反卷积_上采样对比

最新推荐文章于 2022-09-14 19:38:04 发布
最新推荐文章于 2022-09-14 19:38:04 发布
阅读量623 收藏 1
点赞数 1
文章标签： 深度学习 tensorflow
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/helldoger/article/details/106266510
版权 
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.compat.v1 import ConfigProto
from tensorflow.compat.v1 import InteractiveSession

#config = ConfigProto()
#config.gpu_options.allow_growth = True
#session = InteractiveSession(config=config)

(x_train,x_label),(y_train,y_label)=tf.keras.datasets.mnist.load_data()
x_train.shape,y_train.shape

((60000, 28, 28), (10000, 28, 28))

x_train = x_train.astype('float32')
y_train = y_train.astype('float32')
x_train = x_train/255
y_train = y_train/255
x_train = np.reshape(x_train,(len(x_train),28,28,1))
y_train = np.reshape(y_train,(len(y_train),28,28,1))
y_train.shape

(10000, 28, 28, 1)

noise_factor = 0.5
noise_x = x_train + noise_factor*np.random.normal(loc=0.0, scale=1.0, size=x_train.shape)
noise_y = y_train + noise_factor*np.random.normal(loc=0.0, scale=1.0, size=y_train.shape)

n = 20
plt.figure(figsize=(60, 12))    # 指定宽为20，高为4
for i in range(1, 11):
    ax = plt.subplot(2, n/2, i)  # 表示将设置为2行，n/2列，当前位置在i。
    plt.imshow(x_train[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
    ax = plt.subplot(2, n/2, 10+i)
    plt.imshow(noise_x[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

在这里插入图片描述

input = tf.keras.layers.Input(shape=(28,28,1))
x1 = tf.keras.layers.Conv2D(32,(3,3),padding='same',activation = 'relu',name = 'x1')(input) #变为28*28*32
x2 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2),strides=(2, 2),padding='same',name = 'x2')(x1) #变为14*14*32
x3 = tf.keras.layers.Conv2D(64,(3,3),padding='same',activation = 'relu',name = 'x3')(x2)#变为14*14*64
x4 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2),strides=(2, 2), padding="same",name = 'x4')(x3)#变为7*7*64
x5 = tf.keras.layers.Conv2D(64,(3,3),padding="same",activation = 'relu',name = 'x5')(x4)
x6 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2),strides=(2, 2), padding="same",name = 'x6')(x5) #变为4*4*64

x1.get_shape,x2.get_shape,x3.get_shape,x4.get_shape,x5.get_shape,x6.get_shape

(<bound method Tensor.get_shape of <tf.Tensor 'x1/Identity:0' shape=(None, 28, 28, 32) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'x2/Identity:0' shape=(None, 14, 14, 32) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'x3/Identity:0' shape=(None, 14, 14, 64) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'x4/Identity:0' shape=(None, 7, 7, 64) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'x5/Identity:0' shape=(None, 7, 7, 64) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'x6/Identity:0' shape=(None, 4, 4, 64) dtype=float32>>)

x7 = tf.keras.layers.Conv2DTranspose(32,(4,4),strides=(1, 1), padding='valid',name = 'x7')(x6) 
#用4*4的卷积核对齐上一层,shape=(None, 7, 7, 32)  ,padding = valid 自动填充,扩大了图片
x8 = tf.keras.layers.Conv2DTranspose(16,(2,2),strides=(2, 2), padding='same',name = 'x8')(x7) #放大图像为8*8*16
x9 = tf.keras.layers.Conv2DTranspose(1,(2,2),strides=(2, 2), padding='same',name = 'x9')(x8)#放大图像为16*16*8
x10 = tf.keras.layers.Conv2D(1,(3,3),padding="same",activation = 'sigmoid',name = 'x10')(x9)

x7.get_shape,x8.get_shape,x9.get_shape,x10.shape

(<bound method Tensor.get_shape of <tf.Tensor 'x7/Identity:0' shape=(None, 7, 7, 32) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'x8/Identity:0' shape=(None, 14, 14, 16) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'x9/Identity:0' shape=(None, 28, 28, 1) dtype=float32>>,
 TensorShape([None, 28, 28, 1]))

XXX = tf.keras.Model(inputs = input,outputs = x9)

optimizer_1 = tf.keras.optimizers.Adam(learning_rate=0.001)

