tensorflow.python.framework.errors_impl.InvalidArgumentError：不兼容的形状：[128,784] 与 [96,784]

Question

我正在使用 python Keras 运行 VAE 模型。 我在拟合自动编码器模型时遇到了错误（这里命名为“vae”）。

错误：

tensorflow.python.framework.errors_impl.InvalidArgumentError：不兼容的形状：[128,784] 与 [96,784]

但我认为这两种形状是相同的，如下所示。 如果有人能给我一些建议，非常有帮助。谢谢。

我的代码：

from keras.layers import Lambda, Input, Dense
from keras.models import Model
from keras.datasets import mnist
from keras.losses import mse, binary_crossentropy
from keras.utils import plot_model
from keras import backend as K


original_dim = 784 #28 * 28
input_shape = (original_dim, )
intermediate_dim = 512
batch_size = 128
latent_dim = 2
epochs = 50
epsilon_std = 1.0


#encoding
x = Input(batch_shape=(batch_size, original_dim))
h = Dense(intermediate_dim, activation = 'relu')(x) #h = hidden layer

#creating latent variable = z
z_mean = Dense(latent_dim)(h) #mean of latent variable (2 dimension)
z_log_sigma = Dense(latent_dim)(h) #log_sigma of latent variable (2 dimension)

def sampling(args):
    z_mean, z_log_sigma = args
    epsilon = K.random_normal(shape=(batch_size, latent_dim),
                              mean=0., 
                              #std=epsilon_std
                              )
    return(z_mean + K.exp(z_log_sigma) * epsilon)

z = Lambda(sampling, output_shape = (latent_dim,))([z_mean, z_log_sigma])

#encoder
encoder = Model(x, z_mean)

#decoding with z
decoder_h = Dense(intermediate_dim, activation='relu')
decoder_mean = Dense(original_dim, activation='sigmoid')

h_decoded = decoder_h(z)#input encoded z
x_decoded_mean = decoder_mean(h_decoded)

#autoencoder
vae = Model(x, x_decoded_mean)

#decoding to generate images
decoder_input = Input(shape=(latent_dim,))
_h_decoded = decoder_h(decoder_input)
_x_decoded_mean = decoder_mean(_h_decoded)

#generator
generator = Model(decoder_input, _x_decoded_mean)


#training
def vae_loss(x,x_decoded_mean):
    xent_loss = binary_crossentropy(x,x_decoded_mean)
    kl_loss = -0.5 * K.mean(1+z_log_sigma-K.square(z_mean)-K.exp(z_log_sigma), axis=-1)
    return(xent_loss+kl_loss)

vae.compile(optimizer = 'rmsprop', loss=vae_loss)

#creating dataset
(x_train, y_train), (x_test, y_test) = mnist.load_data()

x_train = x_train.astype('float32') / 255.
x_test = x_test.astype('float32') / 255.
x_train = x_train.reshape((len(x_train), np.prod(x_train.shape[1:])))
x_test = x_test.reshape((len(x_test), np.prod(x_test.shape[1:])))
print('x_train', x_train.shape)
print('x_test', x_test.shape)

vae.fit(x_train, x_train,
        shuffle=True,
        epochs=epochs,
        batch_size=batch_size,
        validation_data=(x_test, x_test))

Answer 1

鉴于您的代码，您似乎在模型中硬编码批量大小，这通常是个坏主意。

由于您将输入定义为x = Input(batch_shape=(batch_size, original_dim))，因此网络期望数据完全符合(128, 784) 的形状。如果您输入的数据不完全是 128 的倍数，这是有问题的，因为 vae.fit() 将分批拆分数据，对于最后一批，如果它没有足够的数据来填充一个完整的批次，它将发送一个部分的。看来，您的最后一批样本只有 96 个而不是 128 个，因此模型引发了形状不匹配错误。

正确的做法是：不要在模型的任何地方指定batch size，只将参数传递给fit。实际上，这归结为永远不要使用 Keras 层的 batch_shape 参数，而是使用 shape 并以一个样本的形状传递：

x = Input(shape=(original_dim,))

如果您有任何操作需要您指定完整的形状（意思是，包括批量大小），请使用None：

epsilon = K.random_normal(shape=(None, latent_dim),
                          mean=0., 
                          #std=epsilon_std
                          )