Sparsity Preserving DP-SGD in TF Privacy

Move get_registry_generator_fn from clip_grads.py to gradient_clipping_utils.py and change return type to dataclass.

See https://research.google/blog/sparsity-preserving-differentially-private-training/ for more details on the algorithm.

PiperOrigin-RevId: 660548431

tensorflow_privacy/privacy/fast_gradient_clipping/clip_grads.py

@@ -32,43 +32,6 @@ from tensorflow_privacy.privacy.fast_gradient_clipping import layer_registry as
 from tensorflow_privacy.privacy.fast_gradient_clipping import type_aliases
 
 
-def get_registry_generator_fn(
-    tape: tf.GradientTape,
-    layer_registry: lr.LayerRegistry,
-    num_microbatches: Optional[type_aliases.BatchSize] = None,
-):
-  """Creates the generator function for `compute_gradient_norms()`."""
-  if layer_registry is None:
-    # Needed for backwards compatibility.
-    registry_generator_fn = None
-  else:
-
-    def registry_generator_fn(layer_instance, args, kwargs):
-      if layer_instance.trainable_variables:
-        # Only trainable variables factor into the gradient.
-        if not layer_registry.is_elem(layer_instance):
-          raise NotImplementedError(
-              'Layer %s is not in the registry of known layers that can '
-              'be used for efficient gradient clipping.'
-              % layer_instance.__class__.__name__
-          )
-        registry_fn = layer_registry.lookup(layer_instance)
-        (layer_vars, layer_outputs, layer_sqr_norm_fn) = registry_fn(
-            layer_instance, args, kwargs, tape, num_microbatches
-        )
-        return layer_outputs, (
-            str(id(layer_instance)),
-            layer_vars,
-            layer_sqr_norm_fn,
-            layer_instance.trainable_weights,
-        )
-      else:
-        # Non-trainable layer.
-        return layer_instance(*args, **kwargs), None
-
-  return registry_generator_fn
-
-
 def _infer_per_example_loss_fn(model: tf.keras.Model):
   """Infer the per-example loss from model config."""
 
@@ -190,7 +153,7 @@ def compute_gradient_norms(
     are applied prior to clipping.
   """
   tape = tf.GradientTape(persistent=True, watch_accessed_variables=False)
-  registry_generator_fn = get_registry_generator_fn(
+  registry_generator_fn = gradient_clipping_utils.get_registry_generator_fn(
       tape, layer_registry, num_microbatches
   )
   # First loop computes the model outputs, summed loss, and generator outputs.
@@ -241,12 +204,17 @@ def compute_gradient_norms(
   #   $sqrt(k * \sum_i c_i^2)$ where $c_i$ is the norm estimate of its i-th
   #   occurrence. This is an over-estimate of the actual norm. For more details,
   #   see the explanation in go/dp-sgd-shared-weights.
-  for layer_id, v, f, weights_list in filtered_outputs:
+  for registry_fn_output in filtered_outputs:
     if trainable_vars is None or any(
-        w.ref() in trainable_vars for w in weights_list
+        w.ref() in trainable_vars
+        for w in registry_fn_output.layer_trainable_weights
     ):
-      layer_vars[layer_id].append(v)
-      layer_sqr_norm_fns[layer_id].append(f)
+      layer_vars[registry_fn_output.layer_id].append(
+          registry_fn_output.layer_vars
+      )
+      layer_sqr_norm_fns[registry_fn_output.layer_id].append(
+          registry_fn_output.layer_sqr_norm_fn
+      )
   # Second loop evaluates the squared L2 norm functions and appends the results.
   layer_grad_vars = tape.gradient(
       summed_loss,

tensorflow_privacy/privacy/fast_gradient_clipping/gradient_clipping_utils.py

@@ -13,14 +13,23 @@
 # limitations under the License.
 """Utility functions that help in the computation of per-example gradient norms."""
 
-from collections.abc import Sequence, Set
-from typing import Any, Optional
+from collections.abc import Callable, Sequence, Set
+import dataclasses
+from typing import Any, Optional, Tuple
 
 import tensorflow as tf
 from tensorflow_privacy.privacy.fast_gradient_clipping import layer_registry as lr
 from tensorflow_privacy.privacy.fast_gradient_clipping import type_aliases
 
 
+@dataclasses.dataclass(frozen=True)
+class RegistryGeneratorFunctionOutput:
+  layer_id: str
+  layer_vars: Optional[Sequence[tf.Variable]]
+  layer_sqr_norm_fn: Optional[type_aliases.SquareNormFunction]
+  layer_trainable_weights: Optional[Sequence[tf.Variable]]
+
+
 def has_internal_compute_graph(input_object: Any):
   """Checks if input is a TF model and has a TF internal compute graph."""
   return (
@@ -32,6 +41,63 @@ def has_internal_compute_graph(input_object: Any):
   )
 
 
+def get_registry_generator_fn(
+    tape: tf.GradientTape,
+    layer_registry: lr.LayerRegistry,
+    num_microbatches: Optional[type_aliases.BatchSize] = None,
+) -> Optional[Callable[..., Tuple[tf.Tensor, RegistryGeneratorFunctionOutput]]]:
+  """Creates the generator function for `model_forward_backward_pass()`.
+
+  Args:
+    tape: The `tf.GradientTape` to use for the gradient computation.
+    layer_registry: A `dict` of layers that support "fast" gradient norm
+      computations. The key is the class of the layer and the value is a
+      function that returns a `tuple` `(output, sqr_grad_norms, vars)`, where
+      `output` is the pre-activator tensor, `sqr_grad_norms` is related to the
+      squared norms of a layer's pre-activation tensor, and `vars` are relevant
+      trainable
+    num_microbatches: An optional number or scalar `tf.Tensor` for the number of
+      microbatches. If not None, indicates that the loss is grouped into
+      num_microbatches (in this case, the batch dimension needs to be a multiple
+      of num_microbatches).
+
+  Returns:
+    A function that returns a `tuple` `(output, sqr_grad_norms, vars)`, where
+    `output` is the pre-activator tensor, `sqr_grad_norms` is related to the
+    squared norms of a layer's pre-activation tensor, and `vars` are relevant
+    trainable variables.
+  """
+  if layer_registry is None:
+    # Needed for backwards compatibility.
+    registry_generator_fn = None
+  else:
+
+    def registry_generator_fn(layer_instance, args, kwargs):
+      if layer_instance.trainable_variables:
+        # Only trainable variables factor into the gradient.
+        if not layer_registry.is_elem(layer_instance):
+          raise NotImplementedError(
+              'Layer %s is not in the registry of known layers that can '
+              'be used for efficient gradient clipping.'
+              % layer_instance.__class__.__name__
+          )
+        registry_fn = layer_registry.lookup(layer_instance)
+        (layer_vars, layer_outputs, layer_sqr_norm_fn) = registry_fn(
+            layer_instance, args, kwargs, tape, num_microbatches
+        )
+        return layer_outputs, RegistryGeneratorFunctionOutput(
+            layer_id=str(id(layer_instance)),
+            layer_vars=layer_vars,
+            layer_sqr_norm_fn=layer_sqr_norm_fn,
+            layer_trainable_weights=layer_instance.trainable_weights,
+        )
+      else:
+        # Non-trainable layer.
+        return layer_instance(*args, **kwargs), None
+
+  return registry_generator_fn
+
+
 def model_forward_pass(
     input_model: tf.keras.Model,
     inputs: type_aliases.PackedTensors,