XXX.compile(optimizer=optimizer_1,
            loss='binary_crossentropy',
            metrics=['accuracy'])

XXX.fit(noise_x,x_train,
        batch_size=64,
        epochs=50,
        shuffle=True,
        validation_data=(noise_y,y_train)
       )

Train on 60000 samples, validate on 10000 samples
Epoch 1/50
60000/60000 [==============================] - 9s 157us/sample - loss: 0.2303 - accuracy: 0.7986 - val_loss: 0.1954 - val_accuracy: 0.7994
Epoch 2/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.2417 - accuracy: 0.7973 - val_loss: 0.1983 - val_accuracy: 0.7980
Epoch 3/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1989 - accuracy: 0.8009 - val_loss: 0.1812 - val_accuracy: 0.8023
Epoch 4/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1857 - accuracy: 0.8026 - val_loss: 0.1743 - val_accuracy: 0.8009
Epoch 5/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1845 - accuracy: 0.8037 - val_loss: 0.1679 - val_accuracy: 0.8051
Epoch 6/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1759 - accuracy: 0.8049 - val_loss: 0.1675 - val_accuracy: 0.8040
Epoch 7/50
60000/60000 [==============================] - 5s 85us/sample - loss: 0.1631 - accuracy: 0.8064 - val_loss: 0.1553 - val_accuracy: 0.8063
Epoch 8/50
60000/60000 [==============================] - 5s 82us/sample - loss: 0.1793 - accuracy: 0.8051 - val_loss: 0.1756 - val_accuracy: 0.8041
Epoch 9/50
60000/60000 [==============================] - 5s 84us/sample - loss: 0.1788 - accuracy: 0.8047 - val_loss: 0.1698 - val_accuracy: 0.8054
Epoch 10/50
60000/60000 [==============================] - 5s 82us/sample - loss: 0.1675 - accuracy: 0.8060 - val_loss: 0.1622 - val_accuracy: 0.8050
Epoch 11/50
60000/60000 [==============================] - 5s 82us/sample - loss: 0.1796 - accuracy: 0.8049 - val_loss: 0.1792 - val_accuracy: 0.8030
Epoch 12/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1683 - accuracy: 0.8058 - val_loss: 0.1568 - val_accuracy: 0.8059
Epoch 13/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1677 - accuracy: 0.8063 - val_loss: 0.1617 - val_accuracy: 0.8048
Epoch 14/50
60000/60000 [==============================] - 5s 82us/sample - loss: 0.1629 - accuracy: 0.8074 - val_loss: 0.1528 - val_accuracy: 0.8068
Epoch 15/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1554 - accuracy: 0.8077 - val_loss: 0.1480 - val_accuracy: 0.8079
Epoch 16/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1591 - accuracy: 0.8075 - val_loss: 0.1489 - val_accuracy: 0.8063
Epoch 17/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1526 - accuracy: 0.8082 - val_loss: 0.1459 - val_accuracy: 0.8079
Epoch 18/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1665 - accuracy: 0.8068 - val_loss: 0.1522 - val_accuracy: 0.8066
Epoch 19/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1552 - accuracy: 0.8075 - val_loss: 0.1487 - val_accuracy: 0.8073
Epoch 20/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1512 - accuracy: 0.8086 - val_loss: 0.1416 - val_accuracy: 0.8083
Epoch 21/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1515 - accuracy: 0.8084 - val_loss: 0.1431 - val_accuracy: 0.8088
Epoch 22/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1507 - accuracy: 0.8089 - val_loss: 0.1520 - val_accuracy: 0.8077
Epoch 23/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1711 - accuracy: 0.8060 - val_loss: 0.1567 - val_accuracy: 0.8067
Epoch 24/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1572 - accuracy: 0.8071 - val_loss: 0.1455 - val_accuracy: 0.8085
Epoch 25/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1535 - accuracy: 0.8079 - val_loss: 0.1468 - val_accuracy: 0.8081
Epoch 26/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1544 - accuracy: 0.8077 - val_loss: 0.1477 - val_accuracy: 0.8082
Epoch 27/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1491 - accuracy: 0.8091 - val_loss: 0.1448 - val_accuracy: 0.8097
Epoch 28/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1558 - accuracy: 0.8076 - val_loss: 0.1543 - val_accuracy: 0.8065
Epoch 29/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1634 - accuracy: 0.8073 - val_loss: 0.1620 - val_accuracy: 0.8062
Epoch 30/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1489 - accuracy: 0.8085 - val_loss: 0.1470 - val_accuracy: 0.8081
Epoch 31/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1654 - accuracy: 0.8065 - val_loss: 0.1569 - val_accuracy: 0.8074
Epoch 32/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1520 - accuracy: 0.8078 - val_loss: 0.1454 - val_accuracy: 0.8074
Epoch 33/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1575 - accuracy: 0.8074 - val_loss: 0.1510 - val_accuracy: 0.8074
Epoch 34/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1523 - accuracy: 0.8081 - val_loss: 0.1439 - val_accuracy: 0.8081
Epoch 35/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1478 - accuracy: 0.8087 - val_loss: 0.1434 - val_accuracy: 0.8087
Epoch 36/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1466 - accuracy: 0.8090 - val_loss: 0.1420 - val_accuracy: 0.8089
Epoch 37/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1456 - accuracy: 0.8093 - val_loss: 0.1465 - val_accuracy: 0.8077
Epoch 38/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1615 - accuracy: 0.8070 - val_loss: 0.1455 - val_accuracy: 0.8078
Epoch 39/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1639 - accuracy: 0.8061 - val_loss: 0.1461 - val_accuracy: 0.8080
Epoch 40/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1546 - accuracy: 0.8078 - val_loss: 0.1511 - val_accuracy: 0.8065
Epoch 41/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1491 - accuracy: 0.8085 - val_loss: 0.1540 - val_accuracy: 0.8039
Epoch 42/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1511 - accuracy: 0.8090 - val_loss: 0.1442 - val_accuracy: 0.8080
Epoch 43/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1502 - accuracy: 0.8089 - val_loss: 0.1498 - val_accuracy: 0.8078
Epoch 44/50
60000/60000 [==============================] - 5s 85us/sample - loss: 0.1454 - accuracy: 0.8095 - val_loss: 0.1513 - val_accuracy: 0.8087
Epoch 45/50
27392/60000 [============>.................] - ETA: 2s - loss: 0.1655 - accuracy: 0.8057

