In-depth understanding of the core of Pytorch, the road to Tensor’s breakthrough!

Today I will make a record of Pytorch’s tensor content.

At the same time, I hope I can provide you with some help!

Because the content shared today is definitely some examples of very useful information.

Let’s give a brief introduction first. In PyTorch, tensor is the core data structure. It is a multi-dimensional array, similar to the array in NumPy. Tensors are not only containers for storing data, but also the basis for various mathematical operations and deep learning operations.

The following is a summary from three aspects:

• The concept of tensor
• The principle of tensor
• The operation of tensor

Picture

The concept of tensor

1. The definition of tensor

Tensor is a multi-dimensional An array, which can be a scalar (a zero-dimensional array), a vector (a one-dimensional array), a matrix (a two-dimensional array), or an array with a higher dimension.

In PyTorch, a tensor is an instance of torch.Tensor and can be created in different ways, such as directly from a Python list, a NumPy array, or through a specific function.

import torch# 创建一个标量scalar_tensor = torch.tensor(3.14)# 创建一个向量vector_tensor = torch.tensor([1, 2, 3])# 创建一个矩阵matrix_tensor = torch.tensor([[1, 2, 3], [4, 5, 6]])# 创建一个3D张量tensor_3d = torch.rand((2, 3, 4))# 2行3列4深度

2. Attributes of tensors

Each tensor has some important attributes, including shape (shape), data type (dtype) and device (device).

# 获取张量的形状shape = tensor_3d.shape# 获取张量的数据类型dtype = tensor_3d.dtype# 获取张量所在的设备device = tensor_3d.device

3. The shape of a tensor

The shape of a tensor defines its dimensions and the size in each dimension. For example, a tensor of shape (2, 3, 4) has 2 rows, 3 columns, and 4 depths. Shape is very important for understanding and manipulating tensors.

# 获取张量的形状shape = tensor_3d.shape# 改变张量的形状reshaped_tensor = tensor_3d.view(3, 8)# 将原始形状(2, 3, 4)变为(3, 8)

The principle of tensors

Tensors in PyTorch are implemented based on the Tensor class, which provides an abstraction of the underlying storage.

Tensors contain three main components:

• storage
• shape
• stride

1. Storage

(Storage) Storage is the place where data is actually stored. It is a continuous memory area. Multiple tensors can share the same storage, reducing memory consumption. The data in storage is arranged according to the shape of the tensor.

# 获取张量的存储storage = tensor_3d.storage()

2. Shape

The shape of a tensor defines its dimensions and the size in each dimension. Shape information helps explain how data in storage is organized.

# 获取张量的形状shape = tensor_3d.shape

3. Stride

Stride refers to the number of steps required to move to the next element in storage. Understanding strides helps understand performance when indexing and slicing within tensors.

# 获取张量的步幅stride = tensor_3d.stride()

Tensor operations

PyTorch provides a wealth of tensor operations, including mathematical operations, logical operations, indexing, slicing, etc.

Here are the most common centralized operations:

1. Mathematical operations

# 加法result_add = tensor_3d + 2# 乘法result_mul = tensor_3d * 3# 矩阵乘法matrix_a = torch.rand((2, 3))matrix_b = torch.rand((3, 4))result_matmul = torch.mm(matrix_a, matrix_b)

2. Logical operations

# 大小比较result_compare = tensor_3d > 0.5# 逻辑运算result_logical = torch.logical_and(result_add, result_compare)

3. Indexing and slicing

# 索引element = tensor_3d[0, 1, 2]# 切片sliced_tensor = tensor_3d[:, 1:3, :]

4. Shape operation

# 改变形状reshaped_tensor = tensor_3d.view(3, 8)# 转置transposed_tensor = tensor_3d.transpose(0, 2)

5. Broadcast

Broadcasting is an operation that automatically expands tensors so that tensors with different shapes can be processed element by element. computation.

# 广播tensor_a = torch.rand((1, 3, 1))tensor_b = torch.rand((2, 1, 4))result_broadcast = tensor_a + tensor_b

Finally

Today I introduce the basic concepts, principles and common operations of tensors in PyTorch.

Tensors, as the basic data structure in deep learning, are very critical for understanding and implementing neural networks.

The above is the detailed content of In-depth understanding of the core of Pytorch, the road to Tensor’s breakthrough!. For more information, please follow other related articles on the PHP Chinese website!