗的圖像訓(xùn)練一個(gè)自動(dòng)編碼器,那么它會(huì)給貓帶來糟糕的表現(xiàn)。自動(dòng)編碼器計(jì)劃學(xué)習(xí)表示對(duì)整個(gè)數(shù)據(jù)集的編碼。這可能會(huì)導(dǎo)致訓(xùn)練網(wǎng)絡(luò)降低維數(shù)。重構(gòu)部分也是通過這個(gè)學(xué)習(xí)的。

有損操作意味著重建圖像的質(zhì)量通常不如原始圖像清晰或高分辨率,并且對(duì)于具有更大損失的重建,差異更大,這被稱為有損操作。下圖顯示了如何使用特定損失因子對(duì)圖像進(jìn)行編碼和解碼。

自動(dòng)編碼器是一種特殊類型的前饋神經(jīng)網(wǎng)絡(luò),輸入應(yīng)該與輸出相似。因此,我們需要一種編碼方法、損失函數(shù)和解碼方法。最終目標(biāo)是以最小的損失完美地復(fù)制輸入。輸入將通過一層編碼器,它實(shí)際上是一個(gè)完全連接的神經(jīng)網(wǎng)絡(luò),它也構(gòu)成代碼解碼器,因此像 ANN 一樣使用相同的代碼進(jìn)行編碼和解碼。代碼實(shí)現(xiàn)通過反向傳播訓(xùn)練的 ANN 的工作方式與自動(dòng)編碼器相同。在本文中,我們將討論 3 種類型的自動(dòng)編碼器,如下所示:簡(jiǎn)單的自動(dòng)編碼器深度 CNN 自動(dòng)編碼器去噪自動(dòng)編碼器對(duì)于自動(dòng)編碼器的實(shí)現(xiàn)部分,我們將使用流行的 MNIST 數(shù)字?jǐn)?shù)據(jù)集。

1． 簡(jiǎn)單的自動(dòng)編碼器我們首先導(dǎo)入所有必要的庫:import all the dependencies
from keras．layers import Dense,Conv2D,MaxPooling2D,UpSampling2D
from keras import Input, Model
from keras．datasets import mnist
import numpy as np
import matplotlib．pyplot as plt
然后我們將構(gòu)建我們的模型,我們將提供決定輸入將被壓縮多少的維度數(shù)。維度越小,壓縮越大。encoding_dim = 15
input_img = Input(shape=(784,))
# encoded representation of input
encoded = Dense(encoding_dim, activation='relu')(input_img)
# decoded representation of code
decoded = Dense(784, activation='sigmoid')(encoded)
# Model which take input image and shows decoded images
autoencoder = Model(input_img, decoded)
然后我們需要分別構(gòu)建編碼器模型和解碼器模型,以便我們可以輕松區(qū)分輸入和輸出。# This model shows encoded images
encoder = Model(input_img, encoded)
# Creating a decoder model
encoded_input = Input(shape=(encoding_dim,))
# last layer of the autoencoder model
decoder_layer = autoencoder．layers[-1]
# decoder model
decoder = Model(encoded_input, decoder_layer(encoded_input))
然后我們需要用ADAM優(yōu)化器和交叉熵?fù)p失函數(shù)擬合來編譯模型。autoencoder．compile(optimizer='adam', loss='binary_crossentropy')
然后你需要加載數(shù)據(jù):(x_train, y_train), (x_test, y_test) = mnist．load_data()
x_train = x_train．a(chǎn)stype('float32') / 255．
x_test = x_test．a(chǎn)stype('float32') / 255．
x_train = x_train．reshape((len(x_train), np．prod(x_train．shape[1:])))
x_test = x_test．reshape((len(x_test), np．prod(x_test．shape[1:])))
print(x_train．shape)
print(x_test．shape)
輸出 :(60000, 784)
(10000, 784)
如果你想查看數(shù)據(jù)的實(shí)際情況,可以使用以下代碼行:plt．imshow(x_train[0]．reshape(28,28))
輸出 :

