Add files via upload
This commit is contained in:
woodyx218
GitHub
parent
47a984dc25
commit
1d5c5ac2fc
3 changed files with 435 additions and 258 deletions
178
tutorials/adult_tutorial.py
Normal file
178
tutorials/adult_tutorial.py
Normal file
|
|
@ -0,0 +1,178 @@
|
|||
# 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.
|
||||
# =============================================================================
|
||||
|
||||
"""Training a one-layer NN on Adult data with differentially private SGD optimizer."""
|
||||
|
||||
from __future__ import absolute_import
|
||||
from __future__ import division
|
||||
from __future__ import print_function
|
||||
|
||||
from absl import app
|
||||
from absl import flags
|
||||
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
import pandas as pd
|
||||
from sklearn.model_selection import KFold
|
||||
|
||||
# from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
|
||||
# from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
|
||||
from tensorflow_privacy.privacy.optimizers import dp_optimizer
|
||||
|
||||
from tensorflow_privacy.privacy.analysis.gdp_accountant import *
|
||||
|
||||
#### 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', .15, 'Learning rate for training')
|
||||
flags.DEFINE_float('noise_multiplier', 0.55,
|
||||
'Ratio of the standard deviation to the clipping norm')
|
||||
flags.DEFINE_float('l2_norm_clip', 1, 'Clipping norm')
|
||||
flags.DEFINE_integer('epochs', 20, 'Number of epochs')
|
||||
flags.DEFINE_integer('max_mu', 2, 'GDP upper limit')
|
||||
flags.DEFINE_string('model_dir', None, 'Model directory')
|
||||
|
||||
microbatches = 256
|
||||
|
||||
def nn_model_fn(features, labels, mode):
|
||||
''' Define CNN architecture using tf.keras.layers.'''
|
||||
input_layer = tf.reshape(features['x'], [-1, 123])
|
||||
y = tf.keras.layers.Dense(16, activation='relu').apply(input_layer)
|
||||
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.DPGradientDescentGaussianOptimizer(
|
||||
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.GradientDescentOptimizer(
|
||||
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_adult():
|
||||
"""Loads ADULT a2a as in LIBSVM and preprocesses to combine training and validation data."""
|
||||
# https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary.html
|
||||
|
||||
X = pd.read_csv("adult.csv")
|
||||
kf = KFold(n_splits=10)
|
||||
for train_index, test_index in kf.split(X):
|
||||
train, test = X.iloc[train_index, :], X.iloc[test_index, :]
|
||||
train_data = train.iloc[:, range(X.shape[1]-1)].values.astype('float32')
|
||||
test_data = test.iloc[:, range(X.shape[1]-1)].values.astype('float32')
|
||||
|
||||
train_labels = (train.iloc[:, X.shape[1]-1] == 1).astype('int32').values
|
||||
test_labels = (test.iloc[:, X.shape[1]-1] == 1).astype('int32').values
|
||||
|
||||
return train_data, train_labels, test_data, test_labels
|
||||
|
||||
|
||||
def main(unused_argv):
|
||||
'''main'''
|
||||
tf.compat.v1.logging.set_verbosity(0)
|
||||
|
||||
# Load training and test data.
|
||||
train_data, train_labels, test_data, test_labels = load_adult()
|
||||
|
||||
# Instantiate the tf.Estimator.
|
||||
adult_classifier = tf.compat.v1.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 = 29305 // 256
|
||||
test_accuracy_list = []
|
||||
for epoch in range(1, FLAGS.epochs + 1):
|
||||
for step in range(steps_per_epoch):
|
||||
whether = np.random.random_sample(29305) > (1-256/29305)
|
||||
subsampling = [i for i in np.arange(29305) 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=True)
|
||||
# Train the model for one step.
|
||||
adult_classifier.train(input_fn=train_input_fn, steps=1)
|
||||
|
||||
# Evaluate the model and print results
|
||||
eval_results = adult_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, 29305, 256, 1e-5)
|
||||
mu = compute_mu_Poisson(epoch, FLAGS.noise_multiplier, 29305, 256)
|
||||
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)
|
||||
|
|
@ -19,46 +19,47 @@ from __future__ import absolute_import
|
|||
from __future__ import division
|
||||
from __future__ import print_function
|
||||
|
||||
from absl import app
|
||||
from absl import flags
|
||||
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
from scipy.stats import norm
|
||||
from keras.preprocessing import sequence
|
||||
|
||||
from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
|
||||
from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
|
||||
#from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
|
||||
#from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
|
||||
from tensorflow_privacy.privacy.optimizers import dp_optimizer
|
||||
|
||||
from GDprivacy_accountants import *
|
||||
from tensorflow_privacy.privacy.analysis.gdp_accountant import *
|
||||
|
||||
|
||||
from keras.preprocessing import sequence
|
||||
#### FLAGS
|
||||
tf.flags.DEFINE_boolean('dpsgd', True, 'If True, train with DP-SGD. If False, '
|
||||
FLAGS = flags.FLAGS
|
||||
flags.DEFINE_boolean('dpsgd', True, 'If True, train with DP-SGD. If False, '
|
||||
'train with vanilla SGD.')
