Tensorflow GAN：“没有为任何变量提供梯度”

Question

我正在尝试使用 TF 设置 GAN，但我太愚蠢了。 我搜索了网络，但找不到答案。

当我运行提供的代码时，我得到：

gen_optimize = tf.train.AdamOptimizer(learning_rate, beta1).minimize(gen_loss, var_list=gen_vars)

ValueError: 没有为任何变量提供梯度，请检查您的图表中变量之间不支持梯度的操作

def generator(z, activation=tf.nn.relu, reuse=False):
    shape = z.get_shape().as_list()
    weight_init = tf.random_normal_initializer(stddev=0.1)
    bias_init = tf.constant_initializer(0.0)

    fc1_units = 256
    fc1_weights = tf.get_variable('gen_fc1_weights', (shape[1], fc1_units), dtype=tf.float32, initializer=weight_init)
    fc1_biases = tf.get_variable('gen_fc1_biases', (fc1_units), initializer=bias_init)
    fc1 = activation(tf.matmul(z, fc1_weights) + fc1_biases)

    fc2_units = 784
    fc2_weights = tf.get_variable('gen_fc2_weights', (fc1_units, fc2_units), dtype=tf.float32, initializer=weight_init)
    fc2_biases = tf.get_variable('gen_fc2_biases', (fc2_units), initializer=bias_init)
    fc2 = activation(tf.matmul(fc1, fc2_weights) + fc2_biases)

    output = tf.nn.sigmoid(fc2, name='gen_sigmoid_output')
    return output

def discriminator(X, activation=tf.nn.relu):
    shape = z.get_shape().as_list()
    weight_init = tf.random_normal_initializer(stddev=0.1)
    bias_init = tf.constant_initializer(0.0)

    with tf.variable_scope('discriminator', reuse=tf.AUTO_REUSE):

        fc1_units = 1024
        fc1_weights = tf.get_variable('dis_fc1_weights', (shape[1], fc1_units), dtype=tf.float32, initializer=weight_init)
        fc1_biases = tf.get_variable('dis_fc1_biases', (fc1_units), initializer=bias_init)
        fc1 = activation(tf.matmul(z, fc1_weights) + fc1_biases)

        fc2_units = 1
        fc2_weights = tf.get_variable('dis_fc2_weights', (fc1_units, fc2_units), dtype=tf.float32, initializer=weight_init)
        fc2_biases = tf.get_variable('dis_fc2_biases', (fc2_units), initializer=bias_init)
        fc2 = tf.matmul(fc1, fc2_weights) + fc2_biases

        sigmoid_out = tf.nn.sigmoid(fc2, name='dis_sigmoid_output')
    return sigmoid_out, fc2

X = tf.placeholder(tf.float32, shape=(real_batch_size, 28*28), name='X')
z = tf.placeholder(dtype=tf.float32, shape=(fake_batch_size, 100), name='z')

gen = generator(z)
dis_real, dis_real_logits = discriminator(X)
dis_fake, dis_fake_logits = discriminator(gen)


dis_real_loss = tf.reduce_mean(
                    tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(dis_real), logits=dis_real_logits))
dis_fake_loss = tf.reduce_mean(
                    tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.zeros_like(dis_fake), logits=dis_fake_logits))
dis_loss = dis_real_loss + dis_fake_loss
gen_loss = tf.reduce_mean(
                    tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(dis_fake), logits=dis_fake_logits))

train_vars = tf.trainable_variables()
dis_vars = [var for var in train_vars if 'dis_' in var.name]
gen_vars = [var for var in train_vars if 'gen_' in var.name]

dis_optimize = tf.train.AdamOptimizer(learning_rate, beta1).minimize(dis_loss, var_list=dis_vars)
gen_optimize = tf.train.AdamOptimizer(learning_rate, beta1).minimize(gen_loss, var_list=gen_vars)

Answer 1

您的问题在于如何过滤变量：

dis_vars = [var for var in train_vars if 'dis_' in var.name]
gen_vars = [var for var in train_vars if 'gen_' in var.name]

您在discriminator 范围内定义了鉴别器的变量，在/（无前缀）范围内定义了生成器的变量，因此您的过滤器只是过滤掉了每个变量。

您可以收集正确过滤的变量：

dis_vars = [var for var in train_vars if 'discriminator' in var.name]
gen_vars = [var for var in train_vars if 'discriminator' not in var.name]

