PyTorch is a python package that simplifies neural network training processes through the use of tensor computations, automatic differentiation, and dynamic computation graphs. It is primarily implemented in C++, but can be run on python. This allows it to have extreme coding simplicity of python, while still being computationally efficient.
The simplest way to download PyTorch would be through the use of pip (python package manager)
pip should already be installed for python versions 3.4+
Run the following command in the terminal to download torch and additional packages (torchvision and torchaudio are optional, but have helpful functionalities for computer vision or audio purposes)
pip install torch torchvision torchaudio
For the purpose of this tutorial, we will be using premade dataset in torch from the
[torch.utils.data](<http://torch.utils.data>) module. In order to properly test the NN, we don’t only need a data set for training the model, but also for testing it. Having overlap between both datasets isn’t ideal as the model can overfit the training data, giving inaccurate results.
The data also needs to be organized for it to be usable. The provided data sets are organized and can further be customized with torch functions.
You have probably seen a graphical representation of a neural network like these before.
The input is the information that needs to be evaluated. For example, for a number recognizer for a 100 by 100-pixel image, there would be 10,000 inputs for each pixel. These inputs are fed into a series of weights that scales the given input by a factor (image each line to have a multiplicative effect), and a bias (so each node would be shifted by a number).