# Copyright 2020 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. # ============================================================================= """Training a deep NN on IMDB reviews with differentially private Adam optimizer.""" from __future__ import absolute_import from __future__ import division from __future__ import print_function from absl import app from absl import flags from keras.preprocessing import sequence import numpy as np import tensorflow as tf from tensorflow import estimator as tf_estimator from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator from tensorflow_privacy.privacy.analysis.gdp_accountant import compute_eps_poisson from tensorflow_privacy.privacy.analysis.gdp_accountant import compute_mu_poisson from tensorflow_privacy.privacy.optimizers import dp_optimizer #### FLAGS FLAGS = flags.FLAGS flags.DEFINE_boolean( 'dpsgd', True, 'If True, train with DP-SGD. If False, ' 'train with vanilla SGD.') flags.DEFINE_float('learning_rate', 0.02, 'Learning rate for training') flags.DEFINE_float('noise_multiplier', 0.56, 'Ratio of the standard deviation to the clipping norm') flags.DEFINE_float('l2_norm_clip', 1, 'Clipping norm') flags.DEFINE_integer('epochs', 25, 'Number of epochs') flags.DEFINE_integer('max_mu', 2, 'GDP upper limit') flags.DEFINE_string('model_dir', None, 'Model directory') sampling_batch = 512 microbatches = 512 max_features = 10000 maxlen = 256 num_examples = 25000 def nn_model_fn(features, labels, mode): """Define NN architecture using tf.keras.layers.""" input_layer = tf.reshape(features['x'], [-1, maxlen]) y = tf.keras.layers.Embedding(max_features, 16).apply(input_layer) y = tf.keras.layers.GlobalAveragePooling1D().apply(y) y = tf.keras.layers.Dense(16, activation='relu').apply(y) logits = tf.keras.layers.Dense(2).apply(y) # Calculate loss as a vector (to support microbatches in DP-SGD). vector_loss = tf.nn.sparse_softmax_cross_entropy_with_logits( 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 GradientDescentOptimizer. Other optimizers are # available in dp_optimizer. Most optimizers inheriting from # tf.train.Optimizer should be wrappable in differentially private # counterparts by calling dp_optimizer.optimizer_from_args(). optimizer = dp_optimizer.DPAdamGaussianOptimizer( l2_norm_clip=FLAGS.l2_norm_clip, noise_multiplier=FLAGS.noise_multiplier, num_microbatches=microbatches, learning_rate=FLAGS.learning_rate) opt_loss = vector_loss else: optimizer = tf.compat.v1.train.AdamOptimizer( learning_rate=FLAGS.learning_rate) opt_loss = scalar_loss global_step = tf.compat.v1.train.get_global_step() train_op = optimizer.minimize(loss=opt_loss, global_step=global_step) # In the following, we pass the mean of the loss (scalar_loss) rather than # the vector_loss because tf.estimator requires a scalar loss. This is only # used for evaluation and debugging by tf.estimator. The actual loss being # minimized is opt_loss defined above and passed to optimizer.minimize(). return tf_estimator.EstimatorSpec( mode=mode, loss=scalar_loss, train_op=train_op) # Add evaluation metrics (for EVAL mode). if mode == tf_estimator.ModeKeys.EVAL: eval_metric_ops = { 'accuracy': tf.compat.v1.metrics.accuracy( labels=labels, predictions=tf.argmax(input=logits, axis=1)) } return tf_estimator.EstimatorSpec( mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops) return None def load_imdb(): """Load IMDB movie reviews data.""" (train_data, train_labels), (test_data, test_labels) = tf.keras.datasets.imdb.load_data( num_words=max_features) train_data = sequence.pad_sequences( train_data, maxlen=maxlen).astype('float32') test_data = sequence.pad_sequences(test_data, maxlen=maxlen).astype('float32') return train_data, train_labels, test_data, test_labels def main(unused_argv): tf.compat.v1.logging.set_verbosity(3) # Load training and test data. train_data, train_labels, test_data, test_labels = load_imdb() # Instantiate the tf.Estimator. imdb_classifier = tf_estimator.Estimator( model_fn=nn_model_fn, model_dir=FLAGS.model_dir) # Create tf.Estimator input functions for the training and test data. eval_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn( x={'x': test_data}, y=test_labels, num_epochs=1, shuffle=False) # Training loop. steps_per_epoch = num_examples // sampling_batch test_accuracy_list = [] for epoch in range(1, FLAGS.epochs + 1): for _ in range(steps_per_epoch): whether = np.random.random_sample(num_examples) > ( 1 - sampling_batch / num_examples) subsampling = [i for i in np.arange(num_examples) if whether[i]] global microbatches microbatches = len(subsampling) train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn( x={'x': train_data[subsampling]}, y=train_labels[subsampling], batch_size=len(subsampling), num_epochs=1, shuffle=False) # Train the model for one step. imdb_classifier.train(input_fn=train_input_fn, steps=1) # Evaluate the model and print results eval_results = imdb_classifier.evaluate(input_fn=eval_input_fn) test_accuracy = eval_results['accuracy'] test_accuracy_list.append(test_accuracy) print('Test accuracy after %d epochs is: %.3f' % (epoch, test_accuracy)) # Compute the privacy budget expended so far. if FLAGS.dpsgd: eps = compute_eps_poisson(epoch, FLAGS.noise_multiplier, num_examples, sampling_batch, 1e-5) mu = compute_mu_poisson(epoch, FLAGS.noise_multiplier, num_examples, sampling_batch) print('For delta=1e-5, the current epsilon is: %.2f' % eps) print('For delta=1e-5, the current mu is: %.2f' % mu) if mu > FLAGS.max_mu: break else: print('Trained with vanilla non-private SGD optimizer') if __name__ == '__main__': app.run(main)