decoded_imgs = XXX.predict(noise_y)

n = 10
plt.figure(figsize=(20, 4))
for i in range(1, n):
    # display original
    ax = plt.subplot(2, n, i)
    plt.imshow(decoded_imgs[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)

    # display reconstruction
    ax = plt.subplot(2, n, i + n)
    plt.imshow(noise_y[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-MVWyCFW2-1590062786038)(output_14_0.png)]

QQ = tf.keras.utils.plot_model(XXX, show_shapes=True)
QQ

上采样法

input_2 = tf.keras.Input(shape=(28,28,1))
xx1 = tf.keras.layers.Conv2D(16,(3,3),padding='same',activation='relu',name = 'xx1')(input_2) #28*28*16
xx2 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2), padding="same",name = 'xx2')(xx1)#strides = 2*2,14*14*16
xx3 = tf.keras.layers.Conv2D(8,(3,3),padding='same',activation='relu',name = 'xx3')(xx2) #14*14*8
xx4 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2), padding="same",name = 'xx4')(xx3)#7*7*8
xx5 = tf.keras.layers.Conv2D(8,(3,3),padding='same',activation='relu',name = 'xx5')(xx4) #7*7*8
xx6 = tf.keras.layers.MaxPooling2D(pool_size=(2, 2), padding="same",name = 'xx6')(xx5)#4*4*8

xx1.get_shape,xx2.get_shape,xx3.get_shape,xx4.get_shape,xx5.get_shape,xx6.get_shape

