我想在Keras中實現本文:https://arxiv.org/pdf/1603.09056.pdf使用Conv-Deconv跳過連接創建圖像去噪網絡。如果我在相應的Conv-Deconv層之間建立對稱的跳過連接,但是如果我在輸入和輸出之間添加連接(如在論文中),我的網絡工作得非常好,我的網絡無法訓練。難道我不懂紙嗎?使用跳過連接時丟失不會發展
「但是,我們的網絡學習用於從輸入所述添加劑腐敗由於在輸入和網絡輸出之間的跳躍連接」
以下是在本文中描述的網絡:
這裏是我的網絡:
input_img = Input(shape=(None,None,3))
############################
####### CONVOLUTIONS #######
############################
c1 = Convolution2D(64, (3, 3))(input_img)
a1 = Activation('relu')(c1)
c2 = Convolution2D(64, (3, 3))(a1)
a2 = Activation('relu')(c2)
c3 = Convolution2D(64, (3, 3))(a2)
a3 = Activation('relu')(c3)
c4 = Convolution2D(64, (3, 3))(a3)
a4 = Activation('relu')(c4)
c5 = Convolution2D(64, (3, 3))(a4)
a5 = Activation('relu')(c5)
############################
###### DECONVOLUTIONS ######
############################
d1 = Conv2DTranspose(64, (3, 3))(a5)
a6 = Activation('relu')(d1)
m1 = add([a4, a6])
a7 = Activation('relu')(m1)
d2 = Conv2DTranspose(64, (3, 3))(a7)
a8 = Activation('relu')(d2)
m2 = add([a3, a8])
a9 = Activation('relu')(m2)
d3 = Conv2DTranspose(64, (3, 3))(a9)
a10 = Activation('relu')(d3)
m3 = add([a2, a10])
a11 = Activation('relu')(m3)
d4 = Conv2DTranspose(64, (3, 3))(a11)
a12 = Activation('relu')(d4)
m4 = add([a1, a12])
a13 = Activation('relu')(m4)
d5 = Conv2DTranspose(3, (3, 3))(a13)
a14 = Activation('relu')(d5)
m5 = add([input_img, a14]) # Everything goes well without this line
out = Activation('relu')(m5)
model = Model(input_img, out)
model.compile(optimizer='adam', loss='mse')
如果我訓練它,這裏是我得到:
Epoch 1/10
31250/31257 [============================>.] - ETA: 0s - loss: 0.0015
Current PSNR: 28.1152534485
31257/31257 [==============================] - 89s - loss: 0.0015 - val_loss: 0.0015
Epoch 2/10
31250/31257 [============================>.] - ETA: 0s - loss: 0.0015
Current PSNR: 28.1152534485
31257/31257 [==============================] - 89s - loss: 0.0015 - val_loss: 0.0015
Epoch 3/10
31250/31257 [============================>.] - ETA: 0s - loss: 0.0015
Current PSNR: 28.1152534485
31257/31257 [==============================] - 89s - loss: 0.0015 - val_loss: 0.0015
Epoch 4/10
31250/31257 [============================>.] - ETA: 0s - loss: 0.0015
Current PSNR: 28.1152534485
31257/31257 [==============================] - 89s - loss: 0.0015 - val_loss: 0.0015
Epoch 5/10
31250/31257 [============================>.] - ETA: 0s - loss: 0.0015
Current PSNR: 28.1152534485
什麼是錯我的網絡?
但是你的損失是相當低的。你爲什麼聲稱它不是在學習? –
因爲損失不會演變?它是不是應該逐步減少? –