ruby/tensorflow_privacy
1
0
Fork 0
forked from 626_privacy/tensorflow_privacy
Code Pull requests Activity

update dp keras model

Browse source
This commit is contained in:
pranav subramani 2021-01-08 00:22:44 -07:00
parent 9d871b28c1
commit 6982e027b5
1 changed files with 7 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
tensorflow_privacy/privacy/keras_models/dp_keras_model.py
View file

@ -27,7 +27,7 @@ def make_dp_model_class(cls):
noise = tf.random.normal( noise = tf.random.normal(
tf.shape(input=summed_grads), stddev=noise_stddev) tf.shape(input=summed_grads), stddev=noise_stddev)
noised_grads = summed_grads + noise noised_grads = summed_grads + noise
return noised_grads / tf.cast(stacked_grads.shape[0], noised_grads.dtype) return noised_grads / tf.cast(stacked_grads.shape[0], tf.float32)
def compute_per_example_grads(self, data): def compute_per_example_grads(self, data):
x, y = data x, y = data
@ -35,23 +35,24 @@ def make_dp_model_class(cls):
# We need to add the extra dimension to x and y because model # We need to add the extra dimension to x and y because model
# expects batched input. # expects batched input.
y_pred = self(x[None], training=True) y_pred = self(x[None], training=True)
loss = self.compiled_loss(y[None], y_pred, loss = self.compiled_loss(y[None], y_pred,
regularization_losses=self.losses) regularization_losses=self.losses)
grads_list = tape.gradient(loss, self.trainable_variables) grads_list = tape.gradient(loss, self.trainable_variables)
clipped_grads = self.process_per_example_grads(grads_list) clipped_grads = self.process_per_example_grads(grads_list)
return tf.squeeze(y_pred, axis=0), loss, clipped_grads return tf.squeeze(y_pred, axis=0), loss, clipped_grads
def train_step(self, data): def train_step(self, data):
x, y = data x, y = data
y_pred, per_eg_loss, per_eg_grads = tf.vectorized_map( y_pred, per_eg_loss, per_eg_grads = tf.vectorized_map(
self.compute_per_example_grads, data) self.compute_per_example_grads, data)
loss = tf.reduce_mean(per_eg_loss, axis=0) loss = tf.reduce_mean(per_eg_loss, axis=0)
grads = tf.nest.map_structure(self.reduce_per_example_grads, per_eg_grads) grads = tf.nest.map_structure(
self.reduce_per_example_grads, per_eg_grads)
self.optimizer.apply_gradients(zip(grads, self.trainable_variables)) self.optimizer.apply_gradients(zip(grads, self.trainable_variables))
self.compiled_metrics.update_state(y, y_pred) self.compiled_metrics.update_state(y, y_pred)
return {m.name: m.result() for m in self.metrics} return {m.name: m.result() for m in self.metrics}
return DPModelClass return DPModelClass