有人可以帮助我从DCGAN获得正确的结果/图像吗？ 我从一个颜色的图片中得到了不同的颜色（噪音）图片 迭代到另一个，但在任何地方都不接近我应该 得到。我要喂...

Question

def generator(latent_dim, n_classes):
    
    initializer = tf.random_normal_initializer(0., 0.021)
    in_label = Input(shape=(1,))
    li = Embedding(n_classes, 25)(in_label)
    n_nodes = 64 * 64
    
    li = Dense(n_nodes)(li)
    
    li = Reshape((64, 64, 1))(li)
    
    
    in_lat = Input(shape=(latent_dim,))
    n_nodes = 128 * 64 * 64
    
    gen = Dense(n_nodes)(in_lat)
    
    gen = LeakyReLU(alpha=0.2)(gen)
    gen = BatchNormalization(axis=-1)(gen)
    gen = Reshape((64, 64, 128))(gen)

    merge = Concatenate()([gen, li])
    gen = Conv2DTranspose(128, (4,4), strides=(2,2), padding='same',kernel_initializer=initializer,use_bias=False)(merge)
    gen = LeakyReLU(alpha=0.2)(gen)
    gen = BatchNormalization(axis=-1)(gen)
    

    gen = Conv2DTranspose(128, (4,4), strides=(2,2), padding='same',kernel_initializer=initializer,use_bias=False)(gen)
    gen = LeakyReLU(alpha=0.2)(gen)
    gen = BatchNormalization(axis=-1)(gen)

    
    out_layer = Conv2D(3, (7,7), activation='tanh', padding='same',kernel_initializer=initializer,use_bias=False)(gen)
    model = Model([in_lat, in_label], out_layer, name="generator")
    
    opt = Adam(lr=0.0002, beta_1=0.5)
    model.compile(optimizer=opt,loss='binary_crossentropy', metrics='accuracy')
    
    return model



def define_discriminator(n_classes,in_shape=(256,256,3)):
    
    initializer = tf.random_normal_initializer(0., 0.021)
    in_label = Input(shape=(1,))
    li = Embedding(n_classes, 25)(in_label)     
    n_nodes = in_shape[0] * in_shape[1]
    
    li = Dense(n_nodes)(li)

    li = Reshape((in_shape[0], in_shape[1],1))(li)
    
    in_image = Input(shape=in_shape)
    merge = Concatenate()([in_image, li])
    
    fe = Conv2D(128, (3,3), strides=(2,2), padding='same',kernel_initializer=initializer,use_bias=False)(merge)
    fe = LeakyReLU(alpha=0.2)(fe)
    fe = BatchNormalization(axis=-1)(fe)

    fe = Conv2D(128, (3,3), strides=(2,2), padding='same',kernel_initializer=initializer,use_bias=False)(fe)
    fe = LeakyReLU(alpha=0.2)(fe)
    fe = BatchNormalization(axis=-1)(fe)
    
    
    fe = Flatten()(fe)
    fe = Dropout(0.4)(fe)

    out_layer = Dense(1, activation='sigmoid')(fe)
    model = Model([in_image, in_label], out_layer, name="discriminator")

    opt = Adam(lr=0.0002, beta_1=0.5)
    model.compile(optimizer=opt, loss='binary_crossentropy', metrics=['accuracy'])
    
    return model



def define_gan(g_model, d_model):
    
    g_model.trainable = True
    d_model.trainable = False
    
    gen_noise, gen_label = g_model.input
    gen_output = g_model.output
    
    gan_output = d_model([gen_output, gen_label])
    model = Model([gen_noise, gen_label], gan_output, name="gan")
    opt = Adam(lr=0.0002, beta_1=0.5)
    model.compile(optimizer=opt,  loss='binary_crossentropy', metrics='accuracy')
    return model

below是不同迭代的结果/图像。 16个正方形是16个 不同的潜在载体种子。

没有看到任何代码很难推测，但是您可以检查thisRepo以进行参考。 关注此注意：

分开，出于逃脱我们的原因，本笔记本中的歧视者几乎总是无法学习，如果您仅在CPU上训练。由于这种失败，gan很少学习如何生成草图 - 即，它将输出仅是随机噪声的图像。我们已经确定了两种方法来解决这种情况：

使用gpu

Answer 1

改变歧视器的优化器。正如本笔记本的歧视汇编步骤中的评论所指出的那样，从默认的RMSPROP优化器切换到另一个（例如Adam或Adadelta），使歧视器可以有效地学习，因此GAN生成了草图。无论您仅使用CPU，GPU还是TPU，此解决方案都是有效的。 （也就是说，用GPU训练GAN的速度仍然比仅使用CPU或TPU快。） 您还可以研究此讨论：

Https：//github.com/the-deep-learners/deep-learning-ullustrated/issues/2