Pytorch Zero to GANS Assignment 1

5 Tensor Operations

Pytorch is a tensor operations library used mainly for deep learning. In this notebook I introduce five Tensor operations.

torch.Tensor.contiguous
torch.Tensor.expand
torch.Tensor.fill_diagonal_
torch.Tensor.index_add_
torch.Tensor.index_fill_

# Import torch and other required modules
import torch

Function 1 - torch.Tensor.contiguous

Each newly created tensor is contiguous, but if you transpose it, expand it etc. you get a non-contiguous Tensor, Which in some cases is problematic. The contiguous function creates a contiguous copy of the Tensor if it isn't contiguous already, otherwise it returns self.

# Example 1 - working
x = torch.Tensor([[1, 2], [3, 4.]])
y = torch.transpose(x,0 , 1)
y = y.contiguous()
y

tensor([[1., 3.],
        [2., 4.]])

A basic use of contiguous. When y is created, its simply a different way to look at x. After using contiguous, a copy of y is created that looks just like y for a human but has a different representation in memory, specifically, just like if we would create a whole new Tensor with the data stored in y.