GTX950 4G 4核
使用的是第一个
观察到全程占用率只有6%而已。
哈哈，程序还是太简单了

# coding=utf-8
import numpy as np
import tensorflow as tf
import os

save_path ="modules/xaddy.ckpt"

train_times = 100000

# 获取训练数据和测试数据
def get_data(number):
    list_x = []
    list_y = []
    for i in range(number):
        x = np.random.randn(1)
        # 这里构建的数据的分布满足 y=2*x+10
        y = x ** 2 + 2 * x + 10 + np.random.randn(1) * 0.01
        list_x.append(x)
        list_y.append(y)
    return list_x, list_y

#在下面这个with里的变量都会使用gpu0来训练
with tf.device("/gpu:0"):
    def inference(x):
        global weight,bias
        weight = tf.get_variable("weight", [2])
        bias = tf.get_variable("bise", [1])
        y = x ** 2 * weight[0] + x * weight[1] + bias
        return y

train_x = tf.placeholder(tf.float32)
train_lable = tf.placeholder(tf.float32)

test_x = tf.placeholder(tf.float32)
test_lable = tf.placeholder(tf.float32)

with tf.variable_scope("inference"):
    train_y = inference(train_x)
    # 在此处定义相同名字的变量是共享变量
    # 此句之后的tf.get_variable获取的变量是需要根据变量的名字共享前面已经定义的变量
    # 如果之前没有相同名字的变量则会报错
    tf.get_variable_scope().reuse_variables()
    test_y = inference(test_x)

train_loss = tf.square(train_y - train_lable)
test_loss = tf.square(test_y - test_lable)
opt = tf.train.GradientDescentOptimizer(0.002)
train_op = opt.minimize(train_loss)

init = tf.global_variables_initializer()

train_data_x, train_data_lable = get_data(train_times)
test_data_x, test_data_lable = get_data(1)

saver = tf.train.Saver()

with tf.Session() as sess:
    sess.run(init)
    if os.path.exists(save_path):
        saver.restore(sess, save_path)
    for i in range(train_times):
        sess.run(train_op, feed_dict={train_x: train_data_x[i], train_lable: train_data_lable[i]})

        if i % 10 == 0:
            test_loss_value = sess.run(test_loss, feed_dict={test_x: test_data_x[0], test_lable: test_data_lable[0]})
            print("step %d eval loss is %.3f" % (i, test_loss_value))
            print(sess.run(weight))
    saver.save(sess, save_path)