tensorFlow 神经网络2

learnrate 太大容易跑飞，设置激活函数 可以一定程度上增加learnrate，不跑飞

self.saver = tf.train.Saver() 和 self.init_variable = tf.global_variables_initializer()   self.sess.run(self.init_variable)一定要写在所有变量的最后

GradientDescentOptimizer优化器可以继续训练 AdadeltaOptimizer优化器的训练结果，反之则不行

# -*- coding: utf-8 -*-

import numpy as np
import tensorflow as tf
from tensorflow.python.framework import ops

is_continue = True
path_prefix = "../../../"


class neural_network:
    def __init__(self):
        ops.reset_default_graph()
        self.model_path = path_prefix + "data/model/model"
        self.batch_size = 2000
        self.learning_rate = 0.5
        self.input_size = 15193
        self.output_size = 1
        self.sess = tf.InteractiveSession()
        self.input = tf.placeholder(tf.float32, [None, self.input_size])  # 15193个维度
        self.label = tf.placeholder(tf.float32, [None, self.output_size])
        self.add_layers(self.input_size, self.output_size, tf.nn.relu)
        self.loss = tf.reduce_mean(tf.reduce_sum(tf.square(self.output - self.label),
                                                 reduction_indices=[1]))

        self.optimizer = tf.train.AdadeltaOptimizer(learning_rate=self.learning_rate).minimize(loss=self.loss)
        self.saver = tf.train.Saver()
        if is_continue:
            self.saver.restore(self.sess, self.model_path)
        else:
            self.init_variable = tf.global_variables_initializer()
            self.sess.run(self.init_variable)
        self.data_generator = self.data_generator_f(
            path_prefix + ("data/trainData_local"),
            50000)

    def add_layers(self, input_size, output_size, activation_function=None):
        layer1_size = 200
        weight1 = tf.Variable(tf.zeros([input_size, layer1_size]) + 0.1)
        biaes1 = tf.Variable(tf.zeros([1, layer1_size]) + 0.1)
        output_layer1 = tf.matmul(self.input, weight1) + biaes1  # 200 50
        input_layer2 = output_layer1
        if activation_function is not None:
            input_layer2 = activation_function(output_layer1)

        layer2_size = 50
        weight2 = tf.Variable(tf.zeros([layer1_size, layer2_size]) + 0.01)
        biaes2 = tf.Variable(tf.zeros([1, layer2_size]) + 0.1)
        output_layer2 = tf.matmul(input_layer2, weight2) + biaes2  # 200 10
        input_layer3 = output_layer2
        if activation_function is not None:
            input_layer3 = activation_function(output_layer2)

        weight3 = tf.Variable(tf.zeros([layer2_size, self.output_size]) + 0.01)
        biaes3 = tf.Variable(tf.zeros([1, self.output_size]) + 0.1)
        output_layer3 = tf.matmul(input_layer3, weight3) + biaes3

        if activation_function is not None:
            self.output = activation_function(output_layer3)
        else:
            self.output = output_layer3

        return self.output

    def train(self):
        for i in range(0, 20000000):
            labels, feature = self.data_generator.next()
            self.sess.run(self.optimizer, feed_dict={
                self.input: feature,
                self.label: np.asarray(labels).reshape(self.batch_size, 1)
            })

            if (i % 10) == 0:
                print(self.sess.run(self.loss, feed_dict={
                    self.input: feature,
                    self.label: np.asarray(labels).reshape(self.batch_size, 1)
                }))
                if (i % 100) == 0:
                    self.saver.save(self.sess, self.model_path)
                    print("save complete... ...")

