FizzBuzz with keras, because why not

fizzbuzz_keras.py

# These are the things we will use. Implicitly, we'll assume tensorflow is the keras backend
import keras
import keras.backend as K
import numpy as np
from keras.models import Sequential
from keras.layers import Dense

# This is necessary only because of some version incompatibility where the default
# argument to argmax is broken with some combination of Python 3.7, keras 2.1.3 (or 2.1.6),
# and tensorflow 1.12. So we'll just reimplement it until the demons of dependency hell
# are vanquished once and for all
def get_categorical_accuracy_keras(y_true, y_pred):
    return K.mean(K.equal(K.argmax(y_true, axis=1), K.argmax(y_pred, axis=1)))

# we need to do feature extraction, so what features does an integer have?
# clearly, we will take the modulus of an integer from every value between 1 and 15
# this makes complete sense and isn't absurd at all
def extract_features(i):
    return [i % (k + 1) for k in range(15)]

# We'll generate our own truth, because it's 2018 and that's how the world works
def fizzbuzz(i):
    # fizzbuzz is class 3
    if i % 15 == 0:
        return 3
    # buzz is class 2
    if i % 5 == 0:
        return 2
    # fizz is class 1
    if i % 3 == 0:
        return 1
    # fallthrough is class 0
    return 0
  
# generate the training data and the labels
# obviously we don't want to train on the numbers between 1 and 100, so let's pick an
# arbitrary range. 500 to 2500 sounds excellent
data = np.array([extract_features(i) for i in range(500,2500)])
labels = np.array([fizzbuzz(i) for i in range(500,2500)])
dim = np.shape(data)[1]

# Convert labels to categorical one-hot encoding, as this is what keras expects
one_hot_labels = keras.utils.to_categorical(labels)

# specify the network
# why these network parameters? NO, that is the wrong question. The right question
# is "why NOT these network parameters?"
model = Sequential()
model.add(Dense(10, activation='relu', input_dim=dim))
model.add(Dense(4, activation='softmax'))
model.compile(optimizer='rmsprop',
              loss='categorical_crossentropy',
              metrics=[get_categorical_accuracy_keras])

# Train the model, iterating on the data in batches of 64 samples
model.fit(data, one_hot_labels, epochs=50, batch_size=64, shuffle=True)

# Uncomment this block if you want to run things in jupyter to see how things did
# score = model.evaluate(data, one_hot_labels, batch_size=64)
# print(score)
# model.metrics_names

# Finally, we will see if we pass the interview!

def classify(i):
    prediction = model.predict(np.array([extract_features(i)]))
    result = np.argmax(prediction)
    if result == 1:
        return 'fizz'
    if result == 2:
        return 'buzz'
    if result == 3:
        return 'fizzbuzz'
    return str(i)

[classify(i) for i in range(1, 100 + 1)]