# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

"""Simple, end-to-end, LeNet-5-like convolutional MNIST model example.
This should achieve a test error of 0.7%. Please keep this model as simple and
linear as possible, it is meant as a tutorial for simple convolutional models.
Run with --self_test on the command line to execute a short self-test.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import gzip
import os
import sys
import time

from six.moves import urllib
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.python.ops import control_flow_ops

SOURCE_URL = 'http://yann.lecun.com/exdb/mnist/'
WORK_DIRECTORY = 'data'
IMAGE_SIZE = 28
NUM_CHANNELS = 1
PIXEL_DEPTH = 255
NUM_LABELS = 10
VALIDATION_SIZE = 5000  # Size of the validation set.
SEED = 66478  # Set to None for random seed.
BATCH_SIZE = 64
NUM_EPOCHS = 10
EVAL_BATCH_SIZE = 64
EVAL_FREQUENCY = 100  # Number of steps between evaluations.


tf.app.flags.DEFINE_boolean("self_test", False, "True if running a self test.")
tf.app.flags.DEFINE_boolean('use_fp16', False,
                            "Use half floats instead of full floats if True.")
FLAGS = tf.app.flags.FLAGS


def data_type():
    """Return the type of the activations, weights, and placeholder variables."""
    if FLAGS.use_fp16:
        return tf.float16
    else:
        return tf.float32


def maybe_download(filename):
    """Download the data from Yann's website, unless it's already here."""
    if not tf.gfile.Exists(WORK_DIRECTORY):
        tf.gfile.MakeDirs(WORK_DIRECTORY)
    filepath = os.path.join(WORK_DIRECTORY, filename)
    if not tf.gfile.Exists(filepath):
        filepath, _ = urllib.request.urlretrieve(SOURCE_URL + filename, filepath)
        with tf.gfile.GFile(filepath) as f:
            size = f.size()
        print('Successfully downloaded', filename, size, 'bytes.')
    return filepath


def extract_data(filename, num_images):
    """Extract the images into a 4D tensor [image index, y, x, channels].
    Values are rescaled from [0, 255] down to [-0.5, 0.5].
    """
    print('Extracting', filename)
    with gzip.open(filename) as bytestream:
        bytestream.read(16)
        buf = bytestream.read(IMAGE_SIZE * IMAGE_SIZE * num_images * NUM_CHANNELS)
        data = np.frombuffer(buf, dtype=np.uint8).astype(np.float32)
        data = (data - (PIXEL_DEPTH / 2.0)) / PIXEL_DEPTH
        data = data.reshape(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS)
        return data


def extract_labels(filename, num_images):
    """Extract the labels into a vector of int64 label IDs."""
    print('Extracting', filename)
    with gzip.open(filename) as bytestream:
        bytestream.read(8)
        buf = bytestream.read(1 * num_images)
        labels = np.frombuffer(buf, dtype=np.uint8).astype(np.int64)
    return labels


def fake_data(num_images):
    """Generate a fake dataset that matches the dimensions of MNIST."""
    data = np.ndarray(
        shape=(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS),
        dtype=np.float32)
    labels = np.zeros(shape=(num_images,), dtype=np.int64)
    for image in range(num_images):
        label = image % 2
        data[image, :, :, 0] = label - 0.5
        labels[image] = label
    return data, labels


def error_rate(predictions, labels):
    """Return the error rate based on dense predictions and sparse labels."""
    return 100.0 - (
        100.0 *
        np.sum(np.argmax(predictions, 1) == labels) /
        predictions.shape[0])


def main(argv=None):  # pylint: disable=unused-argument
    if FLAGS.self_test:
        print('Running self-test.')
        train_data, train_labels = fake_data(256)
        validation_data, validation_labels = fake_data(EVAL_BATCH_SIZE)
        test_data, test_labels = fake_data(EVAL_BATCH_SIZE)
        num_epochs = 1
    else:
        # Get the data.
        train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
        train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
        test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
        test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')
    
        # Extract it into numpy arrays.
        train_data = extract_data(train_data_filename, 60000)
        train_labels = extract_labels(train_labels_filename, 60000)
        test_data = extract_data(test_data_filename, 10000)
        test_labels = extract_labels(test_labels_filename, 10000)
    