|
||||
tf.flags.DEFINE_float('learning_rate', 0.02, 'Learning rate for training')
|
||||
tf.flags.DEFINE_float('noise_multiplier', 0.56,
|
||||
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')
|
||||
tf.flags.DEFINE_float('l2_norm_clip', 1, 'Clipping norm')
|
||||
tf.flags.DEFINE_integer('epochs', 25, 'Number of epochs')
|
||||
tf.flags.DEFINE_integer('max_mu', 2, 'GDP upper limit')
|
||||
tf.flags.DEFINE_string('model_dir', None, 'Model directory')
|
||||
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')
|
||||
|
||||
FLAGS = tf.flags.FLAGS
|
||||
|
||||
microbatches=512
|
||||
np.random.seed(0)
|
||||
tf.set_random_seed(0)
|
||||
microbatches = 512
|
||||
|
||||
max_features = 10000
|
||||
# cut texts after this number of words (among top max_features most common words)
|
||||
maxlen = 256
|
||||
|
||||
|
||||
def rnn_model_fn(features, labels, mode):
|
||||
# Define CNN 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)
|
||||
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(
|
||||
|
|
@ -68,7 +69,6 @@ def rnn_model_fn(features, labels, mode):
|
|||
|
||||
# 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
|
||||
|
|
@ -81,11 +81,11 @@ def rnn_model_fn(features, labels, mode):
|
|||
learning_rate=FLAGS.learning_rate)
|
||||
opt_loss = vector_loss
|
||||
else:
|
||||
optimizer = tf.train.AdamOptimizer(
|
||||
optimizer = tf.compat.v1.train.AdamOptimizer(
|
||||
learning_rate=FLAGS.learning_rate)
|
||||
opt_loss = scalar_loss
|
||||
|
||||
global_step = tf.train.get_global_step()
|
||||
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
|
||||
|
|
@ -96,39 +96,43 @@ def rnn_model_fn(features, labels, mode):
|
|||
train_op=train_op)
|
||||
|
||||
# Add evaluation metrics (for EVAL mode).
|
||||
elif mode == tf.estimator.ModeKeys.EVAL:
|
||||
if mode == tf.estimator.ModeKeys.EVAL:
|
||||
eval_metric_ops = {
|
||||
'accuracy':
|
||||
tf.metrics.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():
|
||||
(train_data,train_labels), (test_data,test_labels) = tf.keras.datasets.imdb.load_data(num_words=max_features)
|
||||
'''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
|
||||
return train_data, train_labels, test_data, test_labels
|
||||
|
||||
|
||||
def main(unused_argv):
|
||||
tf.logging.set_verbosity(3)
|
||||
'''main'''
|
||||
tf.compat.v1.logging.set_verbosity(3)
|
||||
|
||||
# Load training and test data.
|
||||
train_data,train_labels,test_data,test_labels = load_imdb()
|
||||
train_data, train_labels, test_data, test_labels = load_imdb()
|
||||
|
||||
# Instantiate the tf.Estimator.
|
||||
imdb_classifier = tf.estimator.Estimator(model_fn=rnn_model_fn,
|
||||
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.estimator.inputs.numpy_input_fn(
|
||||
eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
|
||||
x={'x': test_data},
|
||||
y=test_labels,
|
||||
num_epochs=1,
|
||||
|
|
@ -139,15 +143,13 @@ def main(unused_argv):
|
|||
test_accuracy_list = []
|
||||
|
||||
for epoch in range(1, FLAGS.epochs + 1):
|
||||
np.random.seed(epoch)
|
||||
for step in range(steps_per_epoch):
|
||||
tf.set_random_seed(0)
|
||||
whether=np.random.random_sample(25000)>(1-512/25000)
|
||||
subsampling=[i for i in np.arange(25000) if whether[i]]
|
||||
whether = np.random.random_sample(25000) > (1-512/25000)
|
||||
subsampling = [i for i in np.arange(25000) if whether[i]]
|
||||
global microbatches
|
||||
microbatches=len(subsampling)
|
||||
microbatches = len(subsampling)
|
||||
|
||||
train_input_fn = tf.estimator.inputs.numpy_input_fn(
|
||||
train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
|
||||
x={'x': train_data[subsampling]},
|
||||
y=train_labels[subsampling],
|
||||
batch_size=len(subsampling),
|
||||
|
|
@ -164,16 +166,16 @@ def main(unused_argv):