(<bound method Tensor.get_shape of <tf.Tensor 'xx1_1/Identity:0' shape=(None, 28, 28, 16) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx2_1/Identity:0' shape=(None, 14, 14, 16) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx3_1/Identity:0' shape=(None, 14, 14, 8) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx4_1/Identity:0' shape=(None, 7, 7, 8) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx5_1/Identity:0' shape=(None, 7, 7, 8) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx6_1/Identity:0' shape=(None, 4, 4, 8) dtype=float32>>)

xx7 = tf.keras.layers.Conv2D(8, (3, 3), activation="relu", padding="same",name = 'xx7')(xx6) #4*4*8
xx8 = tf.keras.layers.UpSampling2D((2, 2),name = 'xx8')(xx7) #上采样扩大图片 8*8*8
xx9 = tf.keras.layers.Conv2D(8, (3, 3), activation="relu", padding="same",name = 'xx9')(xx8) #8*8*8
xx10 = tf.keras.layers.UpSampling2D((2, 2),name = 'xx10')(xx9) #上采样扩大图片 16*16*8
xx11 = tf.keras.layers.Conv2D(16,(3,3),padding='valid',activation='relu',name = 'xx11')(xx10) #16*16*8## z注意,padding自适应恢复图形形状
xx12 = tf.keras.layers.UpSampling2D((2,2),name = 'xx12')(xx11) #上采样扩大图片 32*32*8
xx13 = tf.keras.layers.Conv2D(1,(3,3),padding='same',activation='sigmoid',name = 'xx13')(xx12) #16*16*1



xx7.get_shape,xx8.get_shape,xx9.get_shape,xx10.get_shape,xx11.get_shape,xx12.get_shape,xx13.get_shape

(<bound method Tensor.get_shape of <tf.Tensor 'xx7_5/Identity:0' shape=(None, 4, 4, 8) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx8_5/Identity:0' shape=(None, 8, 8, 8) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx9_5/Identity:0' shape=(None, 8, 8, 8) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx10_5/Identity:0' shape=(None, 16, 16, 8) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx11_5/Identity:0' shape=(None, 14, 14, 16) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx12_5/Identity:0' shape=(None, 28, 28, 16) dtype=float32>>,
 <bound method Tensor.get_shape of <tf.Tensor 'xx13_5/Identity:0' shape=(None, 28, 28, 1) dtype=float32>>)

XDX = tf.keras.Model(inputs = input_2,outputs = xx13)

XDX.summary()

Model: "model_5"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_3 (InputLayer)         [(None, 28, 28, 1)]       0         
_________________________________________________________________
xx1 (Conv2D)                 (None, 28, 28, 16)        160       
_________________________________________________________________
xx2 (MaxPooling2D)           (None, 14, 14, 16)        0         
_________________________________________________________________
xx3 (Conv2D)                 (None, 14, 14, 8)         1160      
_________________________________________________________________
xx4 (MaxPooling2D)           (None, 7, 7, 8)           0         
_________________________________________________________________
xx5 (Conv2D)                 (None, 7, 7, 8)           584       
_________________________________________________________________
xx6 (MaxPooling2D)           (None, 4, 4, 8)           0         
_________________________________________________________________
xx7 (Conv2D)                 (None, 4, 4, 8)           584       
_________________________________________________________________
xx8 (UpSampling2D)           (None, 8, 8, 8)           0         
_________________________________________________________________
xx9 (Conv2D)                 (None, 8, 8, 8)           584       
_________________________________________________________________
xx10 (UpSampling2D)          (None, 16, 16, 8)         0         
_________________________________________________________________
xx11 (Conv2D)                (None, 14, 14, 16)        1168      
_________________________________________________________________
xx12 (UpSampling2D)          (None, 28, 28, 16)        0         
_________________________________________________________________
xx13 (Conv2D)                (None, 28, 28, 1)         145       
=================================================================
Total params: 4,385
Trainable params: 4,385
Non-trainable params: 0
_________________________________________________________________

optimizer_2 = tf.keras.optimizers.Adam(learning_rate=0.001)

XDX.compile(optimizer=optimizer_2,
            loss = 'binary_crossentropy',
            metrics = ['acc']
           )

history_2 = XDX.fit(noise_x,x_train,
                    batch_size=64,epochs=50,
                    shuffle=True,
                    validation_data=(noise_y,y_train))

