Question

我尝试使用theta作为下面显示的代码，仅训练一个Tensorflow变量tf.select。

import numpy as np
import tensorflow as tf
from numpy.random import RandomState
from tensorflow.contrib.losses import mean_squared_error

tf.logging.set_verbosity(tf.logging.INFO)

random_state = RandomState()


def generate_data(n_train, n_test, theta=0.2):
    X_train = random_state.uniform(size=[n_train])
    X_test = random_state.uniform(size=[n_test])
    y_train = np.where((X_train > theta), np.sin(X_train), np.cos(X_train))
    y_test = np.where((X_test > theta), np.sin(X_test), np.cos(X_test))
    return (X_train, y_train), (X_test, y_test)


(X_train, y_train), (X_test, y_test) = generate_data(1000, 300)

intN = np.int64
floatN = np.float32
intT = tf.int64
floatT = tf.float32

x = tf.placeholder(dtype=floatT, name=r'x', shape=[None])
y_true = tf.placeholder(dtype=floatT, name=r'y_true', shape=[None])
theta = tf.get_variable(dtype=floatT, name=r'theta', initializer=tf.zeros_initializer(shape=[], dtype=floatT))

y_pred = tf.select(
    tf.greater(x, theta),
    tf.sin(x),
    tf.cos(x), name=r'y_pred')

loss = mean_squared_error(y_pred, y_true)
optimizer = tf.train.AdamOptimizer(0.01).minimize(loss)

with tf.Session() as session:
    all_variables = tf.global_variables() + tf.local_variables()
    all_variables_initializer = tf.variables_initializer(all_variables)
    session.run(all_variables_initializer)

    _, batch_loss = session.run([optimizer, loss], feed_dict={
        x: X_train,
        y_true: y_train,
    })

    tf.logging.info(r'batch loss {:.4f}'.format(batch_loss))

但我得到了

ValueError：没有为任何变量提供渐变，检查图表中不支持渐变的ops，变量之间[＆＃39; Tensor（＆＃34; theta / read：0＆＃34;，shape =（）， dtype = float32）＆＃39;]和丢失Tensor（＆＃34; mean_squared_error / value：0＆＃34;，shape =（），dtype = float32）。

那怎么办呢？如果我真的想使用tf.cond或tf.where或tf.select或tf.case而不是像sigmoid这样的话，我的意思是如何做条件分支（分段函数） Tensorflow一般吗？

如何在Tensorflow中编写分段函数？

0 个答案: