Question

我在基于DCGAN的GAN上工作。

我听说Generator和Discriminator需要冻结其变量而另一个正在训练。

所以我之前就这样做了。

G             = generator(batch_size, z, in_h)
D, D_logits   = discriminator(X)
D_, D_logits_ = discriminator(G, reuse=True)

# Variable Grouping
t_vars = tf.trainable_variables()
d_var = [var for var in t_vars if 'discriminator' in var.name]
g_var = [var for var in t_vars if 'generator' in var.name]

op_D = tf.train.AdamOptimizer(learning_rate=learning_rate_D, 
                          beta1=beta1).minimize(D_loss, var_list=d_var)
op_G = tf.train.AdamOptimizer(learning_rate=learning_rate_G, 
                          beta1=beta1).minimize(G_loss, var_list=g_var)

这是我的模特

def generator(batch_size, z, im_size, is_train=True):
with tf.variable_scope('generator', reuse=not is_train):

    alpha = 0.2
    #half_size, quat_size = im_size//2, im_size//4

    x1 = tf.layers.dense(z, 4*4*1024)

    x1 = tf.reshape(x1, (-1, 4, 4, 1024))
    x1 = tf.layers.batch_normalization(x1, training=is_train)
    x1 = tf.maximum(alpha*x1, x1) # leaky relu
    # 4*4*256 now

    x2 = tf.layers.conv2d_transpose(x1, 512, 5, strides=2, padding='same')
    x2 = tf.layers.batch_normalization(x2, training=is_train)
    x2 = tf.maximum(alpha*x2, x2)
    # 8*8*512 now

    x3 = tf.layers.conv2d_transpose(x2, 256, 5, strides=2, padding='same')
    x3 = tf.layers.batch_normalization(x3, training=is_train)
    x3 = tf.maximum(alpha*x3, x3)
    # 16*16*256 now

    x4 = tf.layers.conv2d_transpose(x3, 128, 5, strides=2, padding='same')
    x4 = tf.layers.batch_normalization(x4, training=is_train)
    x4 = tf.maximum(alpha*x4, x4)
    # 32*32*128 now

    logits = tf.layers.conv2d_transpose(x4, in_ch, 5, strides=2, padding='same')
    logits = tf.reshape(logits, (batch_size, im_size, im_size, in_ch))
    # 64*64*3 now

    out = tf.tanh(logits)

    return out


def discriminator(image, reuse=False, is_train=True):
with tf.variable_scope('discriminator', reuse=reuse):

    alpha = 0.2
    dr_rate = 0.3
    #input 64*64*3

    x1 = tf.layers.conv2d(image, 64, 5, strides=2, padding='same')
    x1 = tf.maximum(alpha*x1, x1)
    x1 = tf.layers.dropout(x1, rate=dr_rate, training=is_train)
    # now 32*32*64

    x2 = tf.layers.conv2d(x1, 128, 5, strides=2, padding='same')
    x2 = tf.layers.batch_normalization(x2, training=is_train)
    x2 = tf.maximum(alpha*x2, x2)
    x2 = tf.layers.dropout(x2, rate=dr_rate, training=is_train)
    # now 16*16*128

    x3 = tf.layers.conv2d(x2, 256, 5, strides=2, padding='same')
    x3 = tf.layers.batch_normalization(x3, training=is_train)
    x3 = tf.maximum(alpha*x3, x3)
    x3 = tf.layers.dropout(x3, rate=dr_rate, training=is_train)
    # now 8*8*256

    x4 = tf.layers.conv2d(x3, 512, 5, strides=2, padding='same')
    x4 = tf.layers.batch_normalization(x4, training=is_train)
    x4 = tf.maximum(alpha*x4, x4)
    x4 = tf.layers.dropout(x4, rate=dr_rate, training=is_train)
    # now 4*4*512

    flat = tf.reshape(x4, (-1, 4*4*512))
    logits = tf.layers.dense(flat, 1)
    out = tf.sigmoid(logits)

    return out, logits

请忽略范围缩进。

如果我制作第二个鉴别器，则is_train = False，结果会发生变化。

D_, D_logits_ = discriminator(G, reuse=True, is_train=False)

像这样。改变is_train = False后，G似乎无法学好。

任何人都知道为什么会这样吗？我认为此代码只会优化＆＃39; g_var＆＃39;在op_G期间，

如此改变歧视者＆＃39; is_train＆＃39;不应该影响结果。

Answer 1

让第二版鉴别器可以训练对于GAN正常工作至关重要。如果鉴别器不能根据发生器的输出更新其权重，它将永远不会学会区分真假，因此它永远不会为生成器学习提供有意义的反馈。通过将鉴别器的第二个副本设置为不可训练，您试图通过仅显示其中一个类来训练两类问题的分类器。

作为参考，您可以看到here TensorFlow的TFGAN库如何设置发生器/鉴别器网络。

tensorflow GAN，＆＃39; is_train = False＆＃39;或者＆＃39; opt.minimize（var_list = specific_vars）＆＃39;？

1 个答案: