deepfriedConvnetMnist.py

"""
Convolutional Neural Netwok implementation in tensorflow whith multiple representations possible after the convolution:
    - Fully connected layer
    - Random Fourier Features layer
    - Fast Food layer where Fast Hadamard Transform has been replaced by dot product with Hadamard matrix.

See:
"Deep Fried Convnets" by
Zichao Yang, Marcin Moczulski, Misha Denil, Nando de Freitas, Alex Smola, Le Song, Ziyu Wang

"""

import tensorflow as tf
import numpy as np
import skluc.mldatasets as dataset

tf.logging.set_verbosity(tf.logging.ERROR)

import time as t

from sklearn.preprocessing import LabelBinarizer

enc = LabelBinarizer()
mnist = dataset.MnistDataset()
mnist = mnist.load()
X_train, Y_train = mnist["train"]
X_train = np.array(X_train / 255)
enc.fit(Y_train)
Y_train = np.array(enc.transform(Y_train))
X_test, Y_test = mnist["test"]
X_test = np.array(X_test / 255)
Y_test = np.array(enc.transform(Y_test))
X_train = X_train.astype(np.float32)
X_test = X_test.astype(np.float32)
Y_train = Y_train.astype(np.float32)
Y_test = Y_test.astype(np.float32)

from fasfood_layer import fast_food

# --- Usual functions --- #

def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial, name="weights")


def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial, name="biases")


def conv2d(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')


def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                          strides=[1, 2, 2, 1], padding='SAME')


def convolution_mnist(input):
    with tf.name_scope("conv_pool_1"):
        # 32 is the number of filter we'll use. e.g. the number of different
        # shapes this layer is able to recognize
        W_conv1 = weight_variable([5, 5, 1, 20])
        tf.summary.histogram("weights conv1", W_conv1)
        b_conv1 = bias_variable([20])
        tf.summary.histogram("biases conv1", b_conv1)
        # -1 is here to keep the total size constant (784)
        h_conv1 = tf.nn.relu(conv2d(input, W_conv1) + b_conv1)
        tf.summary.histogram("act conv1", h_conv1)
        h_pool1 = max_pool_2x2(h_conv1)

    with tf.name_scope("conv_pool_2"):
        W_conv2 = weight_variable([5, 5, 20, 50])
        tf.summary.histogram("weights conv2", W_conv2)
        b_conv2 = bias_variable([50])
        tf.summary.histogram("biases conv2", b_conv2)
        h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
        tf.summary.histogram("act conv2", h_conv2)
        h_pool2 = max_pool_2x2(h_conv2)

    return h_pool2


# --- Random Fourier Features --- #

def random_variable(shape, sigma):
    W = np.random.normal(size=shape, scale=sigma).astype(np.float32)
    return tf.Variable(W, name="random_Weights", trainable=False)


def random_biases(shape):
    b = np.random.uniform(0, 2 * np.pi, size=shape).astype(np.float32)
    return tf.Variable(b, name="random_biase", trainable=False)


# --- Representation Layer --- #

def random_features(conv_out, sigma):
    with tf.name_scope("random_features"):
        init_dim = np.prod([s.value for s in conv_out.shape if s.value is not None])
        conv_out2 = tf.reshape(conv_out, [-1, init_dim])

        W = random_variable((init_dim, init_dim), sigma)
        b = random_biases(init_dim)
        h1 = tf.matmul(conv_out2, W, name="Wx") + b
        h1_cos = tf.cos(h1)
        h1_final = tf.scalar_mul(np.sqrt(2.0 / init_dim).astype(np.float32), h1_cos)
        return h1_final


def fully_connected(conv_out):
    with tf.name_scope("fc_1"):
        h_pool2_flat = tf.reshape(conv_out, [-1, 7 * 7 * 64])
        W_fc1 = weight_variable([7 * 7 * 64, 4096*2])
        b_fc1 = bias_variable([4096*2])
        h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
        tf.summary.histogram("weights", W_fc1)
        tf.summary.histogram("biases", b_fc1)

    return h_fc1


def get_next_batch(full_set, batch_nbr, batch_size):
    """
    Return the next batch of a dataset.

    This function assumes that all the previous batches of this dataset have been taken with the same size.

