2-sgd

随机梯度下降

导入包，数据

%matplotlib inline
from fastai.basics import *

n=100
x = torch.ones(n,2) 
x[:,0].uniform_(-1.,1)
x[:5]

tensor([[ 0.9955,  1.0000],
        [-0.0753,  1.0000],
        [ 0.4918,  1.0000],
        [-0.0823,  1.0000],
        [ 0.7418,  1.0000]])

a = tensor(3.,2); a

tensor([3., 2.])

• @运算符做内积，a被转置

y = x@a + torch.rand(n)

x[:,0]

tensor([ 0.9955, -0.0753,  0.4918, -0.0823,  0.7418,  0.5482,  0.0317, -0.0224,
        -0.7323,  0.6282,  0.3766, -0.4949, -0.1994, -0.9810,  0.6939,  0.1182,
         0.6041, -0.7308, -0.5721,  0.3003, -0.4328, -0.1355, -0.2422, -0.6378,
         0.4063, -0.7544, -0.3286,  0.5062,  0.8027, -0.2050, -0.8538,  0.5771,
        -0.4250,  0.3671,  0.3434,  0.4788,  0.9859, -0.7919,  0.1072, -0.8707,
        -0.7484, -0.1305,  0.4626,  0.5997,  0.1537, -0.5728,  0.8755,  0.9620,
         0.2826,  0.0041,  0.5956, -0.5758,  0.1405,  0.3512, -0.1403,  0.4057,
         0.4607, -0.7323, -0.5207,  0.6755,  0.1890,  0.2442, -0.2409,  0.9330,
        -0.3626,  0.1813,  0.3186, -0.2791,  0.2366,  0.0091, -0.1328,  0.9189,
        -0.4149,  0.7616,  0.7230,  0.8930, -0.5025, -0.6394, -0.8798,  0.9016,
         0.9926,  0.0196, -0.8414,  0.8805, -0.3590, -0.1537, -0.7058, -0.9996,
        -0.8218,  0.4228,  0.6947, -0.4348,  0.4422,  0.6000,  0.4775,  0.8744,
        -0.6983, -0.4930,  0.0187, -0.0805])

plt.scatter(x[:,0], y);

构造损失函数

def mse(y_hat, y): return ((y_hat-y)**2).mean()

a = tensor(-1.,1)
y_hat = x@a
mse(y_hat, y)

tensor(8.2766)

plt.scatter(x[:,0],y)
plt.scatter(x[:,0],y_hat);

训练

a = nn.Parameter(a); a#开启梯度

Parameter containing:
tensor([-1.,  1.], requires_grad=True)

def update():
    y_hat = x@a 
    loss = mse(y,y_hat)
    if t % 10 == 0: print(loss)
    loss.backward()
    with torch.no_grad():
        a.sub_(lr * a.grad)
        a.grad.zero_()

lr = 1e-1
for t in range(100): update()

tensor(8.2766, grad_fn=<MeanBackward0>)
tensor(1.3620, grad_fn=<MeanBackward0>)
tensor(0.4088, grad_fn=<MeanBackward0>)
tensor(0.1706, grad_fn=<MeanBackward0>)
tensor(0.1088, grad_fn=<MeanBackward0>)
tensor(0.0928, grad_fn=<MeanBackward0>)
tensor(0.0886, grad_fn=<MeanBackward0>)
tensor(0.0875, grad_fn=<MeanBackward0>)
tensor(0.0873, grad_fn=<MeanBackward0>)
tensor(0.0872, grad_fn=<MeanBackward0>)

• 这里关闭了自动求梯度，不然画图会报错

plt.scatter(x[:,0],y)

a.requires_grad_(False)
plt.scatter(x[:,0],x@a);

动画

from matplotlib import animation, rc
rc('animation', html='jshtml')

• 源代码使用的是‘a’直接画图，但这里不知道为什么报错（根据报错的信息，应该是自动求梯度的问题）
• 这里使用temp=a.data存储a的值，不带梯度的
• 注意：不能temp=a，这样temp只是a的副本，还是带梯度的，即使使用命令temp.requires_grad_(False),那么a的求梯度也会关闭

a = nn.Parameter(tensor(-1.,1))
temp = a.data
fig = plt.figure()
plt.scatter(x[:,0], y, c='orange')
line, = plt.plot(x[:,0], x@temp)
plt.close()

def animate(i):
    update()
    temp = a.data
    line.set_ydata(x@temp)
    return line,

animation.FuncAnimation(fig, animate, np.arange(0, 100), interval=20)