|
|||
|
||||
# Compute the privacy budget expended so far.
|
||||
if FLAGS.dpsgd:
|
||||
eps = compute_epsP(epoch,FLAGS.noise_multiplier,25000,512,1e-5)
|
||||
mu= compute_muP(epoch,FLAGS.noise_multiplier,25000,512)
|
||||
eps = compute_eps_Poisson(epoch, FLAGS.noise_multiplier, 25000, 512, 1e-5)
|
||||
mu = compute_mu_Poisson(epoch, FLAGS.noise_multiplier, 25000, 512)
|
||||
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:
|
||||
if mu > FLAGS.max_mu:
|
||||
break
|
||||
else:
|
||||
print('Trained with vanilla non-private SGD optimizer')
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
tf.app.run()
|
||||
app.run(main)
|
||||
|
|
|
|||
|
|
@ -24,47 +24,45 @@ from absl import flags
|
|||
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
from scipy.stats import norm
|
||||
import pandas as pd
|
||||
from scipy.stats import rankdata
|
||||
from sklearn.model_selection import train_test_split
|
||||
|
||||
from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
|
||||
from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
|
||||
#from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
|
||||
#from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
|
||||
from tensorflow_privacy.privacy.optimizers import dp_optimizer
|
||||
|
||||
from GDprivacy_accountants import *
|
||||
from tensorflow_privacy.privacy.analysis.gdp_accountant import *
|
||||
|
||||
#### FLAGS
|
||||
tf.flags.DEFINE_boolean('dpsgd', True, 'If True, train with DP-SGD. If False, '
|
||||
FLAGS = flags.FLAGS
|
||||
flags.DEFINE_boolean('dpsgd', True, 'If True, train with DP-SGD. If False, '
|
||||
'train with vanilla SGD.')
|
||||
tf.flags.DEFINE_float('learning_rate', .01, 'Learning rate for training')
|
||||
tf.flags.DEFINE_float('noise_multiplier', 0.55,
|
||||
flags.DEFINE_float('learning_rate', .01, 'Learning rate for training')
|
||||
flags.DEFINE_float('noise_multiplier', 0.55,
|
||||
'Ratio of the standard deviation to the clipping norm')
|
||||
tf.flags.DEFINE_float('l2_norm_clip', 5, 'Clipping norm')
|
||||
tf.flags.DEFINE_integer('epochs', 25, 'Number of epochs')
|
||||
tf.flags.DEFINE_integer('max_mu', 2, 'GDP upper limit')
|
||||
tf.flags.DEFINE_string('model_dir', None, 'Model directory')
|
||||
flags.DEFINE_float('l2_norm_clip', 5, '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')
|
||||
|
||||
FLAGS = tf.flags.FLAGS
|
||||
|
||||
microbatches=10000
|
||||
np.random.seed(0)
|
||||
tf.set_random_seed(0)
|
||||
|
||||
n_users=6040
|
||||
n_movies=3706
|
||||
microbatches = 10000
|
||||
|
||||
def nn_model_fn(features, labels, mode):
|
||||
# Adapted from https://github.com/hexiangnan/neural_collaborative_filtering
|
||||
'''NN adapted from github.com/hexiangnan/neural_collaborative_filtering'''
|
||||
n_latent_factors_user = 10
|
||||
n_latent_factors_movie = 10
|
||||
n_latent_factors_mf = 5
|
||||
|
||||
user_input = tf.reshape(features['user'], [-1,1])
|
||||
item_input = tf.reshape(features['movie'], [-1,1])
|
||||
user_input = tf.reshape(features['user'], [-1, 1])
|
||||
item_input = tf.reshape(features['movie'], [-1, 1])
|
||||
|
||||
mf_embedding_user = tf.keras.layers.Embedding(n_users,n_latent_factors_mf,input_length=1)
|
||||
mf_embedding_item = tf.keras.layers.Embedding(n_movies,n_latent_factors_mf,input_length=1)
|
||||
mlp_embedding_user = tf.keras.layers.Embedding(n_users,n_latent_factors_user,input_length=1)
|
||||
mlp_embedding_item = tf.keras.layers.Embedding(n_movies,n_latent_factors_movie,input_length=1)
|
||||
# number of users: 6040; number of movies: 3706
|
||||
mf_embedding_user = tf.keras.layers.Embedding(6040, n_latent_factors_mf, input_length=1)