然后你需要訓(xùn)練你的模型:autoencoder．fit(x_train, x_train,
               epochs=15,
               batch_size=256,
               validation_data=(x_test, x_test))
輸出 :Epoch 1/15
235/235 [==============================] - 14s 5ms/step - loss: 0．4200 - val_loss: 0．2263
Epoch 2/15
235/235 [==============================] - 1s 3ms/step - loss: 0．2129 - val_loss: 0．1830
Epoch 3/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1799 - val_loss: 0．1656
Epoch 4/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1632 - val_loss: 0．1537
Epoch 5/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1533 - val_loss: 0．1481
Epoch 6/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1488 - val_loss: 0．1447
Epoch 7/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1457 - val_loss: 0．1424
Epoch 8/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1434 - val_loss: 0．1405
Epoch 9/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1415 - val_loss: 0．1388
Epoch 10/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1398 - val_loss: 0．1374
Epoch 11/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1386 - val_loss: 0．1360
Epoch 12/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1373 - val_loss: 0．1350
Epoch 13/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1362 - val_loss: 0．1341
Epoch 14/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1355 - val_loss: 0．1334
Epoch 15/15
235/235 [==============================] - 1s 3ms/step - loss: 0．1348 - val_loss: 0．1328
訓(xùn)練后,你需要提供輸入,你可以使用以下代碼繪制結(jié)果:encoded_img = encoder．predict(x_test)
decoded_img = decoder．predict(encoded_img)
plt．figure(figsize=(20, 4))
for i in range(5):
   # Display original
   ax = plt．subplot(2, 5, i + 1)
   plt．imshow(x_test[i]．reshape(28, 28))
   plt．gray()
   ax．get_xaxis()．set_visible(False)
   ax．get_yaxis()．set_visible(False)
   # Display reconstruction
   ax = plt．subplot(2, 5, i + 1 + 5)
   plt．imshow(decoded_img[i]．reshape(28, 28))
   plt．gray()
   ax．get_xaxis()．set_visible(False)
   ax．get_yaxis()．set_visible(False)
plt．show()
你可以分別清楚地看到編碼和解碼圖像的輸出,如下所示。

2． 深度 CNN 自動(dòng)編碼器由于這里的輸入是圖像,因此使用卷積神經(jīng)網(wǎng)絡(luò)或 CNN 確實(shí)更有意義。編碼器將由一堆 Conv2D 和最大池化層組成,解碼器將由一堆 Conv2D 和上采樣層組成。代碼 :model = Sequential()
# encoder network
model．a(chǎn)dd(Conv2D(30, 3, activation= 'relu', padding='same', input_shape = (28,28,1)))
model．a(chǎn)dd(MaxPooling2D(2, padding= 'same'))
model．a(chǎn)dd(Conv2D(15, 3, activation= 'relu', padding='same'))
model．a(chǎn)dd(MaxPooling2D(2, padding= 'same'))
#decoder network
model．a(chǎn)dd(Conv2D(15, 3, activation= 'relu', padding='same'))
model．a(chǎn)dd(UpSampling2D(2))
model．a(chǎn)dd(Conv2D(30, 3, activation= 'relu', padding='same'))
model．a(chǎn)dd(UpSampling2D(2))
model．a(chǎn)dd(Conv2D(1,3,activation='sigmoid', padding= 'same')) # output layer
model．compile(optimizer= 'adam', loss = 'binary_crossentropy')
model．summary()
輸出 :Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #  
=================================================================
conv2d (Conv2D)              (None, 28, 28, 30)        300      
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 14, 14, 30)        0        
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 14, 14, 15)        4065      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 7, 7, 15)          0        
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 7, 7, 15)          2040      
_________________________________________________________________
up_sampling2d (UpSampling2D) (None, 14, 14, 15)        0        
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 14, 14, 30)        4080      
_________________________________________________________________
up_sampling2d_1 (UpSampling2 (None, 28, 28, 30)        0        
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 28, 28, 1)         271      
=================================================================
Total params: 10,756
Trainable params: 10,756
Non-trainable params: 0
_________________________________________________________________
現(xiàn)在你需要加載數(shù)據(jù)并訓(xùn)練模型(x_train, _), (x_test, _) = mnist．load_data()
x_train = x_train．a(chǎn)stype('float32') / 255．
x_test = x_test．a(chǎn)stype('float32') / 255．
x_train = np．reshape(x_train, (len(x_train), 28, 28, 1))
x_test = np．reshape(x_test, (len(x_test), 28, 28, 1))
model．fit(x_train, x_train,
               epochs=15,
               batch_size=128,
               validation_data=(x_test, x_test))
輸出 :Epoch 1/15
469/469 [==============================] - 34s 8ms/step - loss: 0．2310 - val_loss: 0．0818
Epoch 2/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0811 - val_loss: 0．0764
Epoch 3/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0764 - val_loss: 0．0739
Epoch 4/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0743 - val_loss: 0．0725
Epoch 5/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0729 - val_loss: 0．0718
Epoch 6/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0722 - val_loss: 0．0709
Epoch 7/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0715 - val_loss: 0．0703
Epoch 8/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0709 - val_loss: 0．0698
Epoch 9/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0700 - val_loss: 0．0693
Epoch 10/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0698 - val_loss: 0．0689
Epoch 11/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0694 - val_loss: 0．0687
Epoch 12/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0691 - val_loss: 0．0684
Epoch 13/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0688 - val_loss: 0．0680
Epoch 14/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0685 - val_loss: 0．0680
Epoch 15/15
469/469 [==============================] - 3s 7ms/step - loss: 0．0683 - val_loss: 0．0676
現(xiàn)在你需要提供輸入并繪制以下結(jié)果的輸出pred = model．predict(x_test)
plt．figure(figsize=(20, 4))
for i in range(5):
   # Display original
   ax = plt．subplot(2, 5, i + 1)
   plt．imshow(x_test[i]．reshape(28, 28))
   plt．gray()
   ax．get_xaxis()．set_visible(False)
   ax．get_yaxis()．set_visible(False)
   # Display reconstruction
   ax = plt．subplot(2, 5, i + 1 + 5)
   plt．imshow(pred[i]．reshape(28, 28))
   plt．gray()
   ax．get_xaxis()．set_visible(False)
   ax．get_yaxis()．set_visible(False)
plt．show()
輸出 :