        # Generate a validation set.
        validation_data = train_data[:VALIDATION_SIZE, ...]
        validation_labels = train_labels[:VALIDATION_SIZE]
        train_data = train_data[VALIDATION_SIZE:, ...]
        train_labels = train_labels[VALIDATION_SIZE:]
        num_epochs = NUM_EPOCHS
    train_size = train_labels.shape[0]

    # This is where training samples and labels are fed to the graph.
    # These placeholder nodes will be fed a batch of training data at each
    # training step using the {feed_dict} argument to the Run() call below.
    train_data_node = tf.placeholder(
        data_type(),
        shape=(BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
    train_labels_node = tf.placeholder(tf.int64, shape=(BATCH_SIZE,))
    eval_data = tf.placeholder(
        data_type(),
        shape=(EVAL_BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))

    # The variables below hold all the trainable weights. They are passed an
    # initial value which will be assigned when we call:
    # {tf.initialize_all_variables().run()}
    conv1_weights = tf.Variable(
        tf.truncated_normal([5, 5, NUM_CHANNELS, 32],  # 5x5 filter, depth 32.
                            stddev=0.1,
                            seed=SEED, dtype=data_type()))
    conv1_biases = tf.Variable(tf.zeros([32], dtype=data_type()))
    conv2_weights = tf.Variable(tf.truncated_normal(
        [5, 5, 32, 64], stddev=0.1,
        seed=SEED, dtype=data_type()))
    conv2_biases = tf.Variable(tf.constant(0.1, shape=[64], dtype=data_type()))
    fc1_weights = tf.Variable(  # fully connected, depth 512.
        tf.truncated_normal([IMAGE_SIZE // 4 * IMAGE_SIZE // 4 * 64, 512],
                            stddev=0.1,
                            seed=SEED,
                            dtype=data_type()))
    fc1_biases = tf.Variable(tf.constant(0.1, shape=[512], dtype=data_type()))
    fc1p_weights = tf.Variable(  # fully connected, depth 512.
        tf.truncated_normal([512, 2],
                            stddev=0.1,
                            seed=SEED,
                            dtype=data_type()))
    fc1p_biases = tf.Variable(tf.constant(0.1, shape=[2], dtype=data_type()))
    fc2_weights = tf.Variable(tf.truncated_normal([2, NUM_LABELS],
                                                  stddev=0.1,
                                                  seed=SEED,
                                                  dtype=data_type()))
    fc2_biases = tf.Variable(tf.constant(
        0.1, shape=[NUM_LABELS], dtype=data_type()))
    
    def batch_norm(x, phase_train):  #pylint: disable=unused-variable
        """
        Batch normalization on convolutional maps.
        Args:
            x:           Tensor, 4D BHWD input maps
            n_out:       integer, depth of input maps
            phase_train: boolean tf.Variable, true indicates training phase
            scope:       string, variable scope
            affn:      whether to affn-transform outputs
        Return:
            normed:      batch-normalized maps
        Ref: http://stackoverflow.com/questions/33949786/how-could-i-use-batch-normalization-in-tensorflow/33950177
        """
        name = 'batch_norm'
        with tf.variable_scope(name):
            phase_train = tf.convert_to_tensor(phase_train, dtype=tf.bool)
            n_out = int(x.get_shape()[-1])
            beta = tf.Variable(tf.constant(0.0, shape=[n_out], dtype=x.dtype),
                               name=name+'/beta', trainable=True, dtype=x.dtype)
            gamma = tf.Variable(tf.constant(1.0, shape=[n_out], dtype=x.dtype),
                                name=name+'/gamma', trainable=True, dtype=x.dtype)
          
            batch_mean, batch_var = tf.nn.moments(x, [0], name='moments')
            ema = tf.train.ExponentialMovingAverage(decay=0.9)
            def mean_var_with_update():
                ema_apply_op = ema.apply([batch_mean, batch_var])
                with tf.control_dependencies([ema_apply_op]):
                    return tf.identity(batch_mean), tf.identity(batch_var)
            mean, var = control_flow_ops.cond(phase_train,
                                              mean_var_with_update,
                                              lambda: (ema.average(batch_mean), ema.average(batch_var)))
            normed = tf.nn.batch_normalization(x, mean, var, beta, gamma, 1e-3)
        return normed
    

