Basic but important Functions Of PyTorch | PyTorch Series | Snehit Vaddi
Introduction
PyTorch is a Python library that facilitates building Deep Learning models. It was introduced by Facebook. PyTorch is used for a number of applications like Natural Language Processing, Computer Vision, Self-driving cars, etc.
The basic building block of PyTorch is Tensors. In simple words, a Tensor is an array of numbers. In the NumPy library, these matrices are called nd-array. In PyTorch, 1d-tensor is a vector, 2d-tensor is a matrix, 3d- Tensor is a cube, and 4d-Tensor is a cube vector.
Basic Functions of PyTorch:
- Creation and Slicing of Tensor -
Import PyTorch module as “import torch”. For creating an n-dimensional tensor, we use the tensor function of torch module ie, torch.tensor([elements]).
Note: In PyTorch, tensors should be symmetric like 3,3. If the elements are of different shape, the compiler raises sequence exception.
- A Tensor can be initialized only with numbers. It can not contain any strings or characters.
- All the elements in a tensor are homogeneous. In case if there are different types of elements, the compiler itself will upcast the elements.
- torch.rand((m,n)) function is used to create a m x n tensor filled with random numbers in the interval [0,-1) ie, it follows Standard Normal Deviation.
2. Tensor — NumPy Bridge (NumPy array to Tensors vise versa) -
- torch.from_numpy() function is used to convert a n-dimensional NumPy array into n-dimensional tensor.
- Similarly, tensor.numpy() is used to convert an n-dimensional tensor to NumPy array.
3. Mathematical Operations On Tensors -
- PyTorch module provides a number of functions to perform different mathematical / matrix operations.
- In PyTorch, matrix multiplication could be calculated using 3 different methods.
- In matrix multiplication, multiplication fails if columns of A not equal to rows of B.
4. Gradient with PyTorch
In PyTorch, a variable initialized with ‘requires_grad= True’ specifies variable will undergo differentiation.
In the given example, we initialized ‘x’ with requires_grad. An equation ‘y’ in terms of x. In the next step, the derivative of the function using a backward() method. The last step is to find a partial derivative of ‘x’ wrt ‘y’.
5. Optimizer
PyTorch provides a number of built-in optimizers. like SGD, Adam, AdamW, Adamax, Rprop, etc.
The Optim module in Pytorch has pre-written codes foremost of the optimizers that are used while building a neural network.
Importing and Uploading into jovian.ml-
Jovian.ml is the best platform to share collaborate and run our notebooks in the cloud. To use jovian.ml, pip install jovian package, import jovian, and commit your notebook.
Conclusion:
In this post, we have covered the creation of tensors, tensor slicing, Tensor-NumPy bridge, Mathematical operations on tensors, Gradients, and Optim module. In the coming weeks, I will be diving deep into advanced topics like Logistic Regression, Neural Networks, etc.
Reference:
Jovian Notebook: https://jovian.ml/v-snehith999/pytorch-week1-blog
Author:🤠
- Snehit Vaddi
I am a Machine Learning enthusiast. I teach Machines how to See, Listen, and Learn.
Linkedin: https://www.linkedin.com/in/snehit-vaddi-73a814158/
Github: https://github.com/snehitvaddi
