Deep Learning with Pytorch: A 60 Minute Blitz

  Blitz ： Air raids with Blitzkrieg

The goal of this document

• Understand at a higher level Pytorch Of Tensor Libraries and neural networks
• Train a small-scale neural network to classify images

This article assumes that you are basically familiar with numpy

What is? Pytorch

Pytorch It is based on scientific computing package Python, Target two groups of people
  1. To make use of GPU The ability to replace NumPy
  2. Provide the maximum flexibility and speed of deep learning research platform

Start

Tensors

Tensors Be similar to Numpy Of narrays, It's just Tensors have access to GPU Accelerate Computing
from __future__ import print_function
import torch

Construct a 5*3 Uninitialized matrix
x = torch.Tensor(5, 3)

Build a random initialization matrix
x = torch.rand(5, 3)

Get the size of the matrix
print(x.size())
Out: torch.Size([5, 3]) ; Output result

operation

Pytorch There are many syntax for operations . In the following example, we will see the addition operation
y = torch.rand(5, 3)
print(x + y)

perhaps , Additive grammar ②
torch.add(x, y)

Add : Provide an output tensor As a parameter
result = torch.Tensor(5, 3)
torch.add(x, y, out=result)
print(result)

In situ ( in-place ) Add
y.add_(x)
Note : Anything that makes one Tensor Operations that change locally should be followed by a ’_’

You can also use things like NumPy The same index means
print(x[:, 1])

Change the size : If you want to resize perhaps reshape tensor, have access to torch.view
x = torch.randn(4, 4)
y = x.view(16)
z = x.view(-1, 8); -1 The value of is derived from another dimension
print(x.size(), y.size(), z.size())
Out: torch.Size([4, 4]), torch.Size([16]), torch.Size([2, 8])

also 100+ In the operation , see here

Numpy Bridge

Will a Torch Tensor convert to NumPy array Or the reverse is a very easy thing

Torch Tensor and NumPy array Will share hidden memory addresses , So changing one will change both

Will a Torch Tensor convert to NumPy array

a = torch.ones(5);
b = a.numpy()

to glance at NumPy array How to change the value
a.add_(1)

Will a NumPy array convert to Torch Tensor

See how to automatically NumPy array convert to Torch Tensor
a = np.ones(5)
b = torch.from_numpy(a)
np.add(a, 1, out=a)

CUDA Tensors

Tensors have access to .cuda Method transferred to GPU operation

appendix : And that 100+ Middle operation ….

Tensors

torch.is_tensor(obj)
  return True If obj It's a Pytorch tensor
torch.is_storage(obj)
  return True If obj yes pytorch Store the object
torch.set_default_tensor_type()
torch.numel(input)
  Return to input Tensor The total number of elements in
torch.set_printoptions(precesion=None, threshold=None, edgeitems=None, linewidth=None, profile=None)

Creation Ops

torch.eye(n, m=None, out=None)
  return 2-D Divide the diagonal by 1 For all other 0 Matrix
torch.from_numpy(ndarray)
  from numpy.ndarray To create a Tensor
torch.linspace(start, end, steps=100, out=None)
  Return to one dimension Tensor,steps Divide equally from start To end
torch.logspace(start, end, steps=100, out=None)
  Return to one dimension tensor,steps Divide equally from 10^start To 10^end
torch.ones(*size,out=None)
  Returns that all elements are scalars 1 Of tensor, from size decision shape
torch.ones_like(input, out=None)
  Return and input Of size All the same elements are 1 Of tensor
torch.arange(start=0, end, step=1, out=None)
  Return to one dimension tensor, Size is (end-start)/step, [start, end), The difference between every two adjacent elements step
torch.range(start, end, step=1, out=None)
  It's more than arange One more. end Elements
Warning : This function is due to arange Has been abandoned
torch.zeros(*size, out=None)
  Returns that all elements are 0 Of Tensor, from size decision shape
torch.zeros_like(input, out=None)
  Refer to the above ones_like

Indexes , section , union , Rotation operation

torch.cat(seq, dim=0, out=None)
  cat for concatenate ( String ) Link
  Will be given Tensor Sequences are linked on a given dimension , In addition to the cat dimension , all Tensor must size identical
x = torch.randn(2, 3)
y = torch.Tensoe(2, 3)
torch.cat((x, y), dim=1)

torch.chunk(tensor, chunks, dim=0)
  Will a Tensor Decompose into a given number of blocks

To be continued…

torch.gather(input, dim, index, out=None)