Tensor

0: scalar

1: vector

2: matrix

3+: tensor

matrix, tensor 变量一般大写命名

---

rand

zeros

ones

arange, 左闭右开

---

ndim

dtype, 默认是float32

size

Zeros_like

tensor datatype is one of 3 big errors with pytorch & deep learning

1. Tensors not right datatype, tensor.dtype
2. Tensors not right shape, tensor.shape
3. Tensors not on right device, tensor.device
   1. cpu
   2. cuda
   3. tpu

tensor.type(torch.int64)强制类型转换

---

Operation

• Addition

  +, touch.add()

• Sub

  -, touch.sub()`

• Multiplication (element wise) 数乘

  *

• Division

  \

• Matrix Multiplication 点乘/矩阵乘法

  @, torch.matmul(), torch.mm()

  这个比for快很多%%time

• Transpose 转置

  tensor.T

Matrix Multiplication的条件

1. inner dimension must match
2. resulting matrix has the shape of the outer dimension

cat

chunk

Aggregation

min, max, mean, sum, etc.

• min,

  x.min(), torch.min(x)

• max,

  x.max(), torch.max(x)

• mean,

  x.mean(), torch.mean(x)

• sum

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

• argmin, 找到tensor中min的坐标

  x.argmin()

• argmax, 找到tensor中max的坐标

  x.argmax()

---

Shape

reshape

view

stack

squeeze

unsqueeze

Permute

Indexing

xxx

Numpy

torch.from_numpy()

x.numpy()

Reproducibility

Pseudo random

RANDOM_SEED = 42
Torch.manual_seed(RANDOM_SEED)
touch.rand(3, 4)

GPU

1. colab, easiest
2. Buy GPU
3. Cloud Computing

torch.cuda.is_available()

tensor.to(device)
tensor.cpu().numpy()