Question

我正在尝试自己在keras中实施VAE，而没有官方的keras教程VAE代码here的指导下-参见下面的代码。

我的损失中有nan，因此我试图找出是什么原因造成的。似乎是由kl损失引起的，因为当我仅以kl损失运行时，我得到ValueError的话说An operation has None for gradient. keras教程中的代码确实对我有用，而且我似乎找不到我的代码与教程之间的区别。

请不要评论此VAE的性能是否良好，我只想先运行它，然后对其进行微调。

from keras.layers import Input, Dense, Flatten, Reshape, Lambda
from keras.models import Model
from keras.losses import mse
from keras.datasets import mnist

import keras.backend as K

import numpy as np


(x_train,y_train),(x_test,y_test) = mnist.load_data()

def sampling(args):

    mean, log_var = args

    batch = K.shape(mean)[0]
    dim = K.int_shape(mean)[1]

    epsilon = K.random_normal(shape=(batch,dim))

    return mean + K.exp(0.5*log_var)*epsilon

latent_dim = 2

#Encoder
enc_inp = Input(shape=(28,28),name='encoder_input')
x_e = Flatten()(enc_inp)
x_e = Dense(16,activation='relu')(x_e)
x_e = Dense(16,activation='relu')(x_e)

z_mean = Dense(latent_dim,activation='linear')(x_e)
z_log_var = Dense(latent_dim,activation='linear')(x_e)
z = Lambda(sampling)([z_mean,z_log_var])
myEnc = Model(enc_inp,outputs=[z_mean,z_log_var,z])

myEnc.summary()

#Decoder
dec_inp = Input(shape=(latent_dim,))
x_d = Dense(16,activation='relu')(dec_inp)
x_d = Dense(16,activation='relu')(x_d)
x_d = Dense(784,activation='sigmoid')(x_d)
x_d = Reshape((28,28))(x_d)
myDec = Model(dec_inp,x_d)

myDec.summary()


#VAE
outputs = myDec(myEnc(enc_inp)[2])
vae = Model(enc_inp,outputs)

vae.summary()

kl_loss = 1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)
kl_loss = K.sum(kl_loss, axis=-1)
kl_loss *= -0.5

#Only kl_loss for testing as it causes problems
vae.add_loss(kl_loss)
vae.compile(optimizer='adam')

vae.fit(x_train,epochs=10,batch_size=32)

ValueError: An operation has `None` for gradient.
 Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.

“操作无梯度” Keras VAE

0 个答案: