Pytorch基础练习

Pytorch基础入门练习

import 导入

import torch # 基本的torch函数
import torch.autograd as autograd # 自动求导
import torch.nn as nn # 神经网络类都在这个里面
import torch.nn.functional as F # 几乎所有的激励函数
import torch.optim as optim # 优化

创建Tensors

# create 1D vector
v_data = [1., 2., 3.]
v = torch.tensor(v_data)
print(v)

# create 2D vector
M_data = [[1., 2., 3.],[4., 5., 6.]]
M = torch.tensor(M_data)
print(M)

# create 3D vector
T_data = [[[1., 2.],[3., 4.]],[[5., 6.],[7., 8.]]]
T = torch.tensor(T_data)
print(T)
var foo = 'bar';

下面展示一些 运行结果。

tensor([1., 2., 3.])
tensor([[1., 2., 3.],
        [4., 5., 6.]])
tensor([[[1., 2.],
         [3., 4.]],

        [[5., 6.],
         [7., 8.]]])

获取Tensor的部分值

# 获取tensor的部分值
# 这里要比tensorflow 好用多了QAQ
print(v[0])
print(M[0])
print(T[0])

下面展示 运行结果。

tensor(1.)
tensor([1., 2., 3.])
tensor([[1., 2.],
        [3., 4.]])

产生随机数

# 随机数
x = torch.randn((3, 4, 5))
print(x)

运行结果如下：

tensor([[[-0.7559, -0.1650, -0.2283,  0.7088, -0.1870],
         [ 0.0914,  0.4764,  1.2566, -0.6386, -0.0722],
         [-0.8834, -0.9981, -1.5278, -0.5788,  0.9297],
         [-0.0450, -1.7502,  0.9460, -1.3609,  1.3793]],

        [[ 1.1496,  1.4190, -0.3966,  0.8439, -0.2627],
         [ 0.5515,  0.2070, -0.8803, -1.5191,  0.7596],
         [-1.0275, -1.1738,  2.0765,  2.5308, -1.0748],
         [-1.6574, -1.1648, -1.9078, -0.5660,  0.5558]],

        [[-1.8764, -0.8884,  1.2439, -2.0752,  0.2381],
         [ 1.2270,  2.7735,  0.0249, -0.4611, -0.9217],
         [ 2.3551, -0.4531,  0.2677,  0.4329,  1.7990],
         [ 2.3443, -1.5808, -1.5181, -0.2277, -0.1178]]])

Tensor运算

a = torch.tensor([1, 2, 3])
b = torch.tensor([4, 5, 6])
c = a+b
print(c)

运行结果如下：

tensor([5, 7, 9])

[res] torch.cat( [res,] x_1, x_2, [dimension] )：

x_1 = torch.randn(2, 5)
y_1 = torch.randn(3, 5)
z_1 =torch.cat([x_1, y_1])#没有最后一个参数，默认是0，则最终维度的第0维度为x_1与y_1第0维度的和，最终维度的其他维度不变.以下同理
print(z_1)

x_2 = torch.randn(2, 3)
y_2 = torch.randn(2, 5)
z_2 = torch.cat([x_2, y_2], 1)
print(z_2)

运行结果如下：

tensor([[ 0.9761,  0.3068,  2.0089,  1.5800, -1.5596],
        [-0.5638, -1.6565,  0.3520, -1.8214,  1.3104],
        [ 2.3029,  2.6305,  1.2756,  0.3129, -1.1587],
        [ 0.2029,  1.6141,  0.2416,  1.7751, -0.0401],
        [ 0.6935, -0.8595, -0.0923,  0.6760,  1.7727]])
tensor([[-0.7705, -1.4831,  0.7549, -0.3064, -1.2749, -1.1668, -0.2255, -0.8785],
        [-0.1929,  0.6399, -0.6258,  0.4568, -0.6885, -0.2360, -0.1745,  0.7706]])

Tensor维度变形reshaping

`print(x.view(2,12)) # 将234 -> 2*12

# Tensor（reshaping）
x = torch.randn(2, 3, 4)
print(x)
print(x.view(2, -1))#-1的话，意味着最后的相乘为维数，这里为2*之后的成绩

下面展示 运行结果。

tensor([[[-0.1259, -0.2706, -0.3104,  1.3265],
         [-0.5548,  1.4568,  0.6568, -1.3463],
         [ 0.1177, -0.5917, -1.0518, -1.5515]],

        [[-1.3372, -0.9601,  0.5557, -0.7694],
         [-0.3657,  1.0275, -1.2712, -2.0646],
         [-0.7945, -0.9716,  0.0876, -2.8072]]])
tensor([[-0.1259, -0.2706, -0.3104,  1.3265, -0.5548,  1.4568,  0.6568, -1.3463,
          0.1177, -0.5917, -1.0518, -1.5515],
        [-1.3372, -0.9601,  0.5557, -0.7694, -0.3657,  1.0275, -1.2712, -2.0646,
         -0.7945, -0.9716,  0.0876, -2.8072]])

计算图和自动微分

x = autograd.Variable(torch.Tensor([1., 2., 3]), requires_grad=True)
print(x)
print(x.data)#.data显示具体数据
y = autograd.Variable( torch.Tensor([4., 5., 6]), requires_grad=True )
z = x + y
print(z.data)

下面展示 运行结果。

tensor([1., 2., 3.], requires_grad=True)
tensor([1., 2., 3.])
tensor([5., 7., 9.])

# 显示z中所有元素之和
s = z.sum()
print(s)
s.backward() # 反向传播
print(x.grad) # 对x求导

下面展示 运行结果。

tensor(21., grad_fn=<SumBackward0>)
tensor([1., 1., 1.])
#答案解释
#x = [1,2,3]
#y = [4,5,6]
#z = x + y = [x0+y0, x1+y1, x2+y2]
#s = z.sum() = x0+y0+x1+y1+x2+y2
#x.grad 在s运算中对x求导   也就是当中的x0,x1,x2求导  为1，1，1 

Deep Learning Building Blocks: Affine maps, non-linearities and objectives（线性映射与非线性映射）

Affine maps
也可以说是线性映射，即为f(x) = Ax + b
nn.Linear(inputSize,outputSize,bias=True)
输入(N, inputSize)
输出(N, outputSize)

lin = nn.Linear(5,3)
data = autograd.Variable(torch.randn(2, 5))
print(lin(data))

下面展示 运行结果。

tensor([[-0.6649,  1.1707, -0.5394],
        [-0.9146,  1.1783,  0.4715]], grad_fn=<AddmmBackward>)

Non-Linearities
非线性，常用的函数有 tanh(x),σ(x),ReLU(x) 这些都是激励函数
在pytorch中大部分激励函数在torch.functional中

data = autograd.Variable( torch.randn(2, 2) )
print(data)
print (F.relu(data))#relu函数是小于零是0，大于零就是它本身

下面展示 运行结果。

tensor([[ 0.1168,  1.3061],
        [ 0.9262, -1.6161]])
tensor([[0.1168, 1.3061],
        [0.9262, 0.0000]])

Softmax and Probabilities
softmax是x_i/sum(x)

data = autograd.Variable( torch.randn(5) )
print(data)
print(F.softmax(data))
print(F.softmax(data).sum())
print(F.log_softmax(data))

下面展示 运行结果。

tensor([ 0.2949, -0.7197, -1.4672, -0.5649,  0.9722])
tensor([0.2547, 0.0923, 0.0437, 0.1078, 0.5014])
tensor(1.0000)
tensor([-1.3676, -2.3822, -3.1297, -2.2274, -0.6903])