Practical machine learning for everyday web apps

Making TensorFlow the brain of your Django app.

A new wave of machine learning is in full swing. This talk gives an overview of the modern Python machine learning stack based on TensorFlow and my practical experiences from using it in a typical Django web app.

You've probably heard of the recent buzz surrounding deep learning, self-driving cars, Amazon Echo's speech interface, Google DeepMind's AlphaGo program beating a human Go master… The modern applications of machine learning are exploding! But can you benefit from the advances in AI in your everyday web apps? How much data do you need to be able to solve some concrete classification problems? What sort of machinery do you need to run it?

This talk will give an overveiw of Google's recently open sourced TensorFlow library and how it can be used for machine learning. To keep things grounded to realistic problems, we will go through an example Django web app for finding events where we want to classify images based on their content. We'll show how to train the desired image classifier using TensorFlow and use it to classify unknown images. We'll also cover the sort of infrastructure you need to use TensorFlow and give an overview of the available cloud services with specialised hardware support for high performance use cases.

PyDays, 6.5.2017.
https://cfp.linuxwochen.at/de/LWW17/public/events/594

Transcript

1. PRACTICAL MACHINE LEARNING
   FOR EVERYDAY WEB APPS
   Making TensorFlow the brain of your Django app
   Dražen Lučanin
   @metakermit
   

   View Slide

2. THE INTERNET IS FOR CAT PHOTOS!
   

   View Slide

3. IS THAT MIFFLES IN YOUR PHOTO?
   

   View Slide

4. AI: THE NEXT BIG THING™
   ●
   Communication was the Internet’s killer app
   – web apps storing data to DBs ruled
   – Social networks, online stores, messaging apps, productivity apps, online courses, …
   ●
   My view… AI will be the next killer app
   – Faster & cheaper GPU hardware
   – Lots of R&D around machine learning
   – Cool applications
   ●
   Self-driving cars
   ●
   Good speech recognition
   ●
   Automating repetitive manual tasks
   ●
   “The secret sauce”
   

   View Slide

5. GETTING INTO AI
   ●
   …probably a good idea
   ●
   Applying AI != researching AI
   ●
   Modern AI frameworks
   – Torch (Facebook)
   – Theano (academy)
   – TensorFlow (Google)
   

   View Slide

6. TENSORFLOW (TF)
   ●
   Great AI framework built in Google
   – Easy for developers and researchers
   – Production-ready
   ●
   MapReduce
   – White paper only
   – Hadoop became the standard
   ●
   TF open sourced to became the standard
   ●
   Model marketplace
   

   View Slide

7. TF OVERVIEW
   ●
   DataFlow programming language
   ●
   describe a graph of interacting operations that run
   entirely outside Python
   – Graph
   – Session
   ●
   Abstraction levels
   – Low-level API (for researchers)
   – High-level API (GTD)
   

   View Slide

8. LOW-LEVEL API
   import numpy as np
   import tensorflow as tf
   W = tf.Variable([.3], tf.float32)
   b = tf.Variable([-.3], tf.float32)
   x = tf.placeholder(tf.float32); y = tf.placeholder(tf.float32)
   linear_model = W * x + b
   loss = tf.reduce_sum(tf.square(linear_model - y))
   optimizer = tf.train.GradientDescentOptimizer(0.01)
   train = optimizer.minimize(loss)
   x_train = [1,2,3,4]
   y_train = [0,-1,-2,-3]
   init = tf.global_variables_initializer()
   sess = tf.Session()
   sess.run(init)
   for i in range(1000):
   sess.run(train, {x:x_train, y:y_train})
   curr_W, curr_b, curr_loss = sess.run([W, b, loss], {x:x_train, y:y_train})
   print("W: %s b: %s loss: %s"%(curr_W, curr_b, curr_loss))
   

   View Slide

9. EVENTS EXAMPLE
   ●
   AI for event discovery
   – Django web app ( https://www.posterbat.com )
   – Side-project I’m working on
   ●
   Baby steps
   – Scrapped some event cover images from the web
   – Does an image have text in it?
   ●
   AI can be used on everyday problems
   ●
   Not only for cutting edge research problems
   – e.g. speech recognition
   