    def predict(self, feature):
        print "--start predict--"
        return self.sess.run(self.output, feed_dict={
            self.input: feature
        })

    def data_generator_f(self, input_dir, count):
        input_filename = input_dir
        while count > 0:
            labels = np.zeros(self.batch_size)
            rets = np.empty(shape=[self.batch_size, 15193])
            i = 0
            for line in open(input_filename, "r"):
                # print "trainData", line
                data = line.split(" ")
                label = int(float(data[0]))
                ids = []
                values = []
                for fea in data[1:]:
                    id, value = fea.split(":")
                    ids.append(int(id))
                    values.append(float(value))
                ret = np.zeros([1, 15193])
                for (index, d) in zip(ids, values):
                    ret[0][index] = d
                labels[i] = int(label)
                rets[i] = ret
                i += 1
                if i > self.batch_size - 1:
                    i = 0
                    yield labels, rets
            print("train count:", count)
            count -= 1


if __name__ == '__main__':
    print('start run... ...')
    trainer = neural_network()
    trainer.train()
    """
    trainer.batch_size = 100
    data = trainer.data_generator_f("/home/panteng/下载/20171231.libsvm", 3)
    threshold = 0.55

    for i in range(0, 100):
        labels, features = data.next()
        pre = trainer.predict(feature=features)
        print ("P:", np.sum(np.array([(labels > threshold)])), " N:", np.sum(np.array([(labels < threshold)])))

        print("accuracy:", np.sum(
            (np.array([(labels > threshold)]).reshape(trainer.batch_size, 1) == (pre > threshold)).reshape(1, trainer.batch_size)
        ) * 1.0 / trainer.batch_size)
        print("precision:",
              np.sum((np.array([(labels > threshold)]).reshape(trainer.batch_size, 1) == (pre > threshold)) & (
                  pre > threshold)) * 1.0 / np.sum(np.array([(pre > threshold)]))
              )
        print("recall:",
              np.sum((np.array([(labels > threshold)]).reshape(trainer.batch_size, 1) == (pre > threshold)) & (
                  pre > threshold)) * 1.0 / np.sum(np.array([(labels > threshold)]))
              )
    """

 

2：

# coding:utf-8
import tensorflow as tf
import numpy as np

tf.app.flags.DEFINE_string('train_dir', '')
tf.app.flags.DEFINE_integer('capacity', 30000, 'indicates training epoch')
FLAGS = tf.app.flags.FLAGS

features = tf.placeholder(tf.float32, shape=[None, 151])
label = tf.placeholder(tf.float32, shape=[None, 1])


def read_csv():
    file_queue = tf.train.string_input_producer(tf.train.match_filenames_once(FLAGS.train_dir + "*"))
    reader = tf.TextLineReader(skip_header_lines=1)
    key, value = reader.read(file_queue)
    data = tf.decode_csv(value, record_defaults=[[0.0] for _ in range(152)], field_delim="\t")
    return tf.train.shuffle_batch(data, batch_size=1024, capacity=FLAGS.capacity,
                                  min_after_dequeue=FLAGS.capacity / 2, num_threads=8)


def build_net():
    deep_layer1 = tf.layers.dense(inputs=features, units=80, activation=tf.nn.relu)
    deep_layer2 = tf.layers.dense(inputs=deep_layer1, units=32, activation=tf.nn.relu)
    deep_layer3 = tf.layers.dense(inputs=deep_layer2, units=8, activation=tf.nn.relu)
    deep_layer4 = tf.layers.dense(inputs=deep_layer3, units=5, activation=tf.nn.relu)
    wide_layer1 = tf.layers.dense(inputs=features, units=1, activation=tf.nn.sigmoid)

    wide_deep_layer = tf.concat([wide_layer1, deep_layer4], axis=1)
    return tf.layers.dense(inputs=wide_deep_layer, units=1, activation=tf.nn.sigmoid)


with tf.Session() as sess:
    output = build_net()
    loss = tf.reduce_mean(tf.reduce_sum(tf.squared_difference(output, label)))
    optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)

    train_data = read_csv()
    init = (tf.global_variables_initializer(), tf.local_variables_initializer())
    sess.run(init)
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)
    data_set = sess.run(train_data)
    label_in = np.array(data_set[0], dtype='int').ravel()
    feature_in = np.transpose(data_set[1:])
    for i in range(100):
        sess.run(optimizer,
                 feed_dict={features: np.array(feature_in).reshape([-1, 151]),
                            label: np.array(label_in).reshape([-1, 1])})
        print sess.run(loss,
                       feed_dict={features: np.array(feature_in).reshape([-1, 151]),
                                  label: np.array(label_in).reshape([-1, 1])})