1 of 16

TDS - #38 Soirée Multi-sujets - 18/06/2019

vfdev-5

Victor FOMIN

DL engineer @ Magellium

2 of 16

your PyTorch neural networks

make the common things easy and the hard things possible

https://pytorch.org/ignite/

https://github.com/pytorch/ignite

joint work with Alykhan Tejani (Twitter), Joshi Anmol (Boeing), Jason Kriss (ML/DL Consultant) and other contributors.

3 of 16

What we’ll cover

​

in 25 minutes

​

​

  • 🔥PyTorch: dynamic autograd library

​

  • Ignite: library to help with training neural networks
    • what and why ?
    • API and usage
    • public use-cases
    • demo

​

Hold tight !

4 of 16

​

​

An open source deep learning platform that provides a seamless path from research prototyping to production deployment.

  • Python-first library for rapid prototyping
  • Rich ecosystem and active community
  • New hybrid front-end seamlessly transitions between eager mode and graph mode to provide both flexibility and speed.
  • Distributed training + Automatic Mixed Precision

​

5 of 16

6 of 16

7 of 16

8 of 16

Today all frameworks provide similar front-ends

9 of 16

🔥PyTorch + Ignite🔥

    • what and why ?
    • API and usage
    • public use-cases
    • demo with EfficientNet-B0

10 of 16

🔥PyTorch + Ignite🔥

You get full-featured training loops without the boilerplate

  • Training loop abstraction
  • Out-of-box tested features
    • Early stopping
    • Model checkpointing
    • Metrics
    • Logging
    • Advanced lr scheduling
    • ...
  • Less code to write
  • More flexible code
  • More factorized code

Pure PyTorch

PyTorch+Ignite

11 of 16

Yet another Keras for PyTorch ?

-> Yes, but better 😉

  • Model fit/predict => Engine with update function -> Trainer/Evaluator

​

​

​

  • Callbacks => Event handlers

12 of 16

Ignite concepts

13 of 16

Demo time

Fine-tuning of pretrained EfficientNet-B0 on CIFAR-100 with Ignite

⚠ Will be technical ⚠

No boring MLP on MNIST

https://github.com/pytorch/ignite/blob/master/examples/notebooks/EfficientNet_Cifar100_finetuning.ipynb

14 of 16

Public use-cases

Public use-cases:

​

15 of 16

Install Ignite

Use it

Feedback us

Contribute

​

​

Feedbacks:

​

Contributions (Issues/PRs) are very welcome 🎉

  • Already 38 contributors
  • Join us 😉

​

​

​

​

​

​

​

​

16 of 16

Thank you !

​

Questions ?

​

​