|
||||
mf_embedding_item = tf.keras.layers.Embedding(3706, n_latent_factors_mf, input_length=1)
|
||||
mlp_embedding_user = tf.keras.layers.Embedding(6040, n_latent_factors_user, input_length=1)
|
||||
mlp_embedding_item = tf.keras.layers.Embedding(3706, n_latent_factors_movie, input_length=1)
|
||||
|
||||
# GMF part
|
||||
# Flatten the embedding vector as latent features in GMF
|
||||
|
|
@ -91,7 +89,6 @@ def nn_model_fn(features, labels, mode):
|
|||
|
||||
# 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
|
||||
|
|
@ -104,11 +101,11 @@ def nn_model_fn(features, labels, mode):
|
|||
learning_rate=FLAGS.learning_rate)
|
||||
opt_loss = vector_loss
|
||||
else:
|
||||
optimizer = tf.train.AdamOptimizer(
|
||||
optimizer = tf.compat.v1.train.AdamOptimizer(
|
||||
learning_rate=FLAGS.learning_rate)
|
||||
opt_loss = scalar_loss
|
||||
|
||||
global_step = tf.train.get_global_step()
|
||||
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
|
||||
|
|
@ -119,56 +116,58 @@ def nn_model_fn(features, labels, mode):
|
|||
train_op=train_op)
|
||||
|
||||
# Add evaluation metrics (for EVAL mode).
|
||||
elif mode == tf.estimator.ModeKeys.EVAL:
|
||||
if mode == tf.estimator.ModeKeys.EVAL:
|
||||
eval_metric_ops = {
|
||||
'rmse':
|
||||
tf.metrics.root_mean_squared_error(
|
||||
tf.compat.v1.metrics.root_mean_squared_error(
|
||||
labels=tf.cast(labels, tf.float32),
|
||||
predictions=tf.tensordot(a=tf.nn.softmax(logits,axis=1),b=tf.constant(np.array([0,1,2,3,4]),dtype=tf.float32),axes=1))
|
||||
predictions=tf.tensordot(a=tf.nn.softmax(logits, axis=1),
|
||||
b=tf.constant(np.array([0, 1, 2, 3, 4]),
|
||||
dtype=tf.float32),
|
||||
axes=1))
|
||||
}
|
||||
return tf.estimator.EstimatorSpec(mode=mode,
|
||||
loss=scalar_loss,
|
||||
eval_metric_ops=eval_metric_ops)
|
||||
return None
|
||||
|
||||
|
||||
def load_adult():
|
||||
import pandas as pd
|
||||
import numpy as np
|
||||
|
||||
data = pd.read_csv('ratings.dat', sep='::', header=None,names=["userId", "movieId", "rating", "timestamp"])
|
||||
n_users=len(set(data['userId']))
|
||||
n_movies=len(set(data['movieId']))
|
||||
print('number of movie: ',n_movies)
|
||||
print('number of user: ',n_users)
|
||||
"""Loads MovieLens 1M as from https://grouplens.org/datasets/movielens/1m"""
|
||||
data = pd.read_csv('ratings.dat', sep='::', header=None,
|
||||
names=["userId", "movieId", "rating", "timestamp"])
|
||||
n_users = len(set(data['userId']))
|
||||
n_movies = len(set(data['movieId']))
|
||||
print('number of movie: ', n_movies)
|
||||
print('number of user: ', n_users)
|
||||
|
||||
# give unique dense movie index to movieId
|
||||
from scipy.stats import rankdata
|
||||
data['movieIndex']=rankdata(data['movieId'], method='dense')
|
||||
data['movieIndex'] = rankdata(data['movieId'], method='dense')
|
||||
# minus one to reduce the minimum value to 0, which is the start of col index
|
||||
|
||||
print('number of ratings:',data.shape[0])
|
||||
print('percentage of sparsity:',(1-data.shape[0]/n_users/n_movies)*100,'%')
|
||||
print('number of ratings:', data.shape[0])
|
||||
print('percentage of sparsity:', (1-data.shape[0]/n_users/n_movies)*100, '%')
|
||||
|
||||
from sklearn.model_selection import train_test_split
|
||||
train,test=train_test_split(data,test_size=0.2,random_state=100)
|
||||
train, test = train_test_split(data, test_size=0.2, random_state=100)
|
||||
|
||||
return train.values-1, test.values-1, np.mean(train['rating'])
|
||||
|
||||
|
||||
def main(unused_argv):
|
||||
tf.logging.set_verbosity(3)
|
||||
'''main'''
|
||||
tf.compat.v1.logging.set_verbosity(3)