    :param full_set: the full dataset from which the batch will be taken
    :param batch_nbr: the number of the batch
    :param batch_size: the size of the batch
    :return:
    """
    index_start = (batch_nbr * batch_size) % full_set.shape[0]
    index_stop = ((batch_nbr + 1) * batch_size) % full_set.shape[0]
    if index_stop > index_start:
        return full_set[index_start:index_stop]
    else:
        part1 = full_set[index_start:]
        part2 = full_set[:index_stop]
        return np.vstack((part1, part2))


if __name__ == '__main__':
    SIGMA = 5.0
    print("Sigma = {}".format(SIGMA))

    with tf.Graph().as_default():
        # todo parametrize dataset
        input_dim, output_dim = X_train.shape[1], Y_train.shape[1]

        x = tf.placeholder(tf.float32, shape=[None, input_dim], name="x")
        y_ = tf.placeholder(tf.float32, shape=[None, output_dim], name="labels")

        # side size is width or height of the images
        side_size = int(np.sqrt(input_dim))
        x_image = tf.reshape(x, [-1, side_size, side_size, 1])
        tf.summary.image("digit", x_image, max_outputs=3)

        # Representation layer
        h_conv = convolution_mnist(x_image)
        # h_conv = x
        # out_fc = fully_connected(h_conv)  # 95% accuracy
        # out_fc = tf.nn.relu(fast_food(h_conv, SIGMA, nbr_stack=1))  # 83% accuracy (conv) | 56% accuracy (noconv)
        out_fc = tf.nn.relu(fast_food(h_conv, SIGMA, nbr_stack=2))
        # out_fc = tf.nn.relu(fast_food(h_conv, SIGMA, nbr_stack=2, trainable=True))
        # out_fc = tf.nn.relu(fast_food(h_conv, SIGMA, trainable=True))  # 84% accuracy (conv) | 59% accuracy (noconv)
        # out_fc = fast_food(h_conv, SIGMA, diag=True, trainable=True)  # 84% accuracy (conv) | 59% accuracy (noconv)
        # out_fc = random_features(h_conv, SIGMA)  # 82% accuracy (conv) | 47% accuracy (noconv)

        # classification
        with tf.name_scope("fc_2"):
            keep_prob = tf.placeholder(tf.float32, name="keep_prob")
            h_fc1_drop = tf.nn.dropout(out_fc, keep_prob)
            dim = np.prod([s.value for s in h_fc1_drop.shape if s.value is not None])
            W_fc2 = weight_variable([dim, output_dim])
            b_fc2 = bias_variable([output_dim])
            tf.summary.histogram("weights", W_fc2)
            tf.summary.histogram("biases", b_fc2)

            y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

        # calcul de la loss
        with tf.name_scope("xent"):
            cross_entropy = tf.reduce_mean(
                tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv, name="xentropy"),
                name="xentropy_mean")
            tf.summary.scalar('loss-xent', cross_entropy)

        # calcul du gradient
        with tf.name_scope("train"):
            global_step = tf.Variable(0, name="global_step", trainable=False)
            train_optimizer = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(cross_entropy, global_step=global_step)

        # calcul de l'accuracy
        with tf.name_scope("accuracy"):
            predictions = tf.argmax(y_conv, 1)
            correct_prediction = tf.equal(predictions, tf.argmax(y_, 1))
            accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
            tf.summary.scalar("accuracy", accuracy)

        merged_summary = tf.summary.merge_all()

        init = tf.global_variables_initializer()
        # Create a session for running Ops on the Graph.
        sess = tf.Session()
        # Instantiate a SummaryWriter to output summaries and the Graph.
        summary_writer = tf.summary.FileWriter("results_deepfried_stacked")
        summary_writer.add_graph(sess.graph)
        # Initialize all Variable objects
        sess.run(init)
        # actual learning
        started = t.time()
        for i in range(1100):
            X_batch = get_next_batch(X_train, i, 64)
            Y_batch = get_next_batch(Y_train, i, 64)
            feed_dict = {x: X_batch, y_: Y_batch, keep_prob: 0.5}
            # le _ est pour capturer le retour de "train_optimizer" qu'il faut appeler
            # pour calculer le gradient mais dont l'output ne nous interesse pas
            _, loss = sess.run([train_optimizer, cross_entropy], feed_dict=feed_dict)
            if i % 100 == 0:
                print('step {}, loss {} (with dropout)'.format(i, loss))
                summary_str = sess.run(merged_summary, feed_dict=feed_dict)
                summary_writer.add_summary(summary_str, i)

        stoped = t.time()
        accuracy, preds = sess.run([accuracy, predictions], feed_dict={
            x: X_test, y_: Y_test, keep_prob: 1.0})
        print('test accuracy %g' % accuracy)
        np.set_printoptions(threshold=np.nan)
        print("Prediction sample: " + str(preds[:50]))
        print("Actual values: " + str(np.argmax(Y_test[:50], axis=1)))
        print("Elapsed time: %.4f s" % (stoped - started))