Train on 60000 samples, validate on 10000 samples
Epoch 1/50
60000/60000 [==============================] - 7s 111us/sample - loss: 0.2205 - acc: 0.7964 - val_loss: 0.1694 - val_acc: 0.8017
Epoch 2/50
60000/60000 [==============================] - 5s 83us/sample - loss: 0.1586 - acc: 0.8020 - val_loss: 0.1493 - val_acc: 0.8038
Epoch 3/50
60000/60000 [==============================] - 5s 85us/sample - loss: 0.1470 - acc: 0.8047 - val_loss: 0.1424 - val_acc: 0.8060
Epoch 4/50
60000/60000 [==============================] - 5s 83us/sample - loss: 0.1413 - acc: 0.8059 - val_loss: 0.1374 - val_acc: 0.8050
Epoch 5/50
60000/60000 [==============================] - 5s 84us/sample - loss: 0.1377 - acc: 0.8065 - val_loss: 0.1353 - val_acc: 0.8076
Epoch 6/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1350 - acc: 0.8071 - val_loss: 0.1332 - val_acc: 0.8049
Epoch 7/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1328 - acc: 0.8076 - val_loss: 0.1303 - val_acc: 0.8072
Epoch 8/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1308 - acc: 0.8080 - val_loss: 0.1290 - val_acc: 0.8058
Epoch 9/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1292 - acc: 0.8083 - val_loss: 0.1266 - val_acc: 0.8078
Epoch 10/50
60000/60000 [==============================] - 5s 83us/sample - loss: 0.1279 - acc: 0.8085 - val_loss: 0.1264 - val_acc: 0.8067
Epoch 11/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1268 - acc: 0.8088 - val_loss: 0.1244 - val_acc: 0.8081
Epoch 12/50
60000/60000 [==============================] - 5s 83us/sample - loss: 0.1258 - acc: 0.8090 - val_loss: 0.1235 - val_acc: 0.8083
Epoch 13/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1249 - acc: 0.8091 - val_loss: 0.1232 - val_acc: 0.8080
Epoch 14/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1242 - acc: 0.8093 - val_loss: 0.1225 - val_acc: 0.8078
Epoch 15/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1235 - acc: 0.8094 - val_loss: 0.1213 - val_acc: 0.8086
Epoch 16/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1228 - acc: 0.8095 - val_loss: 0.1211 - val_acc: 0.8082
Epoch 17/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1223 - acc: 0.8096 - val_loss: 0.1203 - val_acc: 0.8090
Epoch 18/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1218 - acc: 0.8097 - val_loss: 0.1214 - val_acc: 0.8101
Epoch 19/50
60000/60000 [==============================] - 5s 84us/sample - loss: 0.1213 - acc: 0.8098 - val_loss: 0.1194 - val_acc: 0.8084
Epoch 20/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1208 - acc: 0.8099 - val_loss: 0.1192 - val_acc: 0.8085
Epoch 21/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1205 - acc: 0.8100 - val_loss: 0.1191 - val_acc: 0.8097
Epoch 22/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1201 - acc: 0.8100 - val_loss: 0.1180 - val_acc: 0.8091
Epoch 23/50
60000/60000 [==============================] - 5s 83us/sample - loss: 0.1197 - acc: 0.8101 - val_loss: 0.1180 - val_acc: 0.8088
Epoch 24/50
60000/60000 [==============================] - 5s 83us/sample - loss: 0.1194 - acc: 0.8102 - val_loss: 0.1177 - val_acc: 0.8089
Epoch 25/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1191 - acc: 0.8102 - val_loss: 0.1183 - val_acc: 0.8079
