PyTorch

PyTorch GitHub

Libraries

• Genral PyTorch Installation Using pip/conda:

pip install torch
pip install torchvision
pip install torchaudio

• Importing torch library

import torch

• 1D Empty tensor filled with uninitialized data

x = torch.empty(2)

tensor([1.8331e-40, 0.0000e+00])

• Random tensor

x = torch.rand(2)

tensor([[0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00], [0.0000e+00, 0.0000e+00, 4.2531e-05, 1.0802e-05]])

• Zeros tensor

x = torch.zeros(2)

tensor([0., 0.])

• Ones tensor

x = torch.ones(2)

tensor([1., 1.])

• Check data type

x = torch.ones(2)
x.dtype

torch.float32

• Tensor Shape and Size

x = torch.ones(2,4)
x.size()

torch.Size([2, 4])

• Define Tensors

x = torch.tensor([2.5,0.1])

tensor([2.5000, 0.1000])

y=torch.tensor([4.9,4.3])

tensor([4.9000, 4.3000])

• Operation's on Tensor
• Adding Tensor x and y

z = x+y

tensor([7.4000, 4.4000])

y.add_(x)

tensor([7.4000, 4.4000])

• Underscore uses for Inplace operation

x.add_(y)

tensor([9.9000, 4.5000])

x

tensor([9.9000, 4.5000])

y

tensor([7.4000, 4.4000])

x.sub_(y)

tensor([2.5000, 0.1000])

• Built-in operation

z= torch.sub(x,y)

tensor([2.5000, 0.1000])

• Random 2D Tensor

x = torch.rand(5,4)
x

tensor([[0.3094, 0.3055, 0.9537, 0.1301], [0.4614, 0.6939, 0.5546, 0.1630], [0.9947, 0.1444, 0.0529, 0.8653], [0.9795, 0.6218, 0.5568, 0.8080], [0.5672, 0.0596, 0.5012, 0.3082]])

• Slicing operation

x[:,0]

tensor([0.3094, 0.4614, 0.9947, 0.9795, 0.5672])

x[0,:]

tensor([0.3094, 0.3055, 0.9537, 0.1301])

• item() used for pure data representation

x[1,1].item()

0.6939277052879333

• Random 4 by 4 Matrix

x=torch.rand(4,4)
x

tensor([[0.0519, 0.0100, 0.5350, 0.1515], [0.8415, 0.6998, 0.0026, 0.1225], [0.7409, 0.2396, 0.6612, 0.0884], [0.8749, 0.2309, 0.2504, 0.6981]])

• Reshaping Matrix
• We convert 4x4 matrix to 1D Tensor

x.view(16)

tensor([0.0519, 0.0100, 0.5350, 0.1515, 0.8415, 0.6998, 0.0026, 0.1225, 0.7409,0.2396, 0.6612, 0.0884, 0.8749, 0.2309, 0.2504, 0.6981])

• Reshaping Matrix 4x4 to 8x2

x.view(-1,8)

tensor([[0.0519, 0.0100, 0.5350, 0.1515, 0.8415, 0.6998, 0.0026, 0.1225], [0.7409, 0.2396, 0.6612, 0.0884, 0.8749, 0.2309, 0.2504, 0.6981]])

• Reshaping Matrix 4x4 to 2x8

x.view(8,2)

tensor([[0.0519, 0.0100], [0.5350, 0.1515], [0.8415, 0.6998], [0.0026, 0.1225], [0.7409, 0.2396], [0.6612, 0.0884], [0.8749, 0.2309], [0.2504, 0.6981]])

• Reshaping Matrix 4x4 to 2x8

x.view(8,-1)

tensor([[0.0519, 0.0100], [0.5350, 0.1515], [0.8415, 0.6998], [0.0026, 0.1225], [0.7409, 0.2396], [0.6612, 0.0884], [0.8749, 0.2309], [0.2504, 0.6981]])

• Converting Tensor to Numpy and Vice-Versa

Importing numpy

import numpy as np

a = torch.ones(5)
a

tensor([1., 1., 1., 1., 1.])

• Convert Tensor to Array

b = a.numpy()
type(b)

numpy.ndarray

• Tensor data type

type(a)

torch.Tensor

• Convert Numpy Array to Tensor

b = torch.from_numpy(a)
b

tensor([1., 1., 1., 1., 1.], dtype=torch.float64)

• Check if you have CUDA is available for GPU operation

'gpu' if torch.cuda.is_available() else 'cpu'

..