Running experiments to improve accuracy without grid-search or alike.

epochs.py

import argparse

import matplotlib.pyplot as plt
import numpy as np
from keras.layers.core import Dense
from keras.models import Sequential
from numpy import array
from scipy import signal


def parse_arguments():
    parser = argparse.ArgumentParser()

    parser.add_argument('--epochs', type=int, nargs='+',
                        help='The epoches to run', required=True)
    parser.add_argument('--mode', help='The mode: azure or python',
                        choices=['python', 'azure'], default='python')
    args = parser.parse_args()

    # will always be there since it's a required argument
    epochs = np.array(args.epochs)
    epochs_string = ' '.join(str(x) for x in epochs)
    return epochs

def main():
    epochs = parse_arguments()
    if len(epochs) == 0:
        print("Nothing to do: no epochs given.")
    else:
        accs = list()
        for i in epochs:
            print('Running with %s epochs.' % i)
            exp = Experiment()
            acc = exp.learn(i)
            accs.append(acc)

        plt.plot(epochs, accs, alpha=0.5)
        plt.show()

class Experiment():

    def __init__(self):
        self.N = 10
        self.model = self.create_model()
        self.X, self.y = self.create_data(1000)
        self.X_test, self.y_test = self.create_test()

    def create_model(self):
        model = Sequential()
        model.add(Dense(1, input_shape=(1, self.N)))
        model.add(Dense(10))
        model.add(Dense(1))
        model.compile(optimizer='adam', loss='mean_squared_error', metrics=['acc'])
        return model

    def create_data(self, n):
        X = list()
        y = list()
        for i in range(n):
            X.append(array([array(signal.unit_impulse(self.N, i % self.N))]))
            y.append(array([array([i % self.N])]))
        return array(X), array(y)

    def create_test(self):
        a = array([0, 0, 4, 5, 5, 6, 8, 4, 1, 4, 9, 8, 1, 1, 7, 9, 9, 3, 6, 7])
        X = list()
        y = list()
        for i in a:
            X.append(array([array(signal.unit_impulse(self.N, i % self.N))]))
            y.append(array([array([i % self.N])]))
        return array(X), array(y)

    def learn(self, epochs):

        self.model.fit(self.X, self.y, epochs=epochs, verbose=0)
        e = self.model.evaluate(self.X_test, self.y_test, verbose=0)
        print("%s: %.0f%%" % (self.model.metrics_names[1], e[1] * 100))
        return e[1]



# python run_basic.py --epochs 5 10 15 20 25 30 40 50
if __name__ == "__main__":
    main()