From Keras to Production

OUTDATED! FOR THE NEWEST VERSION SEE https://github.com/codecentric/from-jupyter-to-production-workshop (a fork of this repository)

Data Source

https://www.kaggle.com/moltean/fruits

Notebooks

https://github.com/codecentric/from-keras-to-production-workshop.git

Used Docker Images

docker pull codecentric/from-keras-to-production-baseimage
docker pull codecentric/tensorflow-serving-baseimage
docker pull tsabsch/airflow-baseimage

Start Jupyterlab

With docker-compose (recommended, no Docker Toolbox)

# Without Airflow
docker-compose up

# With Airflow
docker-compose -f docker-compose.yml -f optional-airflow.yml up

Without docker-compose

Jupyterlab

# Linux/Mac (Docker version >= 17.06)
docker run -p 8888:8888 --mount type=bind,source=$(pwd)/notebooks,target=/keras2production/notebooks codecentric/from-keras-to-production-baseimage

# Docker for Windows (Docker version >= 17.06)
docker run -p 8888:8888 --mount type=bind,source=%cd%/notebooks,target=/keras2production/notebooks codecentric/from-keras-to-production-baseimage

# Docker for Windows (Docker version < 17.06)
docker run -p 8888:8888 -v %cd%/notebooks:/keras2production/notebooks codecentric/from-keras-to-production-baseimage

# Docker Toolbox (Windows 7, 8 and Windows 10 Home; a separate VM for Docker)
docker run -d -p 8888:8888 codecentric/from-keras-to-production-baseimage
# Copy notebooks manually into the container
## get container id
docker ps
## copy into container
docker cp notebooks <container id>:/keras2production
# After the first day, stop the container
docker stop <container id>
# On the second day, start the container again
docker start <container id>

With Docker Toolbox, the JupyterLab instance might be available at 192.168.99.100:8888, not localhost:8888.

TensorFlow Serving

docker run -p 8501:8501 -p 8500:8500 --mount type=bind,source=$(pwd)/notebooks/12-models/fruits/,target=/models/fruits -e MODEL_NAME=fruits codecentric/tensorflow-serving-baseimage

Airflow

# Docker for Linux/Mac/Windows (Docker Version >= 17.06)
docker run -p 8080:8080 --mount type=bind,source=$(pwd)/notebooks/04-airflow/dags,target=/usr/local/airflow/dags \
                        --mount type=bind,source=$(pwd)/notebooks/04-airflow/exercise-dataset,target=/exercise-dataset \
                        tsabsch/airflow-baseimage

# Docker for Windows (Docker version < 17.06)
docker run -p 8080:8080 -v %cd%/notebooks/04-airflow/dags:/usr/local/airflow/dags \
                        -v %cd%/notebooks/04-airflow/exercise-dataset:/exercise-dataset \
                        tsabsch/airflow-baseimage

# Docker Toolbox (Windows 7, 8 and Windows 10 Home)
docker run -d -p 8080:8080 tsabsch/airflow-baseimage
# Copy notebooks manually into the container
## get container
docker ps
## copy into container
docker cp notebooks/04-airflow/exercise-dataset <container id>:/exercise-dataset
docker cp notebooks/04-airflow/dags <container id>:/usr/local/airflow/dags

In general

• Replace current directory in commands with either %cd% (Windows) or $(pwd) Mac/Linux
• --mount is supported since Docker version 17.06. If you use an older version you have to use -v (Volumes). See the Example in the Airflow section above.

Old Slides

pip install -r requirements.txt
cd slides
jupyter nbconvert end2end_ds.ipynb --to slides --post serve --reveal-prefix=reveal.js

References and Further Information

General

• Cheatsheet for working with IPython/Jupyter

• Cheatsheet for Docker

• Free notebooks from the book "Deep Learning with Python"

• Introduction to Reinforcement Learning (Youtube)

• Keras examples

Convolutional Networks

• Visualization of image kernels

• Visualization of activation maps

• Combining channels in convolutional layers

Natural Language Processing

• Using pre-trained word embeddings in a Keras model

• Text preprocessing

Production-Ready Data Science

• What’s your ML test score? A rubric for ML production systems

• A walkthrough of DVC, DVC dependency management