Introduction to PyTorch

Professor Albert Bifet

• Framework for Machine Learning.
• Compare with other popular frameworks like TensorFlow.

• Consists of two main components

  • Numeric representation optimised for GPUs
  • Deep Learning framework

• Motivation

  • Deep Learning using Python accelerates the path from research prototyping to production deployment
  • Uses a dynamic computation graph approach
  • Uses hardware accelerators as GPUs

• Deep Learning

  • Inspired by Biological neural networks, but not the same
  • Construct improved features representing the input problem
  • Combine features to improve network predictive capability

• Neural Network

  • Example network

• PyTorch audience

  • Replacement for NumPy - using the power of GPUs
  • Deep Learning research platform providing
    • Speed
    • Flexibility

• TensorFlow

  • Playground - https://playground.tensorflow.org

• Stochastic Gradient Descent

  • Strategy for updating weights ('learning')
  • Computed automatically within PyTorch

• Skorch

  • Scikit-Learn wrapper for PyTorch

• PyTorch example

  • MNIST Dataset (National Institute of Standards and Technology)
  • Classical dataset used within Deep Learning
  • PyTorch data Loading
  • PyTorch model
  • Training & Evaluation
  • Evaluating Training Loss
  • Evaluating Testing Loss

• TensorFlow

  • Feature contrast with PyTorch
  • Static topology
  • Model building similar to PyTorch