Epoch 26/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1188 - acc: 0.8103 - val_loss: 0.1170 - val_acc: 0.8088
Epoch 27/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1185 - acc: 0.8103 - val_loss: 0.1167 - val_acc: 0.8097
Epoch 28/50
60000/60000 [==============================] - 5s 82us/sample - loss: 0.1182 - acc: 0.8103 - val_loss: 0.1169 - val_acc: 0.8090
Epoch 29/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1181 - acc: 0.8103 - val_loss: 0.1189 - val_acc: 0.8075
Epoch 30/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1179 - acc: 0.8104 - val_loss: 0.1159 - val_acc: 0.8095
Epoch 31/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1178 - acc: 0.8104 - val_loss: 0.1165 - val_acc: 0.8091
Epoch 32/50
60000/60000 [==============================] - 5s 84us/sample - loss: 0.1176 - acc: 0.8104 - val_loss: 0.1161 - val_acc: 0.8093
Epoch 33/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1174 - acc: 0.8105 - val_loss: 0.1169 - val_acc: 0.8107
Epoch 34/50
60000/60000 [==============================] - 5s 82us/sample - loss: 0.1173 - acc: 0.8105 - val_loss: 0.1157 - val_acc: 0.8095
Epoch 35/50
60000/60000 [==============================] - 5s 82us/sample - loss: 0.1171 - acc: 0.8105 - val_loss: 0.1156 - val_acc: 0.8099
Epoch 36/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1170 - acc: 0.8106 - val_loss: 0.1154 - val_acc: 0.8093
Epoch 37/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1168 - acc: 0.8106 - val_loss: 0.1152 - val_acc: 0.8093
Epoch 38/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1168 - acc: 0.8106 - val_loss: 0.1158 - val_acc: 0.8106
Epoch 39/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1166 - acc: 0.8106 - val_loss: 0.1159 - val_acc: 0.8108
Epoch 40/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1166 - acc: 0.8106 - val_loss: 0.1151 - val_acc: 0.8096
Epoch 41/50
60000/60000 [==============================] - 5s 76us/sample - loss: 0.1164 - acc: 0.8107 - val_loss: 0.1151 - val_acc: 0.8100
Epoch 42/50
60000/60000 [==============================] - 5s 78us/sample - loss: 0.1164 - acc: 0.8107 - val_loss: 0.1151 - val_acc: 0.8090
Epoch 43/50
60000/60000 [==============================] - 5s 79us/sample - loss: 0.1163 - acc: 0.8107 - val_loss: 0.1154 - val_acc: 0.8088
Epoch 44/50
60000/60000 [==============================] - 5s 77us/sample - loss: 0.1162 - acc: 0.8107 - val_loss: 0.1144 - val_acc: 0.8101
Epoch 45/50
60000/60000 [==============================] - 5s 78us/sample - loss: 0.1162 - acc: 0.8107 - val_loss: 0.1144 - val_acc: 0.8101
Epoch 46/50
60000/60000 [==============================] - 5s 81us/sample - loss: 0.1160 - acc: 0.8107 - val_loss: 0.1148 - val_acc: 0.8105
Epoch 47/50
60000/60000 [==============================] - 5s 84us/sample - loss: 0.1160 - acc: 0.8107 - val_loss: 0.1150 - val_acc: 0.8093
Epoch 48/50
60000/60000 [==============================] - 5s 83us/sample - loss: 0.1159 - acc: 0.8107 - val_loss: 0.1149 - val_acc: 0.8107
Epoch 49/50
60000/60000 [==============================] - 5s 80us/sample - loss: 0.1159 - acc: 0.8108 - val_loss: 0.1160 - val_acc: 0.8109
Epoch 50/50
60000/60000 [==============================] - 5s 77us/sample - loss: 0.1159 - acc: 0.8108 - val_loss: 0.1158 - val_acc: 0.8084

