【一起撸个深度学习框架】6 折与曲的相会——激活函数(2)

# example/01_esay/04_relu与二次拟合.py
import sys
sys.path.append('../..')  # 父目录的父目录
from ourdl.core import Varrible
from ourdl.ops import Mul, Add
from ourdl.ops.loss import ValueLoss
from ourdl.ops import LeakyRelu as Relu
import matplotlib.pyplot as plt
import numpy as np
import random
# 1.1 线性变换一
x = Varrible()
w_11 = Varrible()
w_12 = Varrible()
mul_11 = Mul([x, w_11])
mul_12 = Mul([x, w_12])
b_11 = Varrible()
b_12 = Varrible()
add_11 = Add([mul_11, b_11])
add_12 = Add([mul_12, b_12])
# 1.2 激活函数 --> 非线性变换
relu_11 = Relu([add_11])
relu_12 = Relu([add_12])
# 1.3 线性变换二
w_21 = Varrible()
w_22 = Varrible()
mul_21 = Mul([relu_11, w_21])
mul_22 = Mul([relu_12, w_22])
b_21 = Varrible()
add_21 = Add([mul_21, mul_22, b_21])
# 1.4 损失函数
label = Varrible()
loss = ValueLoss([label, add_21])

values = [-0.1571950013426796, -0.1070365984042347, 0.3791639008324807, 0.31960284774415215, 0.4263410176300597, 0.5097967360623379, 0.7597168751185974]
for i in range(len(params)):
    params[i].set_value(values[i])

# example/01_esay/04_relu与二次拟合.py
# 3 生成数据
data_x = [random.uniform(0, 2) for i in range(1500)]  # 似乎实数比离散的[0, 1, 2]要好
data_label = [x * x for x in data_x]

# example/01_esay/04_relu与二次拟合.py
# 4 开始训练
losses = []
for i in range(len(data_x)):
    x.set_value(data_x[i])
    label.set_value(data_label[i])
    loss.forward()
    for param in params:
        param.get_grad()
        param.update(lr=0.01)
    if i % 100 == 0:
        print(f'[{i}]:loss={loss.value},', [param.value for param in params])
    losses.append(loss.value)
    loss.clear()
# 5 画出训练过程中loss的变化曲线
show_x = [i for i in range(len(losses))]
show_y = [_ for _ in losses]
plt.plot(show_x, show_y)
plt.show()

# example/01_esay/04_relu与二次拟合.py
# 4 开始训练
# 4.1 创建画布
fig = plt.figure(figsize=(15,4))
ax = fig.subplots(1,3,sharex=True,sharey=False)  # ax是包含一行三列，一共三块子画布的列表
# 4.2 训练，同时绘制动画
losses = []
for i in range(len(data_x)):
    x.set_value(data_x[i])
    label.set_value(data_label[i])
    loss.forward()
    for param in params:
        param.get_grad()
        param.update(lr=0.01)
    if i % 200 == 0:
        print(f'[{i}]:loss={loss.value},', [param.value for param in params])
    losses.append(loss.value)
    loss.clear()
    if i % 40 == 0:
        show_ax_mul(ax)  # 用于绘制多图动画

# example/01_esay/04_relu与二次拟合.py
def show_ax_mul(ax):
    '''
    用于绘制多图动画\n
  '''
    # 1 画真实的二次函数曲线
    show_x = np.linspace(-0.5, 2.5, 30, endpoint=True)
    show_y = [x_one * x_one for x_one in show_x]
    ax[2].plot(show_x, show_y)
    # 2 画模型拟合的曲线
    show_x = np.linspace(-0.5, 2.5, 10, endpoint=True)
    shows = {'add1':[], 'add2':[], 'mul1':[], 'mul2':[], 'y':[]}
    for x_one in show_x:
        x.set_value(x_one)
        add_21.clear()
        add_21.forward()
        shows['y'].append(add_21.value)
        shows['add1'].append(add_11.value)
        shows['add2'].append(add_12.value)
        shows['mul1'].append(mul_21.value)
        shows['mul2'].append(mul_22.value)
    ax[2].plot(show_x, shows['y'])
    ax[0].plot(show_x, shows['add1'])
    ax[0].plot(show_x, shows['add2'])
    ax[1].plot(show_x, shows['mul1'])
    ax[1].plot(show_x, shows['mul2'])
    y_0 = [0 for _ in show_x]  # 水平的参考线
    ax[2].plot(show_x, y_0)
    ax[0].plot(show_x, y_0)
    ax[1].plot(show_x, y_0)
    plt.pause(0.01)
    for i in range(ax.shape[0]):
        ax[i].cla()