    # We will replicate the model structure for the training subgraph, as well
    # as the evaluation subgraphs, while sharing the trainable parameters.
    def model(data, train=False):
        """The Model definition."""
        # 2D convolution, with 'SAME' padding (i.e. the output feature map has
        # the same size as the input). Note that {strides} is a 4D array whose
        # shape matches the data layout: [image index, y, x, depth].
        conv = tf.nn.conv2d(data,
                            conv1_weights,
                            strides=[1, 1, 1, 1],
                            padding='SAME')
        # Bias and rectified linear non-linearity.
        relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases))
        # Max pooling. The kernel size spec {ksize} also follows the layout of
        # the data. Here we have a pooling window of 2, and a stride of 2.
        pool = tf.nn.max_pool(relu,
                              ksize=[1, 2, 2, 1],
                              strides=[1, 2, 2, 1],
                              padding='SAME')
        conv = tf.nn.conv2d(pool,
                            conv2_weights,
                            strides=[1, 1, 1, 1],
                            padding='SAME')
        relu = tf.nn.relu(tf.nn.bias_add(conv, conv2_biases))
        pool = tf.nn.max_pool(relu,
                              ksize=[1, 2, 2, 1],
                              strides=[1, 2, 2, 1],
                              padding='SAME')
        # Reshape the feature map cuboid into a 2D matrix to feed it to the
        # fully connected layers.
        pool_shape = pool.get_shape().as_list() #pylint: disable=no-member
        reshape = tf.reshape(
            pool,
            [pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]])
        # Fully connected layer. Note that the '+' operation automatically
        # broadcasts the biases.
        hidden = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
        # Add a 50% dropout during training only. Dropout also scales
        # activations such that no rescaling is needed at evaluation time.
        if train:
            hidden = tf.nn.dropout(hidden, 0.5, seed=SEED)

        hidden = tf.matmul(hidden, fc1p_weights) + fc1p_biases

        return tf.nn.relu(tf.matmul(hidden, fc2_weights) + fc2_biases), hidden

    def center_loss_op(logits, labels):
        alfa = 0.05
        nrof_features = logits.get_shape()[1]
        centers = tf.get_variable('centers', shape=(nrof_features), dtype=tf.float32,
            initializer=tf.constant_initializer(value=0.0, dtype=tf.float32), trainable=False)
        # Define center loss
        #center_loss = tf.reduce_sum(tf.pow(tf.abs(logits - centers), 2.0))
        center_loss = tf.nn.l2_loss(logits - centers)
        one_hot = tf.one_hot(labels, nrof_features, axis=1, dtype=tf.float32, name='one_hot')
        delta1 = tf.reduce_mean((centers-logits)*one_hot,0)
        delta2 = 1+tf.reduce_mean(one_hot,0)
        centers_delta = delta1 / delta2
        update_centers = tf.assign_add(centers, -alfa*centers_delta)
        return center_loss, update_centers
  
    # Training computation: logits + cross-entropy loss.
    logits, hidden = model(train_data_node, True)
    #logits = batch_norm(logits, True)
    xent_loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
        logits, train_labels_node))
    beta = 1e-6
    center_loss, update_centers = center_loss_op(hidden, train_labels_node)
    loss = xent_loss + beta * center_loss
  
    # L2 regularization for the fully connected parameters.
    regularizers = (tf.nn.l2_loss(fc1_weights) + tf.nn.l2_loss(fc1_biases) +
                    tf.nn.l2_loss(fc2_weights) + tf.nn.l2_loss(fc2_biases))
    # Add the regularization term to the loss.
    loss += 5e-4 * regularizers
  
    # Optimizer: set up a variable that's incremented once per batch and
    # controls the learning rate decay.
    batch = tf.Variable(0, dtype=data_type())
    # Decay once per epoch, using an exponential schedule starting at 0.01.
    learning_rate = tf.train.exponential_decay(
        0.01,                # Base learning rate.
        batch * BATCH_SIZE,  # Current index into the dataset.
        train_size,          # Decay step.
        0.95,                # Decay rate.
        staircase=True)
    # Use simple momentum for the optimization.
    optimizer = tf.train.MomentumOptimizer(learning_rate,
                                           0.9).minimize(loss,
                                                         global_step=batch)
  
    # Predictions for the current training minibatch.
    train_prediction = tf.nn.softmax(logits)
  
    # Predictions for the test and validation, which we'll compute less often.
    eval_logits, eval_embeddings = model(eval_data)
    eval_prediction = tf.nn.softmax(eval_logits)
    
