TensorFlow Sketch-RNN pre-trained model demo

Discrlaimer: This is not an official Google product.

This is an example application demonstraing how Sketch-RNN: A Generative Model for Vector Drawings can be used to create a sketch generation application.

Products

• TensorFlow

• Google Compute Engine

Language

• Python

Prerequisites

1. A Google Cloud Platform Account

2. A new Google Cloud Platform Project for this lab with billing enabled

Do this first

First you launch a GCE instance with the following configuration.

• vCPU x 4

• Memory 8GB

• Debian GNU/Linux 9 (stretch) as a guest OS

• Allow HTTP traffic

• Assign a static IP address

You can leave other settings as default. Once the instance has started, log in to the guest OS using SSH and change the OS user to root.

$ sudo -i

All remaining operations should be done from the root user.

Install packages

# apt-get update
# apt-get install -y build-essential python-pip unzip python-cairosvg git
# pip install ipython==5.5.0 rdp==0.8 svgwrite==1.1.6 tensorflow==1.3.0 Flask==0.12.2
# pip install magenta

Download the pre-trained model binaries

# mkdir -p /opt/sketch_demo/models
# cd /opt/sketch_demo/models
# curl -OL http://download.magenta.tensorflow.org/models/sketch_rnn.zip
# unzip sketch_rnn.zip

Install the demo application

# cd $HOME
# git clone https://github.com/GoogleCloudPlatform/tensorflow-sketch-rnn-example.git
# cp -a tensorflow-sketch-rnn-example/sketch_demo /opt/
# cp /opt/sketch_demo/sketch_demo.service /etc/systemd/system/

This application provides a simple user authentication mechanism. You can change the username and password by modifying the following part in /opt/sketch_demo/auth_decorator.py.

USERNAME = 'username'
PASSWORD = 'passw0rd'

Launch the demo application

# systemctl daemon-reload
# systemctl enable sketch_demo
# systemctl start sketch_demo
# systemctl status sketch_demo

The last command outputs the application status, as in the following example:

● sketch_demo.service - Sketch-RNN demo
   Loaded: loaded (/etc/systemd/system/sketch_demo.service; enabled; vendor preset: enabled)
   Active: active (running) since Sat 2017-10-21 05:59:12 UTC; 10s ago
 Main PID: 2049 (start_app.sh)
    Tasks: 10 (limit: 4915)
   CGroup: /system.slice/sketch_demo.service
           ├─2049 /bin/bash /opt/sketch_demo/start_app.sh
           ├─2050 /usr/bin/python /opt/sketch_demo/backend.py -p 8081 -d /opt/sketch_demo/models/catbu
           ├─2051 /usr/bin/python /opt/sketch_demo/backend.py -p 8082 -d /opt/sketch_demo/models/eleph
           ├─2052 /usr/bin/python /opt/sketch_demo/backend.py -p 8083 -d /opt/sketch_demo/models/flami
           ├─2053 /usr/bin/python /opt/sketch_demo/backend.py -p 8084 -d /opt/sketch_demo/models/owl/l
           └─2054 /usr/bin/python /opt/sketch_demo/app.py

Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Input dropout mode = 0.
Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Output dropout mode = 0.
Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Recurrent dropout mode = 0.
Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Input dropout mode = 0.
Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Output dropout mode = 0.
Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Recurrent dropout mode = 0.
Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Model using gpu.
Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Input dropout mode = 0.
Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Output dropout mode = 0.
Oct 21 05:59:17 sketch-demo start_app.sh[2049]: INFO:tensorflow:Recurrent dropout mode = 0.

You have to wait around 60secs for the application to finish loading the pre-trained model graph. After that, you can access the instance's static IP address using a web browser. You draw a sample picture on the white canvas, and when you submit it, three children (emulated by a machine learning model) try to imitate your sample.

There are four classrooms and the machine learning model in each class is tranied with a different dataset. As a result, the children in each classroom tend to draw some specific objects such as cats and buses. You can choose the classroom from the buttons on the screen.

• Note: thie example application is tested with only Chrome browser. It is recommended that you would access the application with Chrome browser.