在 tensorflow 1.15 中如何正确传递权重和偏差？

Question

我正在尝试使用 tensorflow 1.15 实现 U_NET 架构，这是第一个卷积层：

import tensorflow as tf

print("############################### VERSION TENSORFLOW ###############################################")
print(tf.__version__)
print("############################### VERSION TENSORFLOW ###############################################")
def u_net_model(feature):

    w_init = tf.truncated_normal_initializer(stddev=0.01)

    print("--------------------------------------------------------------------------------- w_init")
    print(w_init)
    b_init = tf.constant_initializer(value=0.40)
    gamma_init = tf.random_normal_initializer(1., 0.02)

    with tf.variable_scope("u_network",reuse=True):
        x = tf.keras.Input(batch_size = 5,tensor=feature)
        #y = tf.keras.layers.Dense(16, activation='softmax')(x)


        conv1 = tf.keras.layers.Conv2D(64,4,(2,2),activation = 'relu',padding='same',kernel_initializer= w_init,bias_initializer=b_init, name = "convolution1")(x)
        print("conv1")
        print(conv1)

        conv2 = tf.keras.layers.Conv2D(128,4,(2,2),activation = 'relu',padding='same', kernel_initializer= w_init,bias_initializer=b_init, name = "convolution2")(conv1)
        print("conv2")
        print(conv2)
        conv2 = tf.keras.layers.BatchNormalization()(conv2)
        print("conv2")
        print(conv2)

在 main.py 我有：

nw, nh, nz = X_train.shape[1:]

t_image_good = tf.placeholder('float32', [25, nw, nh, nz], name='good_image')
print(t_image_good)

t_image_good_samples = tf.placeholder('float32', [50, nw, nh, nz], name='good_image_samples')
print(t_image_good_samples)

t_PROVA = t_image_good

t_PROVA_samples = t_image_good_samples

g_nmse_a = tf.sqrt(tf.reduce_sum(tf.squared_difference(t_PROVA, t_PROVA), axis=[1, 2, 3]))
g_nmse_b = tf.sqrt(tf.reduce_sum(tf.square(t_PROVA), axis=[1, 2, 3]))
g_nmse = tf.reduce_mean(g_nmse_a / g_nmse_b)

generator_loss  = g_alpha *g_nmse 
print("generator_loss")
#geneator_loss è un tensore
print(generator_loss)
learning_rate = 0.0001
beta = 0.5




print("\n")
generator_variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,'u_network')

print("--------------------------------------- generator_variables")
print(generator_variables)
generator_gradient_optimum = tf.train.AdamOptimizer(learning_rate, beta1=beta).minimize(generator_loss, var_list = generator_variables )




nw, nh, nz = X_train.shape[1:]

t_image_good = tf.placeholder('float32', [25, nw, nh, nz], name='good_image')
print(t_image_good)

t_image_good_samples = tf.placeholder('float32', [50, nw, nh, nz], name='good_image_samples')
print(t_image_good_samples)

t_PROVA = t_image_good

t_PROVA_samples = t_image_good_samples

g_nmse_a = tf.sqrt(tf.reduce_sum(tf.squared_difference(t_PROVA, t_PROVA), axis=[1, 2, 3]))
g_nmse_b = tf.sqrt(tf.reduce_sum(tf.square(t_PROVA), axis=[1, 2, 3]))
g_nmse = tf.reduce_mean(g_nmse_a / g_nmse_b)

generator_loss  = g_alpha *g_nmse 
print("generator_loss")
#geneator_loss è un tensore
print(generator_loss)
learning_rate = 0.0001
beta = 0.5




print("\n")
generator_variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,'u_network')

print("--------------------------------------- generator_variables")
print(generator_variables)
generator_gradient_optimum = tf.train.AdamOptimizer(learning_rate, beta1=beta).minimize(generator_loss, var_list = generator_variables )

当我运行它时，我得到：

ValueError: No gradients provided for any variable, check your graph for ops that do not support gradients, between variables ["<tf.Variable 'u_network/convolution1/kernel:0' shape=(4, 4, 1, 64) dtype=float32>", "<tf.Variable 'u_network/convolution1/bias:0' shape=(64,) dtype=float32>", "<tf.Variable 'u_network/convolution2/kernel:0' shape=(4, 4, 64, 128) dtype=float32>", "<tf.Variable 'u_network/convolution2/bias:0' shape=(128,) dtype=float32>", "<tf.Variable 'u_network/batch_normalization/gamma:0' shape=(128,) dtype=float32>", "<tf.Variable 'u_network/batch_normalization/beta:0' shape=(128,) dtype=float32>", "<tf.Variable 'u_network/convolution3/kernel:0' shape=(4, 4, 128, 256) dtype=float32>", "<tf.Variable 'u_network/convolution3/bias:0' shape=(256,) dtype=float32>", "<tf.Variable 'u_network/batch_normalization_1/gamma:0' shape=(256,) dtype=float32>"

...这种类型的许多行，最后以：

 and loss Tensor("mul_10:0", shape=(), dtype=float32).

我要做的是传递参数、权重和偏差，以便启动 AdamOptimizer。

我究竟做错了什么？

Answer 1

在您提供的代码中，您没有在哪里调用 u_net_model 的u_net_model 。 您提供的代码在图中只有几个占位符，并在其上执行了一些操作。 您使用的操作是tf.square和tf.squared_difference ，它们没有任何可学习的参数，因此优化器没有什么可以最小化（或收敛）的。