-
Notifications
You must be signed in to change notification settings - Fork 198
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
add synthetic expt, confidence estimators, training code
PiperOrigin-RevId: 596072058 Change-Id: If8e1f3743507a9c473609124418d2f1175df0951
- Loading branch information
1 parent
2d68af9
commit 9969a86
Showing
22 changed files
with
2,271 additions
and
6 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,133 @@ | ||
| # Copyright 2023, Google LLC. | ||
| # | ||
| # 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. | ||
| """Evaluation utilities required for auditing.""" | ||
|
|
||
| # from absl import logging | ||
| import numpy as np | ||
| import tensorflow as tf | ||
|
|
||
| from lidp_auditing import constants | ||
| from lidp_auditing import utils | ||
|
|
||
|
|
||
| def get_evaluate_fn(): | ||
| """Return a `tf.function` to evaluate a keras model.""" | ||
|
|
||
| @tf.function | ||
| def evaluate_model(dataset, model, metric): | ||
| # Note: pass in batched dataset | ||
| metric.reset_state() | ||
| for x, y, _ in dataset: | ||
| predictions = model(x, training=False) | ||
| metric.update_state(y, predictions) | ||
| return metric.result() | ||
|
|
||
| return evaluate_model | ||
|
|
||
|
|
||
| def evaluate_canary_dataset( | ||
| canary_type: str, | ||
| canary_dataset: tf.data.Dataset, | ||
| model: tf.keras.Model, | ||
| vector_loss_fn: tf.keras.losses.Loss, | ||
| batch_size: int, | ||
| ) -> np.ndarray: | ||
| """Run the test to see if the canary can be found.""" | ||
| if canary_dataset is None: | ||
| return np.array([]) | ||
| if canary_type == constants.RANDOM_GRADIENT_CANARY: | ||
| # return evaluate_random_gradient_canary(canary_dataset, model) | ||
| return evaluate_random_gradient_canary_batched(canary_dataset, model) | ||
|
|
||
| # Static or adaptive data canary | ||
| return evaluate_data_canary(canary_dataset, model, vector_loss_fn, batch_size) | ||
|
|
||
|
|
||
| def evaluate_data_canary(canary_dataset, model, vector_loss_fn, batch_size): | ||
| """Compute the loss on the canaries.""" | ||
| if canary_dataset is None: | ||
| return np.array([]) | ||
| all_losses = [] | ||
| for x, y, _ in canary_dataset.batch(batch_size, drop_remainder=False): | ||
| predictions = model(x, training=False) | ||
| loss_vector = vector_loss_fn(y, predictions) | ||
| all_losses.append(loss_vector.numpy()) | ||
| return np.concatenate(all_losses) | ||
|
|
||
|
|
||
| def evaluate_random_gradient_canary(canary_dataset, model): | ||
| """Compute the cosines of the parameters with the canaries.""" | ||
| if canary_dataset is None: | ||
| return np.array([]) | ||
| all_cosines = [] | ||
| weights = tf.nest.flatten(model.trainable_variables) | ||
| weight_norm = tf.sqrt( | ||
| tf.add_n( | ||
| tf.nest.map_structure( | ||
| lambda x: tf.linalg.norm(x) ** 2, tf.nest.flatten(weights) | ||
| ) | ||
| ) | ||
| ) | ||
| for _, _, z in canary_dataset: # all examples are canaries | ||
| # Note: We use canaries of norm = 1 because we normalize by the norm of | ||
| # canaries anyway in our final statistic. So the clip norm does not matter. | ||
| noise = utils.get_random_normal_like(weights, z, flat_l2_norm=1) | ||
| dot_product = tf.add_n( | ||
| tf.nest.map_structure(lambda a, b: tf.reduce_sum(a * b), noise, weights) | ||
| ) | ||
| cosine = dot_product / weight_norm | ||
| all_cosines.append(cosine.numpy()) | ||
| return np.array(all_cosines) | ||
|
|
||
|
|
||
| def evaluate_random_gradient_canary_batched( | ||
| canary_dataset, model, max_batch_size=1024 | ||
| ): | ||
| """Batched computation of the cosines of the parameters with the canaries.""" | ||
| # Batching gives a 20x speedup on the evaluation. | ||
| if canary_dataset is None: | ||
| return np.array([]) | ||
| all_cosines = [] | ||
| weights = tf.nest.flatten(model.trainable_variables) | ||
| weight_norm = tf.sqrt( | ||
| tf.add_n( | ||
| tf.nest.map_structure( | ||
| lambda x: tf.linalg.norm(x) ** 2, tf.nest.flatten(weights) | ||
| ) | ||
| ) | ||
| ) | ||
| # All examples are canaries, so no special filtering necessary. | ||
| for _, _, z in canary_dataset.batch(max_batch_size): | ||
| # Note: We use canaries of norm = 1 because we normalize by the norm of | ||
| # canaries anyway in our final statistic. So the clip norm does not matter. | ||
| noise = utils.get_batched_random_normal_like( | ||
| weights, z, flat_l2_norm=tf.constant(1.0) | ||
| ) # list of (batch_size, *weights[i]) | ||
| dot_product = tf.add_n( | ||
| tf.nest.map_structure(batched_dot, noise, weights) | ||
| ) # (batch_size,) | ||
| cosine = dot_product / weight_norm | ||
| all_cosines.append(cosine.numpy()) | ||
| return np.concatenate(all_cosines) | ||
|
|
||
|
|
||
| def batched_dot(a, b): | ||
| """Return [dot(c, b) for c in a] but in TF.""" | ||
| # a: (bsz, s1, s2, ...) | ||
| # b: (s1, s2, ...) | ||
| return tf.tensordot( | ||
| tf.reshape(a, (tf.shape(a)[0], -1)), # (bsz, s1, s2, ...) -> (bsz, s) | ||
| tf.reshape(b, -1), # (s1, s2, ...) -> (s,) | ||
| axes=1, | ||
| ) # (bsz,) |
Oops, something went wrong.