decoded_imgs_1 = XDX.predict(noise_y)

n = 10
plt.figure(figsize=(20, 4))
for i in range(1, n):
    # display original
    ax = plt.subplot(2, n, i)
    plt.imshow(decoded_imgs_1[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)

    # display reconstruction
    ax = plt.subplot(2, n, i + n)
    plt.imshow(noise_y[i].reshape(28, 28))
    plt.gray()
    ax.get_xaxis().set_visible(False)
    ax.get_yaxis().set_visible(False)
plt.show()

[
潘诺西亚的火山
关注
  • 1
    点赞
  • 踩
  • 1
    收藏
    觉得还不错? 一键收藏
  • 打赏
    打赏
  • 1
    评论
上采样之反卷积
qq_37844044的博客
10-06 718
上采样（Upsample） 实现feature map由小分辨率到大分辨率映射的操作 上采样常见有三种方法：（1）双线性插值（bilinear）；（2）反卷积（Transposed Convolution）；（3）反池化（Unpooling）。 反卷积（Transposed Convolution） 反卷积是一种特殊的正向卷积，先按照一定比例通过补0来扩大输入图像尺寸，然后旋转卷积核，进行正向卷积。 反卷积不是一个预先定义好的插值方式，有一些可以学习的参数（learnable parameters）。应用场
上采样、以及反卷积&空洞卷积区别
qq_34009929的博客
11-20 2869
一、 上采样(upsampling) upsampling(上采样)的三种方式： Resize，如双线性插值直接缩放，类似于图像缩放; 反卷积(deconvolution & transposed convolution); 反池化(unpooling)。 上采样upsampling的主要目的是放大图像，几乎都是采用内插值法，即在原有图像像素的基础上，在像素点值之间采用合适的插值算法插入新的元素。 传统的网络是subsampling的，对应的输出尺寸会降低；upsampling的意义在于将小尺寸的高
上采样（upsample）和 反卷积（deconvolution）
学渣的博客
12-15 2万+
目录1 卷积2 上采样3 反卷积3.1 反卷积的数学推导正向卷积的实现过程用矩阵乘法描述卷积反卷积的输入输出尺寸关系**Relationship 1：**3.3 反卷积的应用3.4 利用pytorch验证反卷积的计算参考 1 卷积 可以看一下这篇文章：https://blog.csdn.net/weixin_42468475/article/details/108628474 2 上采样 在应用在计算机视觉的深度学习领域，由于输入图像通过卷积神经网络(CNN)提取特征后，得到的特征图（Feature ma
“上采样”与“反卷积”
m0_50294896的博客
02-21 6770
在阅读YOLO模型和DenseNet网络模型的时候，对“上采样”和“反卷积”的概念不甚理解，查阅了一些资料，整理如下。并附上pytorch实现上采样的源码。 在阅读本文前，默认读者已经了解了深度学习中的卷积操作。 声明：本文用到的部分资料来自简书作者@乔大叶_803e和知乎作者@幽并游侠儿_1425 文章目录上采样（Upsample）反卷积（Transposed Convolution）pytorch实现上采样 上采样（Upsample） 在介绍上采样前，还有一个概念叫做下采样。下采样简单来讲，就是在卷
上采样和反卷积 Up-sampling and Transposed Convolution (Deconvolution)
qq_41340996的博客
05-26 712
文章目录 1. 卷积操作2. 卷积矩阵3. 反卷积4. 反卷积矩阵5. 使用卷积计算反卷积6. 总结参考文献 使用反卷积（Transposed Convolution）来进行上采样操作，不需要借助插值方法，并且可以提供能够学习的参数。 反卷积在DCGAN和图像的语义分割网络中，有很重要的应用。 1. 卷积操作 我们用一个简单的例子来介绍卷积的计算过程。如下图所示，用一个3x3的卷积核对一个4x4的矩阵进行卷积操作，步长为1，不施加padding，则可以得到一个2x2的结果矩阵。 计算输入.
反卷积(Deconvolution)、上采样(UNSampling)与上池化(UnPooling)
qq_38410428的博客
06-10 3134
反卷积(Deconvolution)、上采样(UNSampling)与上池化(UnPooling)相关代码和可视化代码在附录 反卷积(Deconvolution) 反卷积（转置卷积、空洞卷积（微步卷积））近几年用得较多，可以概括为：反卷积=上采样=（转置卷积+微步卷积）⊆ 空洞卷积=一般意义上的广义卷积（包含上采样和下采样）。 这是空洞卷积（微步卷积）的例子，其中下面的图是输入，上面的图是输出，...
超分辨率 | 反卷积deconvolution存在的弊端及引起的棋盘效应
程序员深度学习
10-27 2396
1. 棋盘效应 当我们要用到深度学习来生成图像的时候，是往往是基于一个低分辨率且具有高层语义的图像。这会使得深度学习来对这种低分辨率图像进行填充细节。一般来说，为了执行从低分辨率图像到高分辨率图像之间的转换，我们往往要进行deconvolution。简单来说，deconvolution layer可以允许模型通过每一个点进行绘制高分辨率图像上的一个方块，这种情况的产生与deconvolution的stride、kernel size有关。 然而，在使用deconvlution的时候，如果出现kernel.
