Keras and Deep Learning Agenda : 6:30 - 6:45 PM: News and Introduction 6:45 – 7:30 PM: Introduction to Keras 7:30 - 8:15 PM: A demo of Google Cloud ML Engine for Deep Learning 8:15 - 8:30 PM: Wrap-up

1. Data Science | Design | Technology
(February 20, 2019)
https://www.meetup.com/DSDTMTL

2. Agenda
6:30 - 6:45 Intro & News
6:45 - 7:30 Introduction to Keras
7:30 - 8:15 Google Cloud ML Engine
8:15 - 8:30: Wrap-up
2
Keras and Deep Learning

3. 3
A special thanks to…
Contributors
venue sponsor / snacksbrain

4. 4
Movement to promote Québec’s technology industry:
• Increase knowledge on the innovation ecosystem
• Promote entrepreneurship, careers and education
• Increase international recognition
Technopolys
More than 550 contributing companies
@technopolys_qc
www.linkedin.com/showcase/technopolys

5. 5
Technopolys Ambassadors
50 ambassadors (the voice of Quebec’s technology industry)
Objective for 2019
100
www.technopolys.ca

6. 6
Ecosystem News
● Feb 27: Kubernetes Q1 Meetup 2019 – Cloud Native Computing Foundation update
https://www.meetup.com/Kubernetes-Montreal/events/258883214/
● Feb/March: Desjardins Labs:
https://www.meetup.com/DesjardinsLab/events/
● March 12: Data Driven Montreal - HEC Forecast
https://www.meetup.com/DataDrivenMTL/events/259082491/
● April 10-11: World Summit AI - Americas (Montreal):
https://americas.worldsummit.ai/

7. 7
DSDT on Slideshare
● www.slideshare.net/DSDT_MTL
● All presentations since DSDT creation

8. 8
Keras and Deep Learning
Nicolas Feller
“Keras: From Core concepts
to Advanced
Experimentation”
Florian Soudan
“Demo of Google Cloud ML Engine
for Deep Learning”

9. Keras: From Core
Concepts to
Advanced
Experimentation
Data Science | Design | Technology 9

10. Outline
- What is Keras
- How to use Keras
- Examples and Tutorials
- Advanced(ish) Example
- Upcoming Roadmap

11. Keras: API for specifying & training differentiable
programs (deep learning for humans)
Keras API
Tensorflow or Theano,
MXnet, CNTK
Hardware: CPU, GPU, TPU

12. Official high-level API of Tensorflow
● Tensorflow specific functionality
○ tf.data pipelines
○ Estimators, conceptual abstractions that isolates training, evaluating and deploy as tensorflow
○ Multiple GPUs
■ data parallelism - same model on each device
■ device parallelism - part of the model on each device
○ TPUs
○ Tensorboard - visualize learning
○ Data Augmentation
○ Eager execution (currently limited, improvements on the way)

13. Who makes/uses Keras?
https://discovery.hgdata.com/product/keras

14. Jeff Hale Power Score Criteria
● Online job listings
● KDnuggets usage survey
● Google search volume
● Medium articles
● Amazon books
● arXiv articles
● Github activity

15. Deep learning for real life
● Android tensorflow runtime
● iOS CoreML
● Keras.js and WebDNN GPU accelerated JS runtimes
● Google Cloud via tensorflow serving - ML engine
● Web backend in Flask
● JVM in DL4J
● Raspberry Pi

16. Start Using Keras in seconds
● Start a Jupyter Notebook from
Tensorflow docker
● Regular python download
● Access Google Colabs from
any gmail address

17. Demo: http://bit.ly/2tzaRWJ
● Model types
○ Sequential
○ Functional
○ Model Subclassing
● Visualize model
○ Summary
○ Plot_model
● Extra features
○ Use model
○ tf.data
○ Custom layers
○ Callbacks
○ Saving and restoring model
○ Pretrained Models

18. Upcoming Features
● Eager execution
● Distributed training - tensorflow like performance
○ Parameter strategies
● Tight integration to build and productionize
○ Export to tf life and tfx
● Better tensorboard integration (profiler, displaying graph correctly)
● Canned models
● Improved performance

19. Questions?
19

20. A demo of Google
Cloud ML Engine for
Deep Learning
Data Science | Design | Technology 20
https://github.com/ivado-labs/meetup-googleml-keras

21. 21
Typical Deep Learning Workflow (Iterative)
1. Data pipeline preparation
2. Model development/optimization
3. Training monitoring
4. Result analysis Google Cloud
Buckets
ML Engine
Computer
Local development
Data
Images
Logs &
Weights
Read Logs
Launch
Tensorboard
Server
Launch
Notebook
Server
Submit Training Job
1 4 3
2

22. Merci / Thank You
@DsdtMtl
Data Science | Design | Technology
(Check for next DSDT meetup at https://www.meetup.com/DSDTMTL)
http://bit.ly/dsdtmtl-in