|
||||
|
||||
# Load training and test data.
|
||||
train_data, test_data, mean = load_adult()
|
||||
|
||||
# Instantiate the tf.Estimator.
|
||||
adult_classifier = tf.estimator.Estimator(model_fn=nn_model_fn,
|
||||
ml_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.estimator.inputs.numpy_input_fn(
|
||||
x={'user': test_data[:,0], 'movie': test_data[:,4]},
|
||||
y=test_data[:,2],
|
||||
eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
|
||||
x={'user': test_data[:, 0], 'movie': test_data[:, 4]},
|
||||
y=test_data[:, 2],
|
||||
num_epochs=1,
|
||||
shuffle=False)
|
||||
|
||||
|
|
@ -176,41 +175,39 @@ def main(unused_argv):
|
|||
steps_per_epoch = 800167 // 10000
|
||||
test_accuracy_list = []
|
||||
for epoch in range(1, FLAGS.epochs + 1):
|
||||
np.random.seed(epoch)
|
||||
for step in range(steps_per_epoch):
|
||||
tf.set_random_seed(0)
|
||||
whether=np.random.random_sample(800167)>(1-10000/800167)
|
||||
subsampling=[i for i in np.arange(800167) if whether[i]]
|
||||
whether = np.random.random_sample(800167) > (1-10000/800167)
|
||||
subsampling = [i for i in np.arange(800167) if whether[i]]
|
||||
global microbatches
|
||||
microbatches=len(subsampling)
|
||||
microbatches = len(subsampling)
|
||||
|
||||
train_input_fn = tf.estimator.inputs.numpy_input_fn(
|
||||
x={'user': train_data[subsampling,0], 'movie': train_data[subsampling,4]},
|
||||
y=train_data[subsampling,2],
|
||||
train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
|
||||
x={'user': train_data[subsampling, 0], 'movie': train_data[subsampling, 4]},
|
||||
y=train_data[subsampling, 2],
|
||||
batch_size=len(subsampling),
|
||||
num_epochs=1,
|
||||
shuffle=True)
|
||||
# Train the model for one step.
|
||||
adult_classifier.train(input_fn=train_input_fn, steps=1)
|
||||
ml_classifier.train(input_fn=train_input_fn, steps=1)
|
||||
|
||||
# Evaluate the model and print results
|
||||
eval_results = adult_classifier.evaluate(input_fn=eval_input_fn)
|
||||
eval_results = ml_classifier.evaluate(input_fn=eval_input_fn)
|
||||
test_accuracy = eval_results['rmse']
|
||||
test_accuracy_list.append(test_accuracy)
|
||||
print('Test RMSE after %d epochs is: %.3f' % (epoch, test_accuracy))
|
||||
|
||||
# Compute the privacy budget expended so far.
|
||||
if FLAGS.dpsgd:
|
||||
eps = compute_epsP(epoch,FLAGS.noise_multiplier,800167,10000,1e-6)
|
||||
mu= compute_muP(epoch,FLAGS.noise_multiplier,800167,10000)
|
||||
eps = compute_eps_Poisson(epoch, FLAGS.noise_multiplier, 800167, 10000, 1e-6)
|
||||
mu = compute_mu_Poisson(epoch, FLAGS.noise_multiplier, 800167, 10000)
|
||||
print('For delta=1e-6, the current epsilon is: %.2f' % eps)
|
||||
print('For delta=1e-6, the current mu is: %.2f' % mu)
|
||||
|
||||
if mu>FLAGS.max_mu:
|
||||
if mu > FLAGS.max_mu:
|
||||
break
|
||||
else:
|
||||
print('Trained with vanilla non-private SGD optimizer')
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
tf.app.run()
|
||||
app.run(main)
|
||||
|
|
|
|||
Loading…
Reference in a new issue