CNN概念之上采样，反卷积，Unpooling概念解释
g11d111的博客
09-03 2万+
前言 本文译自quora上对《What is the difference between Deconvolution, Upsampling, Unpooling, and Convolutional Sparse Coding?》的回答。个人觉得这位老哥的回答很简洁干练。所以转到这里。 Upsampling（上采样） 在FCN、U-net等网络结构中，我们见识到了上采样这个东西。...
上采样--学习笔记
oneself的博客
09-14 684
上采样--学习笔记
CNN可视化& 反卷积.zip
04-30
2014 ECCV 深度学习卷积神经网络开始之作，CNN可视化，介绍了CNN不同层的作用及特征图的可视化，2011 ICCV 可视化反卷积论文
深度学习基础模型之FCN介绍以及反卷积、上采样具体操作
qq_39117858的博客
03-20 3185
背景 CNN对图片分类任务有很好的处理方法 https://www.cnblogs.com/gujianhan/p/6030639.html
反卷积(Deconvolution)、上采样(UnSampling)与上池化(UnPooling)
我是天才很好
12-11 2445
文章目录前言1. 图示理解2. 概念详解2.1 Upsampling(上采样)2.2 Unpooling(上池化)2.3 Deconvolution（反卷积）2.4 FCN(U-net)中的low-level和high-level特征的融合2.5 上采样/下采样3.参考文献 前言 在看图像语义分割方面的论文时，发现在网络解码器结构中有的时候使用反卷积、而有的时候使用unpooling或或者un...
膨胀卷积的缺点_卷积、反卷积与膨胀卷积
weixin_39640904的博客
12-20 497
卷积(多---->1 的映射)本质：在对输入做9--->1的映射关系时，保持了输出相对于input中的位置性关系对核矩阵做以下变形：卷积核的滑动步骤变成了卷积核矩阵的扩增卷积的矩阵乘法变成以下形式：核矩阵重排，输入featuremap变形为向量反卷积(1---->多 的映射)deconvlution/transpose-convlution动机想要网络去学出一种最优的上采样方法,...
卷积 反卷积 上采样 下采样 区别
用于记录学习的过程，不做商业用途。
11-05 3286
1.卷积 就是利用卷积核 步长前进 卷积整个图片 2.反卷积 反卷积的具体操作 原图输入尺寸为【1，3，3，3】对应【batch_size,channels,width,height】 反卷积tconv = nn.ConvTranspose2d(3,3,3,stride=2,padding=1) 求反卷积的输出尺寸？ 第一步：对输入的特征图做插值，在原先高宽方向的每两个相邻中间插上（Stride−1）列0，如下图所示，原来的尺寸为3*3，因为stride=2,则是在相邻行（列）之间插入.
FCN于反卷积(Deconvolution)、上采样(UpSampling)
热门推荐
skyfengye的博客
03-01 5万+
最近在研究图像分割方面的内容，图像分割是对图像进行像素进行分类。在这里提到分类，首先想到了 使用深度学习的方法进行分类操作的。 在这里以前的一些传统的方法进行的操作，这里就不细说了。这样使用了深度学习的方法，提到使用深度学习的方法进行图像分割的话。必须要提到 Fully Convolutional Networks for Semantic Segmentation。这篇文章。这也算是深度学
上采样(Upsampling)与反卷积(Deconvolution)的区别
qinglv1的博客
12-03 1万+
1.上采样和反卷积是不一样的，大家要注意这一点。可以参考这篇博客：https://blog.csdn.net/qq_27871973/article/details/82973048 2.对于pad默认为0，自己如果有需要得进行设置。group默认为1. 3.对于卷积和反卷积的维度计算公式：参考博客：https://blog.csdn.net/where_is_my_keyboard/arti...
反卷积（deconvolution）的理解 +上采样(UNSampling)与上池化(UnPooling)
Snoopy_Dream
11-30 1万+
像往常一样，举个梯子： https://blog.csdn.net/A_a_ron/article/details/79181108 上采样( (UNSampling)与上池化(UnPooling) https://www.zhihu.com/question/43609045/answer/132235276  反卷积 deconvolution networks   在网络解码器结构中...
反卷积和上采样+卷积的区别？
qq_38765642的博客
04-12 492
反卷积和上采样+卷积的区别？ 很详细
基于卷积神经网络的遥感图像融合
最新发布
06-02
4. 反卷积操作：将融合后的特征输入到反卷积网络中，通过反卷积和上采样操作，恢复出融合后的图像。 5. 后处理：对恢复出的融合图像进行后处理，如去噪、调整亮度和对比度等操作，得到最终的融合结果。 基于卷积...
评论 1
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

潘诺西亚的火山

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付：¥1
获取中
扫码支付

您的余额不足，请更换扫码支付或充值

打赏作者

实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值