   View Slide

10. GETTING DATA FROM DJANGO
    $ ./manage.py shell -c 'from export_images import run; run()'
    from events.models import Event
    def run():
    for event in Event.objects.all():
    prepare_event_image(event)
    

    View Slide

11. NORMALISING IMAGES USING PIL
    from PIL import Image
    def prepare_event_image(event):
    with open(f'./images/{event.id}.png', 'wb') as f:
    size = (256, 256)
    try:
    image = Image.open(event.img)
    except ValueError:
    return
    image.thumbnail(size, Image.ANTIALIAS)
    region = image.crop((0, 0, *size))
    region.save(f, 'png')
    

    View Slide

12. LABELING CLASSES
    

    View Slide

13. IMPORTING DATA INTO TF
    labels = pd.read_csv('./labels.csv', index_col=0)
    def read_data(folder):
    path = './data/images/' + folder + '/'
    x = []; y = []
    for filename in os.listdir(path):
    image_id = int(filename.split('.')[0])
    # convert input 256x256 image to grayscale
    # flatten to a 1-d array of floats (0-255)
    im = misc.imread(
    path + filename,flatten=True
    ).flatten()
    x.append(im)
    y.append(int(labels.ix[image_id]))
    x = np.array(x)
    return tf.constant(x), tf.constant(y)
    train, train_labels = read_data('train')
    

    View Slide

14. TRAINING THE MODEL
    feature_columns = [tf.contrib.layers.real_valued_column("",
    dimension=65536)]
    classifier = tf.contrib.learn.DNNClassifier(
    feature_columns=feature_columns,
    hidden_units=[1024, 512, 256],
    n_classes=2,
    model_dir="/tmp/model_dir",
    )
    classifier.fit(
    x=train, y=train_labels, steps=20000
    )
    

    View Slide

15. LOAD THE MODEL IN DJANGO
    ●
    Save
    ●
    Add the model directory to your code repository
    ●
    Load in Django
    saver = tf.train.Saver()
    saver.save(session, 'my-model')
    new_saver = tf.train.import_meta_graph('events/my-model.meta')
    new_saver.restore(session, 'events/my-model')
    

    View Slide

16. APPLY THE MODEL
    def get_new_img():
    x = []
    img_path = 'image.png'
    im = misc.imread(img_path, flatten=True).flatten()
    x.append(im)
    x = np.array(x)
    return x
    classifier.predict(input_fn=get_new_img, as_iterable=False)
    

    View Slide

17. TENSORBOARD – MONITORING
    ●
    Open http://localhost:6006
    ●
    Monitor training at runtime
    $ tensorboard --logdir /tmp/model_dir
    

    View Slide

18. TENSORBOARD – GRAPHS
    

    View Slide

19. TENSORBOARD – CLASSES
    

    View Slide

20. TENSORBOARD – CLASSES
    

    View Slide

21. PERFORMANCE
    ●
    CPU (C++ implementation – pretty efficient)
    ●
    GPU – even faster!
    ●
    JIT compiler
    – Speed things up by adding a single line of code
    – Experimental
    ●
    XLA compiler
    – Ahead-of time compilation
    – Run on embedded devices (phones, IoT)
    

    View Slide

22. DEPLOYMENT
    ●
    Google Cloud & AWS offer VMs with GPUs
    ●
    FloydHub
    – Heroku for AI
    – https://www.floydhub.com/
    Provider AWS Google Floyd
    Cost per
    hour ($) 0.99 0.795 0.432
    

    View Slide

23. LEARNING
    ●
    Easy riding
    – https://changelog.com/podcast/219
    – TF Dev Summit ‘17 videos – https://events.withgoogle.com/tensorflow-dev-summit/
    ●
    Docs & tutorial
    – https://www.tensorflow.org/get_started/get_started
    – https://medium.freecodecamp.com/big-picture-machine-learning-classifying-text-with-
    neural-networks-and-tensorflow-d94036ac2274
    ●
    Good free books
    – ESL – http://statweb.stanford.edu/~tibs/ElemStatLearn/
    – Michael Nielsen – http://neuralnetworksanddeeplearning.com/
    ●
    Research
    – http://distill.pub/
    

    View Slide

24. THANKS!
    ●
    Dražen Lučanin
    ●
    @metakermit
    ●
    Building apps with a kick!
    https://punkrockdev.com/
    

    View Slide