【ML笔记】PyTorch Essential（前置知识）

以下是notebook直接转换成markdown，在Google Colab上可以正常执行。
可以在这里直接进入👉https://colab.research.google.com/drive/1oSCI36tAQq9gfTaHmG5PyDwcPzA1P_pl?usp=sharing
内容为pytorch和ML的入门经典内容，各位可以康一康哈，但是主要以笔记为主，因此会有些中英参半

Basic

import torch
ones = torch.ones(size=(3, 4));
ones

zeros = ones * 0;
zeros

torch.arange(1,11)

torch.arange(1,101,2)

torch.rand(size=(10,10,10))

one_to_ten = torch.range(1,11);
one_to_ten

ten_zeros = torch.zeros_like(one_to_ten);
ten_zeros

RAND = torch.rand(size=(5,5));
RAND
RAND.dtype

float_32_tensor = torch.tensor([3.0, 6.0, 9.0],
                               dtype = None,
                               device = None,
                               requires_grad=False)
float_16_tensor = float_32_tensor.type(torch.float16)
float_16_tensor

float_16_tensor * float_32_tensor

int_32_tensor = torch.tensor([3, 6, 9],
                             dtype = torch.int32)
int_32_tensor.dtype

some_tensor = torch.rand(3, 4)
some_tensor

some_tensor.dtype
some_tensor.size()
some_tensor.shape
some_tensor.device

some_tensor.dtype

some_tensor.size()

some_tensor.shape

tensor = torch.tensor([1, 2, 3])
tensor + 1090

tensor *= 10

tensor

Matrix Multiplication

import torch
test = torch.rand(2,4)
test

tensor_a = torch.rand(3,4)
tensor_b = torch.rand(3,4)
tensor_b_t = tensor_b.T
tensor_c = torch.matmul(tensor_a, tensor_b_t)
print(tensor_b, '\n', tensor_b_t)

Tensor Aggregation

import torch
x = torch.arange(10,101,10)
x

# max & min
x.max(), torch.max(x), x.min(), torch.min(x)

# mean（平均数）
# 注意数据类型是浮点
torch.mean(x.type(torch.float)), x.type(torch.float).mean()

# sum
x.sum(), torch.sum(x)

x.argmin(), x.argmax()

x[x.argmax()], x[x.argmin()]

Reshaping, Stacking, Squeezing and Unsqueezing

import torch
x = torch.arange(1., 28.)
x

# reshape（重新构造tensor）
# 需要元素个数正确，e.g. 1*9=9
# 可能与原张量共享内存
x_reshape = x.reshape(3, 3, 3)
x_reshape[0, 0, 0] = 100
x_reshape, x

# change view
# view与原张量共享内存
z = x.view(3, 3, 3)
z[0,0,0] = 100
z, x

# stack
x = torch.arange(1., 10.)
x_stack = torch.hstack([x, x, x, x])
x_stack

# squeeze, unsqueeze
# 其中，unsqueeze的第二个穿传参n代表在n与n+1中插入一个维度
x = torch.arange(1., 10.)
a = torch.unsqueeze(x, 0)
x = torch.rand(2, 1, 2, 1, 4)
a = torch.squeeze(x, (-1, -2))
a = torch.unsqueeze(x, 1)
a.shape

# permute
# 用于调整
x = torch.tensor([[1, 1, 1],
                [2, 2, 2]])
# a = torch.permute(x, (5, 4, 3, 2, 1, 0))
a = torch.permute(x, (1, 0))
a

Index

import torch
x = torch.arange(1, 28).reshape(3, 3, 3)
x[:, 0, :]

Reproducibility

import torch
torch.rand(3, 3)

# manual_seed只在一个代码块中起作用
torch.manual_seed(2233)
a = torch.rand(3, 3)

torch.manual_seed(2233)
b = torch.rand(3, 3)
print(a, '\n', b)
print(a == b)

GPU Accessing & Using

import torch
import numpy as np
torch.cuda.is_available()

device = "cuda" if torch.cuda.is_available() else "cpu"
torch.cuda.device_count(), device

a = torch.rand(3, 3)
a = a.to(device)

# numpy只在CPU上可用
a.cpu().numpy()