3． 去噪自動(dòng)編碼器現(xiàn)在我們將看到模型如何處理圖像中的噪聲。我們所說的噪聲是指模糊的圖像、改變圖像的顏色,甚至是圖像上的白色標(biāo)記。noise_factor = 0．7
x_train_noisy = x_train + noise_factor * np．random．normal(loc=0．0, scale=1．0, size=x_train．shape)
x_test_noisy = x_test + noise_factor * np．random．normal(loc=0．0, scale=1．0, size=x_test．shape)
x_train_noisy = np．clip(x_train_noisy, 0．, 1．)
x_test_noisy = np．clip(x_test_noisy, 0．, 1．)
Here is how the noisy images look right now．
plt．figure(figsize=(20, 2))
for i in range(1, 5 + 1):
   ax = plt．subplot(1, 5, i)
   plt．imshow(x_test_noisy[i]．reshape(28, 28))
   plt．gray()
   ax．get_xaxis()．set_visible(False)
   ax．get_yaxis()．set_visible(False)
plt．show()
輸出 :

現(xiàn)在圖像幾乎無法識(shí)別,為了增加自動(dòng)編碼器的范圍,我們將修改定義模型的層以增加過濾器,使模型性能更好,然后擬合模型。model = Sequential()
# encoder network
model．a(chǎn)dd(Conv2D(35, 3, activation= 'relu', padding='same', input_shape = (28,28,1)))
model．a(chǎn)dd(MaxPooling2D(2, padding= 'same'))
model．a(chǎn)dd(Conv2D(25, 3, activation= 'relu', padding='same'))
model．a(chǎn)dd(MaxPooling2D(2, padding= 'same'))
#decoder network
model．a(chǎn)dd(Conv2D(25, 3, activation= 'relu', padding='same'))
model．a(chǎn)dd(UpSampling2D(2))
model．a(chǎn)dd(Conv2D(35, 3, activation= 'relu', padding='same'))
model．a(chǎn)dd(UpSampling2D(2))
model．a(chǎn)dd(Conv2D(1,3,activation='sigmoid', padding= 'same')) # output layer
model．compile(optimizer= 'adam', loss = 'binary_crossentropy')
model．fit(x_train_noisy, x_train,
               epochs=15,
               batch_size=128,
               validation_data=(x_test_noisy, x_test))
輸出 :Epoch 1/15
469/469 [==============================] - 5s 9ms/step - loss: 0．2643 - val_loss: 0．1456
Epoch 2/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1440 - val_loss: 0．1378
Epoch 3/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1373 - val_loss: 0．1329
Epoch 4/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1336 - val_loss: 0．1305
Epoch 5/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1313 - val_loss: 0．1283
Epoch 6/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1294 - val_loss: 0．1268
Epoch 7/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1278 - val_loss: 0．1257
Epoch 8/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1267 - val_loss: 0．1251
Epoch 9/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1259 - val_loss: 0．1244
Epoch 10/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1251 - val_loss: 0．1234
Epoch 11/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1241 - val_loss: 0．1234
Epoch 12/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1239 - val_loss: 0．1222
Epoch 13/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1232 - val_loss: 0．1223
Epoch 14/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1226 - val_loss: 0．1215
Epoch 15/15
469/469 [==============================] - 4s 8ms/step - loss: 0．1221 - val_loss: 0．1211
訓(xùn)練結(jié)束后,我們將提供輸入并編寫繪圖函數(shù)以查看最終結(jié)果。pred = model．predict(x_test_noisy)
plt．figure(figsize=(20, 4))
for i in range(5):
   # Display original
   ax = plt．subplot(2, 5, i + 1)
   plt．imshow(x_test_noisy[i]．reshape(28, 28))
   plt．gray()
   ax．get_xaxis()．set_visible(False)
   ax．get_yaxis()．set_visible(False)
   # Display reconstruction
   ax = plt．subplot(2, 5, i + 1 + 5)
   plt．imshow(pred[i]．reshape(28, 28))
   plt．gray()
   ax．get_xaxis()．set_visible(False)
   ax．get_yaxis()．set_visible(False)
plt．show()
輸出 :

尾注我們已經(jīng)了解了自動(dòng)編碼器工作的結(jié)構(gòu),并使用了 3 種類型的自動(dòng)編碼器。自動(dòng)編碼器有多種用途,如降維圖像壓縮、電影和歌曲推薦系統(tǒng)等。模型的性能可以通過訓(xùn)練更多的 epoch 或增加我們網(wǎng)絡(luò)的維度來提高。