"""
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    - Fast Food layer where Fast Hadamard Transform has been replaced by dot product with Hadamard matrix.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
from skluc.neural_networks import convolution_mnist, classification_mnist, batch_generator
from skluc.kernel_approximation.fasfood_layer import fast_food

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

import time as t

val_size = 5000
mnist = dataset.MnistDataset(validation_size=val_size)
mnist.load()
mnist.normalize()
mnist.to_one_hot()
mnist.data_astype(np.float32)
mnist.labels_astype(np.float32)
X_train, Y_train = mnist.train
X_test, Y_test = mnist.test
X_val, Y_val = mnist.validation


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

    with tf.Graph().as_default():
        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)
        # 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, nbr_stack=1, trainable=True)  # 84% accuracy (conv) | 59% accuracy (noconv)
        # out_fc = random_features(h_conv, SIGMA)  # 82% accuracy (conv) | 47% accuracy (noconv)

        # classification
        y_conv, keep_prob = classification_mnist(out_fc, output_dim)

        # 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()
        feed_dict_val = {x: X_val, y_: Y_val, keep_prob: 1.0}
        for _ in range(1):
            i = 0
            for X_batch, Y_batch in batch_generator(X_train, Y_train, 64, circle=True):
                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, y_result, x_exp = sess.run([train_optimizer, cross_entropy, y_conv, x_image], feed_dict=feed_dict)
                if i % 100 == 0:
                    print('step {}, loss {} (with dropout)'.format(i, loss))
                    r_accuracy = sess.run([accuracy], feed_dict=feed_dict_val)
                    print("accuracy: {} on validation set (without dropout).".format(r_accuracy))
                    summary_str = sess.run(merged_summary, feed_dict=feed_dict)
                    summary_writer.add_summary(summary_str, i)
                i += 1

        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))