# Copyright 2018, The TensorFlow Authors.
#
# 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.

"""Training a CNN on MNIST with differentially private Adam optimizer."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import numpy as np
import tensorflow as tf

from privacy.analysis.rdp_accountant import compute_rdp
from privacy.analysis.rdp_accountant import get_privacy_spent
from privacy.optimizers import dp_optimizer

tf.flags.DEFINE_boolean('dpsgd', True, 'If True, train with DP-Adam. If False,'
                        'train with vanilla Adam.')
tf.flags.DEFINE_float('learning_rate', 0.0015, 'Learning rate for training')
tf.flags.DEFINE_float('noise_multiplier', 1.0,
                      'Ratio of the standard deviation to the clipping norm')
tf.flags.DEFINE_float('l2_norm_clip', 1.0, 'Clipping norm')
tf.flags.DEFINE_integer('batch_size', 256, 'Batch size')
tf.flags.DEFINE_integer('epochs', 15, 'Number of epochs')
tf.flags.DEFINE_integer('microbatches', 256,
                        'Number of microbatches (must evenly divide batch_size')
tf.flags.DEFINE_string('model_dir', None, 'Model directory')

FLAGS = tf.flags.FLAGS


def cnn_model_fn(features, labels, mode):
  """Model function for a CNN."""

  # Define CNN architecture using tf.keras.layers.
  input_layer = tf.reshape(features['x'], [-1, 28, 28, 1])
  y = tf.keras.layers.Conv2D(16, 8,
                             strides=2,
                             padding='same',
                             kernel_initializer='he_normal').apply(input_layer)
  y = tf.keras.layers.MaxPool2D(2, 1).apply(y)
  y = tf.keras.layers.Conv2D(32, 4,
                             strides=2,
                             padding='valid',
                             kernel_initializer='he_normal').apply(y)
  y = tf.keras.layers.MaxPool2D(2, 1).apply(y)
  y = tf.keras.layers.Flatten().apply(y)
  y = tf.keras.layers.Dense(32, kernel_initializer='he_normal').apply(y)
  logits = tf.keras.layers.Dense(10, kernel_initializer='he_normal').apply(y)

  # Calculate loss as a vector (to support microbatches in DP-SGD).
  vector_loss = tf.nn.softmax_cross_entropy_with_logits_v2(
      labels=labels, logits=logits)
  # Define mean of loss across minibatch (for reporting through tf.Estimator).
  scalar_loss = tf.reduce_mean(vector_loss)

  # Configure the training op (for TRAIN mode).
  if mode == tf.estimator.ModeKeys.TRAIN:

    if FLAGS.dpsgd:
      # Use DP version of AdamOptimizer. For illustration purposes, we do that
      # here by calling make_optimizer_class() explicitly, though DP versions
      # of standard optimizers are available in dp_optimizer.
      dp_optimizer_class = dp_optimizer.make_optimizer_class(
          tf.train.AdamOptimizer)
      optimizer = dp_optimizer_class(
          learning_rate=FLAGS.learning_rate,
          noise_multiplier=FLAGS.noise_multiplier,
          l2_norm_clip=FLAGS.l2_norm_clip,
          num_microbatches=FLAGS.microbatches)
    else:
      optimizer = tf.train.AdamOptimizer(learning_rate=FLAGS.learning_rate)
    global_step = tf.train.get_global_step()
    train_op = optimizer.minimize(loss=vector_loss, global_step=global_step)
    return tf.estimator.EstimatorSpec(mode=mode,
                                      loss=scalar_loss,
                                      train_op=train_op)

  # Add evaluation metrics (for EVAL mode).
  elif mode == tf.estimator.ModeKeys.EVAL:
    eval_metric_ops = {
        'accuracy':
            tf.metrics.accuracy(
                labels=tf.argmax(labels, axis=1),
                predictions=tf.argmax(input=logits, axis=1))
    }
    return tf.estimator.EstimatorSpec(mode=mode,
                                      loss=scalar_loss,
                                      eval_metric_ops=eval_metric_ops)


def load_mnist():
  """Loads MNIST and preprocesses to combine training and validation data."""
  train, test = tf.keras.datasets.mnist.load_data()
  train_data, train_labels = train
  test_data, test_labels = test

  train_data = np.array(train_data, dtype=np.float32) / 255
  test_data = np.array(test_data, dtype=np.float32) / 255

  train_labels = tf.keras.utils.to_categorical(train_labels)
  test_labels = tf.keras.utils.to_categorical(test_labels)

  assert train_data.min() == 0.
  assert train_data.max() == 1.
  assert test_data.min() == 0.
  assert test_data.max() == 1.
  assert train_labels.shape[1] == 10
  assert test_labels.shape[1] == 10

  return train_data, train_labels, test_data, test_labels


def main(unused_argv):
  tf.logging.set_verbosity(tf.logging.INFO)
  if FLAGS.batch_size % FLAGS.microbatches != 0:
    raise ValueError('Number of microbatches should divide evenly batch_size')

  # Load training and test data.
  train_data, train_labels, test_data, test_labels = load_mnist()

  # Instantiate the tf.Estimator.
  mnist_classifier = tf.estimator.Estimator(model_fn=cnn_model_fn,
                                            model_dir=FLAGS.model_dir)

  # Create tf.Estimator input functions for the training and test data.
  train_input_fn = tf.estimator.inputs.numpy_input_fn(
      x={'x': train_data},
      y=train_labels,
      batch_size=FLAGS.batch_size,
      num_epochs=FLAGS.epochs,
      shuffle=True)
  eval_input_fn = tf.estimator.inputs.numpy_input_fn(
      x={'x': test_data},
      y=test_labels,
      num_epochs=1,
      shuffle=False)

  # Define a function that computes privacy budget expended so far.
  def compute_epsilon(steps):
    """Computes epsilon value for given hyperparameters."""
    if FLAGS.noise_multiplier == 0.0:
      return float('inf')
    orders = [1 + x / 10. for x in range(1, 100)] + range(12, 64)
    sampling_probability = FLAGS.batch_size / 60000
    rdp = compute_rdp(q=sampling_probability,
                      stddev_to_sensitivity_ratio=FLAGS.noise_multiplier,
                      steps=steps,
                      orders=orders)
    # Delta is set to 1e-5 because MNIST has 60000 training points.
    return get_privacy_spent(orders, rdp, target_delta=1e-5)[0]

  # Training loop.
  steps_per_epoch = 60000 // FLAGS.batch_size
  for epoch in range(1, FLAGS.epochs + 1):
    # Train the model for one epoch.
    mnist_classifier.train(input_fn=train_input_fn, steps=steps_per_epoch)

    # Evaluate the model and print results
    eval_results = mnist_classifier.evaluate(input_fn=eval_input_fn)
    test_accuracy = eval_results['accuracy']
    print('Test accuracy after %d epochs is: %.3f' % (epoch, test_accuracy))

    # Compute the privacy budget expended so far.
    if FLAGS.dpsgd:
      eps = compute_epsilon(epoch * steps_per_epoch)
      print('For delta=1e-5, the current epsilon is: %.2f' % eps)
    else:
      print('Trained with vanilla non-private Adam optimizer')

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