def linear_model_fn(features, labels, mode):
preds = tf.keras.layers.Dense(1, activation='linear',
name='dense').apply(features['x'])
vector_loss = tf.squared_difference(labels, preds)
scalar_loss = tf.reduce_mean(vector_loss)
dp_average_query = gaussian_query.GaussianAverageQuery(1.0, 0.0, 1)
optimizer = dp_optimizer.DPGradientDescentOptimizer(
dp_average_query, num_microbatches=1, learning_rate=1.0)
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)
def define_training_procedure(self, parameters):
# Define training procedure
self.global_step = tf.Variable(0, name="global_step", trainable=False)
if parameters['optimizer'] == 'adam':
self.optimizer = tf.train.AdamOptimizer(parameters['learning_rate'])
elif parameters['optimizer'] == 'sgd':
# self.optimizer = tf.train.GradientDescentOptimizer(parameters['learning_rate'])
###alteration to make it private
# self.optimizer = dp_optimizer.DPGradientDescentGaussianOptimizer(
self.optimizer = dp_optimizer.DPGradientDescentOptimizer(
l2_norm_clip=1.0,#Clipping norm
noise_multiplier=1.1,
num_microbatches=256,
# ledger=ledger,
learning_rate=parameters['learning_rate'])
# training_hooks = [
# EpsilonPrintingTrainingHook(ledger)
# ]
# opt_loss = vector_loss
###
elif parameters['optimizer'] == 'adadelta':
self.optimizer = tf.train.AdadeltaOptimizer(parameters['learning_rate'])
else:
raise ValueError('The lr_method parameter must be either adadelta, adam or sgd.')
###alteration to make it private
grads_and_vars = self.optimizer.compute_gradients(self.vector_loss,self.input_label_indices_vector)
# grads_and_vars = self.optimizer.compute_gradients(self.loss)
###
if parameters['gradient_clipping_value']:
grads_and_vars = [(tf.clip_by_value(grad, -parameters['gradient_clipping_value'], parameters['gradient_clipping_value']), var)
for grad, var in grads_and_vars if grad is not None]
# grads_and_vars = [(tf.clip_by_value(grad, -parameters['gradient_clipping_value'], parameters['gradient_clipping_value']), var)
# for grad, var in grads_and_vars]
# By defining a global_step variable and passing it to the optimizer we allow TensorFlow handle the counting of training steps for us.
# The global step will be automatically incremented by one every time you execute train_op.
self.train_op = self.optimizer.apply_gradients(grads_and_vars, global_step=self.global_step)
