100% accuracy on example from AI Stack Exchange question

example_dnn.py

import numpy as np
import matplotlib.pyplot as plt

from keras.models import Sequential
from keras.layers import Dense, Activation
from keras.optimizers import SGD
from keras.utils import to_categorical

def get_data():
    X = np.array([[1,2], [1,12], [1,17], [9,33], [48,49], [48,50]])
    y = [ 0, 0, 0, 0, 1, 2 ]
    X = (X/25) - 1
    # This copies indetical data to CV and test data sets, because question is only about fitting training data
    # Don't do this for any real data, otherwise you won't know how well your NN is doing!
    return (X, y, X, y, X, y)

def build_model(layer_sizes, hidden_activation='relu', learning_rate=0.005, momentum=0.9):
    model = Sequential()
    model.add( Dense(layer_sizes[1], input_dim=layer_sizes[0]) )

    for size in layer_sizes[2:]:
        model.add( Activation(hidden_activation) )
        model.add( Dense(size) )

    model.add(Activation('softmax'))

    sgd = SGD(lr=learning_rate, momentum=momentum, nesterov=True)
    model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])

    return model

def train_model(model, X_train, y_train, X_cv, y_cv, epochs=20, batch_size=6):
    y_train = to_categorical(y_train)
    y_cv = to_categorical(y_cv)
    history = model.fit(X_train, y_train, validation_data=(X_cv, y_cv), epochs=epochs, batch_size=batch_size, verbose=0)
    return history.history

def test_model(model, X_test, y_test):
    y_pred = model.predict_classes(X_test, verbose=0)
    accuracy = np.mean( (y_pred == y_test).astype('float') )
    print("Final accuracy {}".format(accuracy * 100))

def plot_history(history):
    plt.plot(history['acc'])
    plt.plot(history['val_acc'])
    plt.title('Accuracy')
    plt.ylabel('Accuracy vs Epoch')
    plt.xlabel('Epoch')
    plt.legend(['train', 'val'], loc='lower right')
    plt.show()

def main():
    np.random.seed(1234)
    X_train, y_train, X_cv, y_cv, X_test, y_test = get_data()

    model = build_model( layer_sizes=[X_train.shape[1], 20, 10, 3],
                         hidden_activation='tanh',
                         learning_rate=0.01,
                         momentum=0.9 )

    history = train_model(model, X_train, y_train, X_cv, y_cv, epochs=500)
    test_model(model, X_test, y_test)
    plot_history(history)

if __name__ == '__main__':
    main()

This is just a hacky script, adapted from other code I had lying around. It is in response to this question: https://ai.stackexchange.com/questions/14099/tensorflow-estimator-dnnclassifier-fails-to-fit-simple-data

tks a lot neil :)