自动编码器的二进制激活函数

Question

我有一个自动编码器，它有两个输出（解码，pred_w），一个输出是重建的输入图像，另一个是重建的二进制图像。我在最后一层使用了 sigmoid 激活函数，但输出是浮点数，我需要每个像素的网络标签为 0 或 1。我在这里附上了我的代码。你能指导我怎么做才能解决这个问题吗？ 谢谢。

from keras.layers import Input, Concatenate, GaussianNoise,Dropout
from keras.layers import Conv2D
from keras.models import Model
from keras.datasets import mnist
from keras.callbacks import TensorBoard
from keras import backend as K
from keras import layers
import matplotlib.pyplot as plt
import tensorflow as tf
import keras as Kr
import numpy as np
import pylab as pl
import matplotlib.cm as cm
import keract
from tensorflow.python.keras.layers import Lambda;

#-----------------building w train---------------------------------------------
w_main = np.random.randint(2,size=(1,4,4,1))
w_main=w_main.astype(np.float32)
w_expand=np.zeros((1,28,28,1),dtype='float32')
w_expand[:,0:4,0:4]=w_main
w_expand.reshape(1,28,28,1)
w_expand=np.repeat(w_expand,49999,0)

#-----------------building w validation---------------------------------------------
w_valid = np.random.randint(2,size=(1,4,4,1))
w_valid=w_valid.astype(np.float32)
wv_expand=np.zeros((1,28,28,1),dtype='float32')
wv_expand[:,0:4,0:4]=w_valid
wv_expand.reshape(1,28,28,1)
wv_expand=np.repeat(wv_expand,9999,0)

#-----------------building w test---------------------------------------------
w_test = np.random.randint(2,size=(1,4,4,1))
w_test=w_test.astype(np.float32)
wt_expand=np.zeros((1,28,28,1),dtype='float32')
wt_expand[:,0:4,0:4]=w_test
wt_expand.reshape(1,28,28,1)
#wt_expand=np.repeat(wt_expand,10000,0)

#-----------------------encoder------------------------------------------------
#------------------------------------------------------------------------------
wtm=Input((28,28,1))
image = Input((28, 28, 1))
conv1 = Conv2D(16, (3, 3), activation='relu', padding='same', name='convl1e')(image)
conv2 = Conv2D(32, (3, 3), activation='relu', padding='same', name='convl2e')(conv1)
conv3 = Conv2D(8, (3, 3), activation='relu', padding='same', name='convl3e')(conv2)
DrO1=Dropout(0.25)(conv3)
encoded =  Conv2D(1, (3, 3), activation='relu', padding='same',name='reconstructed_I')(DrO1)


#-----------------------adding w---------------------------------------
#add_const = Kr.layers.Lambda(lambda x: x + Kr.backend.constant(w_expand))
#encoded_merged=Kr.layers.Add()([encoded,wtm])

add_const = Kr.layers.Lambda(lambda x: x + wtm)
encoded_merged = add_const(encoded)
encoder=Model(inputs=image, outputs= encoded_merged)
encoder.summary()

#-----------------------decoder------------------------------------------------
#------------------------------------------------------------------------------

#encoded_merged = Input((28, 28, 2))
deconv1 = Conv2D(16, (3, 3), activation='relu', padding='same', name='convl1d')(encoded_merged)
deconv2 = Conv2D(32, (3, 3), activation='relu', padding='same', name='convl2d')(deconv1)
deconv3 = Conv2D(8, (3, 3), activation='relu',padding='same', name='convl3d')(deconv2)
DrO2=Dropout(0.25)(deconv3)
decoded = Conv2D(1, (3, 3), activation='relu', padding='same', name='decoder_output')(DrO2) 

#decoder=Model(inputs=encoded_merged, outputs=decoded)
#decoder.summary()
model=Model(inputs=image,outputs=decoded)
#----------------------w extraction------------------------------------
convw1 = Conv2D(16, (3,3), activation='relu', padding='same', name='conl1w')(decoded)
convw2 = Conv2D(32, (3, 3), activation='relu', padding='same', name='convl2w')(convw1)
convw3 = Conv2D(8, (3, 3), activation='relu', padding='same', name='conl3w')(convw2)
DrO3=Dropout(0.25)(convw3)
pred_w = Conv2D(1, (1, 1), activation='sigmoid', padding='same', name='reconstructed_W')(DrO3)  
# reconsider activation (is W positive?)
# should be filter=1 to match W
w_extraction=Model(inputs=[image,wtm],outputs=[decoded,pred_w])


#----------------------training the model--------------------------------------
#------------------------------------------------------------------------------
#----------------------Data preparesion----------------------------------------

(x_train, _), (x_test, _) = mnist.load_data()
x_validation=x_train[1:10000,:,:]
x_train=x_train[10001:60000,:,:]
#
x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
x_validation = x_validation.astype('float32') / 255.
x_train = np.reshape(x_train, (len(x_train), 28, 28, 1))  # adapt this if using `channels_first` image data format
x_test = np.reshape(x_test, (len(x_test), 28, 28, 1))  # adapt this if using `channels_first` image data format
x_validation = np.reshape(x_validation, (len(x_validation), 28, 28, 1))

#---------------------compile and train the model------------------------------
# is accuracy sensible metric for this model?
w_extraction.compile(optimizer='adadelta', loss={'decoder_output':'mse','reconstructed_W':'mse'}, metrics=['mae'])
w_extraction.fit([x_train,w_expand], [x_train,w_expand],
          epochs=100,
          batch_size=128, 
          validation_data=([x_validation,wv_expand], [x_validation,wv_expand]),
          callbacks=[TensorBoard(log_dir='E:/tmp/AutewithW200', histogram_freq=0, write_graph=False)])
model.summary()

Answer 1

如果你在模型中需要这个，你可以使用K.round() from keras.backend。请注意，这将无法区分，并且无法在训练中很好地使用。

如果您只需要结果，您可以简单地定义一个阈值（通常为 0.5）并且：

binary_reslts = results > threshold

为您的模型添加指标

您可以通过添加对数据进行四舍五入的指标来查看结果。 这方面的标准指标可以是"accuracy" 或"categorical_accuracy"。您可以定义自己的指标，例如：

def diceMetric(yTrue, yPred):
    yTrue = K.batch_flatten(yTrue)
    yPred = K.batch_flatten(yPred)

    #round
    yPred = K.greater(yPred, 0.5)
    yPred = K.cast(yPred, K.floatx())

    intersection = yPred * yTrue
    sum = yTrue + yPred

    return (2*intersection + K.epsilon())/(sum + K.epsilon())

指标添加在compile:

model.compile(optimizer=..., loss=..., metrics = [diceMetric, 'categorical_accuracy'])

指标不会影响训练，它们只是让您了解正在发生什么的反馈。

Answer 2

为什么您需要您的网络准确地输出 0 或 1？您可以将网络的输出解释为概率度量，即输入像素对应于 0 类或 1 类的可能性。 因此，在训练过程中，模型会尝试逼近未知的概率分布。

当涉及到预测时，您可以使用像 0.5 这样的阈值，也可以使用像 otsu 阈值这样的东西。然后您将获得二进制输出。不幸的是，阈值会产生一些间隙或缩小某些预测形状的面积。

注意： 通常，您希望在自动编码器中进行下采样和上采样，因为否则模型可能会得知同一性函数是最优的。