    # Small utility function to evaluate a dataset by feeding batches of data to
    # {eval_data} and pulling the results from {eval_predictions}.
    # Saves memory and enables this to run on smaller GPUs.
    def eval_in_batches(data, sess):
        """Get all predictions for a dataset by running it in small batches."""
        size = data.shape[0]
        if size < EVAL_BATCH_SIZE:
            raise ValueError("batch size for evals larger than dataset: %d" % size)
        predictions = np.ndarray(shape=(size, NUM_LABELS), dtype=np.float32)
        for begin in xrange(0, size, EVAL_BATCH_SIZE):
            end = begin + EVAL_BATCH_SIZE
            if end <= size:
                predictions[begin:end, :] = sess.run(
                    eval_prediction,
                    feed_dict={eval_data: data[begin:end, ...]})
            else:
                batch_predictions = sess.run(
                    eval_prediction,
                    feed_dict={eval_data: data[-EVAL_BATCH_SIZE:, ...]})
                predictions[begin:, :] = batch_predictions[begin - size:, :]
        return predictions
  
    def calculate_embeddings(data, sess):
        """Get all predictions for a dataset by running it in small batches."""
        size = data.shape[0]
        if size < EVAL_BATCH_SIZE:
            raise ValueError("batch size for evals larger than dataset: %d" % size)
        predictions = np.ndarray(shape=(size, 2), dtype=np.float32)
        for begin in xrange(0, size, EVAL_BATCH_SIZE):
            end = begin + EVAL_BATCH_SIZE
            if end <= size:
                predictions[begin:end, :] = sess.run(
                    eval_embeddings,
                    feed_dict={eval_data: data[begin:end, ...]})
            else:
                batch_predictions = sess.run(
                    eval_embeddings,
                    feed_dict={eval_data: data[-EVAL_BATCH_SIZE:, ...]})
                predictions[begin:, :] = batch_predictions[begin - size:, :]
        return predictions

    # Create a local session to run the training.
    start_time = time.time()
    with tf.Session() as sess:
        # Run all the initializers to prepare the trainable parameters.
        tf.initialize_all_variables().run() #pylint: disable=no-member
        print('Initialized!')
        # Loop through training steps.
        for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
            # Compute the offset of the current minibatch in the data.
            # Note that we could use better randomization across epochs.
            offset = (step * BATCH_SIZE) % (train_size - BATCH_SIZE)
            batch_data = train_data[offset:(offset + BATCH_SIZE), ...]
            batch_labels = train_labels[offset:(offset + BATCH_SIZE)]
            # This dictionary maps the batch data (as a numpy array) to the
            # node in the graph it should be fed to.
            feed_dict = {train_data_node: batch_data,
                         train_labels_node: batch_labels}
            # Run the graph and fetch some of the nodes.
            #_, l, lr, predictions = sess.run([optimizer, loss, learning_rate, train_prediction], feed_dict=feed_dict)
            _, _, cl, l, lr, predictions = sess.run([update_centers, optimizer, center_loss, loss, learning_rate, train_prediction], feed_dict=feed_dict)
            if step % EVAL_FREQUENCY == 0:
                elapsed_time = time.time() - start_time
                start_time = time.time()
                print('Step %d (epoch %.2f), %.1f ms' %
                      (step, float(step) * BATCH_SIZE / train_size,
                       1000 * elapsed_time / EVAL_FREQUENCY))
                print('Minibatch loss: %.3f  %.3f, learning rate: %.6f' % (l, cl*beta, lr))
                print('Minibatch error: %.1f%%' % error_rate(predictions, batch_labels))
                print('Validation error: %.1f%%' % error_rate(
                    eval_in_batches(validation_data, sess), validation_labels))
                sys.stdout.flush()
        # Finally print the result!
        test_error = error_rate(eval_in_batches(test_data, sess), test_labels)
        print('Test error: %.1f%%' % test_error)
        if FLAGS.self_test:
            print('test_error', test_error)
            assert test_error == 0.0, 'expected 0.0 test_error, got %.2f' % (
                test_error,)
            
        train_embeddings = calculate_embeddings(train_data, sess)
        
        color_list = ['b', 'g', 'r', 'c', 'm', 'y', 'k', 'b', 'g', 'r', 'c' ]
        plt.figure(1)
        for n in range(0,10):
            idx = np.where(train_labels[0:10000]==n)
            plt.plot(train_embeddings[idx,0], train_embeddings[idx,1], color_list[n]+'.')
        plt.show()


if __name__ == '__main__':
    tf.app.run()