626_privacy/tensorflow_privacy
2
0
Fork 1
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions

Skips adding noise when noise_multiplier is 0 for fast clipping.

Browse source 
PiperOrigin-RevId: 522396275
This commit is contained in:
Shuang Song 2023-04-06 11:54:16 -07:00 committed by A. Unique TensorFlower
parent de9836883d
commit c4628d5dbc
1 changed files with 11 additions and 8 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

19
tensorflow_privacy/privacy/keras_models/dp_keras_model.py
View file

@ -231,16 +231,19 @@ def make_dp_model_class(cls):
self._num_microbatches,
)
)
grads = gradient_clipping_utils.add_aggregate_noise(
self,
clipped_grads,
eff_num_microbatches,
self._l2_norm_clip,
self._noise_multiplier,
)
if self._noise_multiplier > 0:
grads = gradient_clipping_utils.add_aggregate_noise(
self,
clipped_grads,
eff_num_microbatches,
self._l2_norm_clip,
self._noise_multiplier,
)
else:
grads = clipped_grads
output_metrics[privatized_loss_name] = weighted_loss
else:
logging.info('Computing gradients using microbatching.')
logging.info('Computing gradients using original clipping algorithm.')
# Computes per-example clipped gradients directly. This is called
# if at least one of the layers cannot use the "fast" gradient clipping
# algorithm.