# Copyright 2019, 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 Keras and the DP SGD optimizer."""

from absl import app
from absl import flags
from absl import logging
from com_google_differential_py.python.dp_accounting
import numpy as np
import tensorflow as tf
from tensorflow_privacy.privacy.optimizers.dp_optimizer_keras import DPKerasSGDOptimizer


flags.DEFINE_boolean(
    'dpsgd', True, 'If True, train with DP-SGD. If False, '
    'train with vanilla SGD.')
flags.DEFINE_float('learning_rate', 0.15, 'Learning rate for training')
flags.DEFINE_float('noise_multiplier', 0.1,
                   'Ratio of the standard deviation to the clipping norm')
flags.DEFINE_float('l2_norm_clip', 1.0, 'Clipping norm')
flags.DEFINE_integer('batch_size', 250, 'Batch size')
flags.DEFINE_integer('epochs', 60, 'Number of epochs')
flags.DEFINE_integer(
    'microbatches', 250, 'Number of microbatches '
    '(must evenly divide batch_size)')
flags.DEFINE_string('model_dir', None, 'Model directory')

FLAGS = flags.FLAGS


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)] + list(range(12, 64))
  accountant = dp_accounting.rdp.RdpAccountant(orders)

  sampling_probability = FLAGS.batch_size / 60000
  event = dp_accounting.SelfComposedDpEvent(
      dp_accounting.PoissonSampledDpEvent(
          sampling_probability,
          dp_accounting.GaussianDpEvent(FLAGS.noise_multiplier)), steps)

  accountant.compose(event)

  # Delta is set to 1e-5 because MNIST has 60000 training points.
  return accountant.get_epsilon(target_delta=1e-5)


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_data = train_data.reshape((train_data.shape[0], 28, 28, 1))
  test_data = test_data.reshape((test_data.shape[0], 28, 28, 1))

  train_labels = np.array(train_labels, dtype=np.int32)
  test_labels = np.array(test_labels, dtype=np.int32)

  train_labels = tf.keras.utils.to_categorical(train_labels, num_classes=10)
  test_labels = tf.keras.utils.to_categorical(test_labels, num_classes=10)

  assert train_data.min() == 0.
  assert train_data.max() == 1.
  assert test_data.min() == 0.
  assert test_data.max() == 1.

  return train_data, train_labels, test_data, test_labels


def main(unused_argv):
  logging.set_verbosity(logging.INFO)
  if FLAGS.dpsgd and 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()

  # Define a sequential Keras model
  model = tf.keras.Sequential([
      tf.keras.layers.Conv2D(
          16,
          8,
          strides=2,
          padding='same',
          activation='relu',
          input_shape=(28, 28, 1)),
      tf.keras.layers.MaxPool2D(2, 1),
      tf.keras.layers.Conv2D(
          32, 4, strides=2, padding='valid', activation='relu'),
      tf.keras.layers.MaxPool2D(2, 1),
      tf.keras.layers.Flatten(),
      tf.keras.layers.Dense(32, activation='relu'),
      tf.keras.layers.Dense(10)
  ])

  if FLAGS.dpsgd:
    optimizer = DPKerasSGDOptimizer(
        l2_norm_clip=FLAGS.l2_norm_clip,
        noise_multiplier=FLAGS.noise_multiplier,
        num_microbatches=FLAGS.microbatches,
        learning_rate=FLAGS.learning_rate)
    # Compute vector of per-example loss rather than its mean over a minibatch.
    loss = tf.keras.losses.CategoricalCrossentropy(
        from_logits=True, reduction=tf.losses.Reduction.NONE)
  else:
    optimizer = tf.keras.optimizers.SGD(learning_rate=FLAGS.learning_rate)
    loss = tf.keras.losses.CategoricalCrossentropy(from_logits=True)

  # Compile model with Keras
  model.compile(optimizer=optimizer, loss=loss, metrics=['accuracy'])

  # Train model with Keras
  model.fit(
      train_data,
      train_labels,
      epochs=FLAGS.epochs,
      validation_data=(test_data, test_labels),
      batch_size=FLAGS.batch_size)

  # Compute the privacy budget expended.
  if FLAGS.dpsgd:
    eps = compute_epsilon(FLAGS.epochs * 60000 // FLAGS.batch_size)
    print('For delta=1e-5, the current epsilon is: %.2f' % eps)
  else:
    print('Trained with vanilla non-private SGD optimizer')


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