此外，错误状态“没有为任何变量提供梯度，请检查您的图表中不支持梯度的操作，在变量 [...] 和损失 Tensor("Mean_2:0", shape=(), dtype =float32)"

其实问题出在损失张量上。

损失张量是输入为生成器输出时判别器的评估。判别器定义是错误的，实际上，您指的是一个从未定义过的变量 z。 因此，如果您使用 X 而不是 z 更新您的鉴别器代码，它会起作用：

import tensorflow as tf


def generator(z, activation=tf.nn.relu, reuse=False):
    with tf.variable_scope("generator"):
        shape = z.get_shape().as_list()
        weight_init = tf.random_normal_initializer(stddev=0.1)
        bias_init = tf.constant_initializer(0.0)

        fc1_units = 256
        fc1_weights = tf.get_variable(
            'gen_fc1_weights', (shape[1], fc1_units),
            dtype=tf.float32,
            initializer=weight_init)
        fc1_biases = tf.get_variable(
            'gen_fc1_biases', (fc1_units), initializer=bias_init)
        fc1 = activation(tf.matmul(z, fc1_weights) + fc1_biases)

        fc2_units = 784
        fc2_weights = tf.get_variable(
            'gen_fc2_weights', (fc1_units, fc2_units),
            dtype=tf.float32,
            initializer=weight_init)
        fc2_biases = tf.get_variable(
            'gen_fc2_biases', (fc2_units), initializer=bias_init)
        fc2 = activation(tf.matmul(fc1, fc2_weights) + fc2_biases)

        output = tf.nn.sigmoid(fc2, name='gen_sigmoid_output')
        return output


def discriminator(X, activation=tf.nn.relu):
    with tf.variable_scope('discriminator', reuse=tf.AUTO_REUSE):
        shape = X.get_shape().as_list()
        weight_init = tf.random_normal_initializer(stddev=0.1)
        bias_init = tf.constant_initializer(0.0)

        fc1_units = 1024
        fc1_weights = tf.get_variable(
            'dis_fc1_weights', (shape[1], fc1_units),
            dtype=tf.float32,
            initializer=weight_init)
        fc1_biases = tf.get_variable(
            'dis_fc1_biases', (fc1_units), initializer=bias_init)
        fc1 = activation(tf.matmul(X, fc1_weights) + fc1_biases)

        fc2_units = 1
        fc2_weights = tf.get_variable(
            'dis_fc2_weights', (fc1_units, fc2_units),
            dtype=tf.float32,
            initializer=weight_init)
        fc2_biases = tf.get_variable(
            'dis_fc2_biases', (fc2_units), initializer=bias_init)
        fc2 = tf.matmul(fc1, fc2_weights) + fc2_biases

        return fc2


### ADDED TO TEST
real_batch_size, fake_batch_size = 10, 10
learning_rate = 1e-5
beta1 = 0.5
###
X = tf.placeholder(tf.float32, shape=(real_batch_size, 28 * 28), name='X')
z = tf.placeholder(dtype=tf.float32, shape=(fake_batch_size, 100), name='z')

gen = generator(z)
dis_real_logits = discriminator(X)
dis_fake_logits = discriminator(gen)

dis_real_loss = tf.reduce_mean(
    tf.nn.sigmoid_cross_entropy_with_logits(
        labels=tf.ones_like(dis_real_logits), logits=dis_real_logits))
dis_fake_loss = tf.reduce_mean(
    tf.nn.sigmoid_cross_entropy_with_logits(
        labels=tf.zeros_like(dis_fake_logits), logits=dis_fake_logits))

dis_loss = dis_real_loss + dis_fake_loss

gen_loss = tf.reduce_mean(
    tf.nn.sigmoid_cross_entropy_with_logits(
        labels=tf.ones_like(dis_fake_logits), logits=dis_fake_logits))

train_vars = tf.trainable_variables()
dis_vars = [var for var in train_vars if 'dis_' in var.name]
gen_vars = [var for var in train_vars if 'gen_' in var.name]

dis_optimize = tf.train.AdamOptimizer(learning_rate, beta1).minimize(
    dis_loss, var_list=dis_vars)
gen_optimize = tf.train.AdamOptimizer(learning_rate, beta1).minimize(
    gen_loss, var_list=gen_vars)