Keras VAE 的输入尺寸不正确

Question

我是机器学习的新手，我使用以下代码在 MNISET 数据集上创建示例 VAE

# We are going to use MINISET Dataset to train our GAN.
# All imports are going to be in this place.
import numpy as np
# Progressbar
from keras.callbacks import TensorBoard
from tqdm import tqdm

import matplotlib.pyplot as plt

from keras.datasets import mnist
from keras.layers import Conv2D
from keras.models import Model
from keras.layers import Input
from keras.layers import Dense
from keras.layers import LeakyReLU
from keras.layers import Dropout
from keras.layers import MaxPooling2D
from keras.layers import BatchNormalization
from keras.layers import Flatten

# All Defs will be defined here

# Importing training data of MINISET
def loadData():
    # y_train contains the labels, numbers are 1, 2 or 5, 7 etc.

    (x_train, y_train), (x_test, y_test) = mnist.load_data()
    # Normalizing based on color 255
    x_train = (x_train.astype(np.float32) - 127.5) / 127.5
    # convert shape of x_train from (60000, 28, 28) to (60000, 784)
    # 784 columns per row
    # x_train = x_train.reshape(60000, 784)
    return (x_train, y_train, x_test, y_test)

def getEncoder():
    # This returns a tensor
    inputs = Input(shape=(28, 28, 1))

    # Making a sample AlexNet Model Layer 1
    encoder = Conv2D(392, (4, 4), padding='same', activation='relu')(inputs)
    encoder = MaxPooling2D((4, 4), strides=(1, 1), padding='same')(encoder)
    encoder = BatchNormalization()(encoder)

    # Making a sample AlexNet Model Layer 2
    encoder = Conv2D(196, (2, 2), padding='same', activation='relu')(encoder)
    encoder = MaxPooling2D((4, 4), strides=(1, 1), padding='same')(encoder)
    encoder = BatchNormalization()(encoder)

    # Making a sample AlexNet Model Layer 3
    encoder = Conv2D(98, (2, 2), padding='same', activation='relu')(encoder)
    encoder = MaxPooling2D((2, 2), strides=(1, 1), padding='same')(encoder)
    encoder = BatchNormalization()(encoder)

    #encoder = Flatten()(encoder)
    #encoder = Dense(2)(encoder)
    #encoder = Dropout(0.4)(encoder)

    model = Model(inputs=inputs, outputs=encoder)
    model.compile(optimizer='adam', loss='binary_crossentropy')

    # model.fit(data, labels)  # starts training
    print(model.summary())
    return model;

def getDecoder():
    # This returns a tensor
    inputs = Input(shape=(98,))

    # a layer instance is callable on a tensor, and returns a tensor
    disc = Dense(196)(inputs)
    disc = LeakyReLU(alpha=0.2)(disc)
    disc = Dropout(0.3)(disc)

    disc = Dense(392)(disc)
    disc = LeakyReLU(alpha=0.2)(disc)
    disc = Dropout(0.3)(disc)

    disc = Dense(784, activation='sigmoid')(disc)
    disc = LeakyReLU(alpha=0.2)(disc)

    model = Model(inputs=inputs, outputs=disc)
    model.compile(optimizer='rmsprop', loss='binary_crossentropy')

    # starts training
    # model.fit(data, labels)
    print(model.summary())
    return model;

def createVAE(decoder, encoder):

    inputs = Input(shape=(28, 28, 1))
    gen = encoder(inputs)
    output = decoder(gen)

    vae = Model(inputs=inputs, outputs=output)
    print(vae.summary())
    vae.compile(loss='binary_crossentropy', optimizer='rmsprop')

    return vae

def doTraining(epochs=1, batchSize=128):
    # Loading the data
    (mniTrainX, mniTrainY, mniTestX, mniTestY) = loadData()
    # Creating GAN
    encoder = getEncoder()
    decoder = getDecoder()
    vae = createVAE(decoder, encoder)

    vae.fit(mniTrainX, mniTrainX,
                    epochs=epochs,
                    batch_size=batchSize,
                    shuffle=True,
                    validation_data=(mniTestX, mniTestX),
                    callbacks = [TensorBoard(log_dir='/tmp/autoencoder')])

doTraining(10, 128)

VAE的模型总结是

None
Model: "model_3"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_3 (InputLayer)         (None, 28, 28, 1)         0         
_________________________________________________________________
model_1 (Model)              (None, 28, 28, 98)        393862    
_________________________________________________________________
model_2 (Model)              multiple                  404740    
=================================================================
Total params: 798,602
Trainable params: 797,230
Non-trainable params: 1,372

现在它给出错误

ValueError: Error when checking input: expected input_3 to have 4 dimensions, but got array with shape (60000, 28, 28)

我在这里缺少什么，第 4 维是什么。

Answer 1

您的问题出在loadData。最后一行需要更改如下。您的编码器需要一个具有一个通道的图像，但您当前的数据没有该通道。

return (np.expand_dims(x_train, -1), y_train, np.expand_dims(x_test, -1), y_test)

但是，您会遇到更多问题。

例如，您的编码器输出以下内容。

batch_normalization_3 (Batch (None, 28, 28, 98)        392       

但您的解码器采用以下输入

input_2 (InputLayer)         (None, 98)  

然后您尝试将编码器的输出（即None, 28, 28, 98）发送到(None, 98) 输入层。哪个行不通。

实现卷积 VAE

警告：如果您刚开始使用 TF/Keras，卷积 VAE 有很多需要学习的地方。我建议使用更简单的模型，例如图像分类/自动编码器等模型，而不是跳到一个复杂的模型，如卷积 VAE。但是，如果您对此感到满意，那没关系。

from tensorflow.keras.layers import Dense, Input, Conv2D, Conv2DTranspose, MaxPooling2D, LeakyReLU, Reshape, BatchNormalization, Flatten, Lambda
from tensorflow.keras.models import Model
import tensorflow.keras.backend as K
from tensorflow.keras.datasets import mnist
import numpy as np

K.clear_session()

def loadData():
    # y_train contains the labels, numbers are 1, 2 or 5, 7 etc.

    (x_train, y_train), (x_test, y_test) = mnist.load_data()
    # Normalizing based on color 255
    x_train = x_train.astype(np.float32) /255.0
    x_test = x_test.astype(np.float32)/255.0
    # convert shape of x_train from (60000, 28, 28) to (60000, 784)
    # 784 columns per row
    # x_train = x_train.reshape(60000, 784)
    return (np.expand_dims(x_train, axis=-1), y_train, np.expand_dims(x_test, axis=-1), y_test)

# use reparameterization trick to push the sampling out as input
# note that "output_shape" isn't necessary with the TensorFlow backend



def sampling(args):
    """Reparameterization trick by sampling from an isotropic unit Gaussian.

    # Arguments
        args (tensor): mean and log of variance of Q(z|X)

    # Returns
        z (tensor): sampled latent vector
    """

    z_mean, z_log_var = args
    batch = K.shape(z_mean)[0]
    dim = K.int_shape(z_mean)[1]
    # by default, random_normal has mean = 0 and std = 1.0
    epsilon = K.random_normal(shape=(batch, dim))
    return z_mean + K.exp(0.5 * z_log_var) * epsilon


def getEncoder():
    # This returns a tensor
    inputs = Input(shape=(28, 28, 1))

    # Making a sample AlexNet Model Layer 1
    encoder = Conv2D(32, (4, 4), padding='same', activation='relu')(inputs)
    encoder = MaxPooling2D((4, 4), strides=(2, 2), padding='same')(encoder) # 14,14
    encoder = BatchNormalization()(encoder)

    # Making a sample AlexNet Model Layer 2
    encoder = Conv2D(64, (2, 2), padding='same', activation='relu')(encoder)
    encoder = MaxPooling2D((4, 4), strides=(2, 2), padding='same')(encoder) # 7, 4
    encoder = BatchNormalization()(encoder)

    latent_dim = 64
    encoder = Flatten()(encoder)
    z_mean = Dense(latent_dim, name='z_mean')(encoder)
    z_log_var = Dense(latent_dim, name='z_log_var')(encoder)

    # This is to perform the reparametrization trick
    z = Lambda(sampling, output_shape=(latent_dim,), name='z')([z_mean, z_log_var])

    encoder = Model(inputs=inputs, outputs=[z_mean, z_log_var, z])
    #encoder.summary()
    return encoder

def getDecoder():
    # This returns a tensor of shape (None, 28, 28, 1) exact same shape as input
    latent_dim = 64
    inputs = Input(shape=(latent_dim,))
    disc = Dense(7*7*1)(inputs)
    disc = LeakyReLU(alpha=0.2)(disc)
    disc = Reshape([7,7,1])(disc)
    # a layer instance is callable on a tensor, and returns a tensor
    disc = Conv2DTranspose(32, (3,3), strides=(2,2), padding='same', activation='relu')(disc)
    disc = Conv2DTranspose(1, (3,3), strides=(2,2), padding='same', activation='sigmoid')(disc)

    decoder = Model(inputs=inputs, outputs=disc)
    #decoder.summary()
    return decoder


def kl_loss(z_mean, z_log_var):

  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
  return K.mean(kl_loss)

def createVAE(decoder, encoder):

    # We are saying that the decoder takes the last output of the encoder as the input
    dec_out = decoder(encoder.outputs[2]) 
    # Defining an end-to-end model with encoder inputs and decoder outputs
    vae = Model(inputs=encoder.inputs, outputs=dec_out)
    print(vae.summary())

    # VAE loss comprises both crossentropy and KL divergence loss
    vae.compile(loss='binary_crossentropy', optimizer='rmsprop')
    vae.add_loss(kl_loss(encoder.outputs[0], encoder.outputs[1]))
    return vae


def doTraining(epochs=1, batchSize=128):
    # Loading the data
    (mniTrainX, mniTrainY, mniTestX, mniTestY) = loadData()
    # Creating GAN
    encoder = getEncoder()
    decoder = getDecoder()
    vae = createVAE(decoder, encoder)

    # I have removed the tensorboard callback. If needed add that.
    vae.fit(mniTrainX, mniTrainX,
                    epochs=epochs,
                    batch_size=batchSize,
                    shuffle=True,
                    validation_data=(mniTestX, mniTestX)
                    )

encoder = getEncoder()
decoder = getDecoder()
vae = createVAE(decoder, encoder)
doTraining()

进一步阅读

• 1 Variational Autoencoder (Not convolutional)
• 2 Convolutional AE

Answer 2

这里有一些问题。

1. 您需要将输入形状从 [BATCH_SIZE, 28, 28] 更改为 [BATCH_SIZE, 28, 28, 1]，以便卷积可以消化：np.expand_dims(x_test, 3)

2. 编码器输出是 4 维 [BATCH_SIZE, 4, 4, 98]，而解码器需要形状为 [BATCH_SIZE, 98] 的输入。您可以添加另一个 maxpooling 层以具有 [BATCH_SIZE, 1, 1, 98]，然后将其展平为 [BATCH_SIZE, 98]，或者将您的解码器重写为也是卷积的。

3. 如果您让解码器完全连接，您需要使用 vae.fit(mniTrainX, mniTrainX.reshape(60000, 784), ...) 扁平化您的参考图像

我建议您从一个非常简单的单层模型（有 1 个卷积层？）开始，然后逐渐将其演变为 VAE。