PyTorch 基本使用方法

创建矩阵

创一个矩阵，有木有觉得很爽，如果用tensorflow的同学可能会这么觉得。。。

x = torch.empty(5, 3)  # 构建未初始化的矩阵；产生很小的数，可以当做 0；结果是一个张量 tensor; 所有的输入和计算都是 tensor 格式。
x

tensor([[ 0.0000e+00, -8.5899e+09,  2.1110e-19],
        [-3.6902e+19,  1.1210e-44,  0.0000e+00],
        [ 0.0000e+00,  0.0000e+00,  0.0000e+00],
        [ 0.0000e+00,  0.0000e+00,  0.0000e+00],
        [ 0.0000e+00,  0.0000e+00,  0.0000e+00]])

x = torch.rand(5, 3)  # 构建随机初始化矩阵
x

tensor([[0.4896, 0.0190, 0.4041],
        [0.7866, 0.5946, 0.2882],
        [0.9009, 0.2157, 0.7783],
        [0.3611, 0.1883, 0.6338],
        [0.6846, 0.9554, 0.6611]])

初始化一个全零的矩阵

x = torch.zeros(5, 3, dtype=torch.long)
x

tensor([[0, 0, 0],
        [0, 0, 0],
        [0, 0, 0],
        [0, 0, 0],
        [0, 0, 0]])

直接传入数据

x = torch.tensor([5.5, 3]) # 从数据直接直接构建tensor
x # tensor([5.5000, 3.0000])

x = x.new_ones(5, 3, dtype=torch.double) # 从一个已有的tensor构建一个tensor，new_ones重用以前的信息，例如数据类型；       

x = torch.randn_like(x, dtype=torch.float) # 产生一个形状一样的矩阵  
x

tensor([[ 0.2848,  0.9409, -0.8545],
        [ 0.0582,  0.0050, -1.1220],
        [-1.1389,  0.2913,  0.1471],
        [-0.6798,  0.2851,  1.5386],
        [-0.7647,  1.0754, -0.7901]])

矩阵属性

展示矩阵大小

x.dtype # torch.float32
x.size() # torch.Size([5, 3]) 
x.shape # torch.Size([5, 3])

# 改变tensor形状

x = torch.randn(4,4)
y = x.view(16)
z = x.view(-1,8)
z

tensor([[-1.6207, -1.0893, -1.1299, -0.4559, -1.8922,  0.4771,  2.0306,  1.6977],
        [-0.8509,  2.1227, -0.3943, -1.8941, -0.6286,  0.3956,  0.5723,  1.4658]])

基本计算方法

x = torch.rand(5, 3)
y = torch.rand(5, 3)
x + y

tensor([[0.9174, 1.3786, 1.0915],
        [1.0750, 1.7506, 1.3954],
        [0.5356, 1.2608, 1.5221],
        [1.4083, 1.0147, 0.9909],
        [0.8153, 0.6314, 0.5632]])

torch.add(x, y)#一样的也是加法

tensor([[0.9174, 1.3786, 1.0915],
        [1.0750, 1.7506, 1.3954],
        [0.5356, 1.2608, 1.5221],
        [1.4083, 1.0147, 0.9909],
        [0.8153, 0.6314, 0.5632]])

result = torch.empty(5,3) 
torch.add(x, y, out=result) 

tensor([[0.9174, 1.3786, 1.0915],
        [1.0750, 1.7506, 1.3954],
        [0.5356, 1.2608, 1.5221],
        [1.4083, 1.0147, 0.9909],
        [0.8153, 0.6314, 0.5632]])

# result = x + y

y.add_(x)

tensor([[0.9174, 1.3786, 1.0915],
        [1.0750, 1.7506, 1.3954],
        [0.5356, 1.2608, 1.5221],
        [1.4083, 1.0147, 0.9909],
        [0.8153, 0.6314, 0.5632]])

x = torch.randn(1) # 得到Python数值

x.item() # -0.49584466218948364

索引

x[:, 1]  # tensor([-1.1208, -0.0828, -0.4144, -0.7263,  0.1368])

view操作改变矩阵维度

x = torch.randn(4, 4)
y = x.view(16) 
z = x.view(-1, 8) # -1 代表自动计算
print(x.size(), y.size(), z.size())

torch.Size([4, 4]) torch.Size([16]) torch.Size([2, 8])

与Numpy的协同操作

# tensor 转 numpy
a = torch.ones(5)
b = a.numpy()
b # array([1., 1., 1., 1., 1.], dtype=float32)

# numpy 转 tensor
import numpy as np
a = np.ones(5)
b = torch.from_numpy(a)
b # tensor([1., 1., 1., 1., 1.], dtype=torch.float64)

# CPU转成GPU

if torch.cuda.is_available():

    device = torch.device("cuda")

    y = torch.ones_like(x, device=device)

    x = x.to(device)

    z = x + y

    print(z)

    print(z.to("cpu", torch.double))


model = model.cuda()

# GPU不能直接转化成numpy，必须先转化为CPU,再转化为numpy

y.to("cpu").data.numpy()

y.cpu().data.numpy()
 

PyTorch 基本使用方法

原文：https://www.cnblogs.com/fldev/p/14466039.html