[Pycon] [new paper] "Luca Antiga" - PyTorch from the ground up

[info a pycon.it]

Title: PyTorch from the ground up
Duration: 240 (includes Q&A)
Q&A Session: 0
Language: en
Type: Training

Abstract: PyTorch is one of the leading frameworks for doing deep learning in Python. With its dynamic computation graph engine, greedy execution, tape-based automatic differentiation and a JIT compiler, PyTorch makes training the latest and greatest deep learning models no harder than it should be. All while ensuring top-notch performance and seamless integration with the rest of the Python data ecosystem.
The training session is aimed at building a working knowledge of how to use PyTorch to write deep learning models and train them successfully, starting from scratch. Introductory working knowledge of Python will be required, some NumPy and a general idea on deep learning will help. We will first introduce the tool as a whole, move on to how to represent various kinds of data such as images and text using tensors and learn how to solve practical tasks using popular deep neural network architectures.
The last hour will be devoted to some of the internals. One of the strengths of PyTorch is the accessibility of its codebase, which is key to a full understanding of what's going on under the hood. We will describe how the project is structured and what it takes to extend it.

Tags: [u'ComputerVision', u'Machine Learning', u'GPUComputing', u'data-analysis', u'machine-learning', u'data', u'machine-learning']