Adding the repeated dense graph patterns.

PiperOrigin-RevId: 668437140

kfac_jax/_src/curvature_estimator/block_diagonal.py

@@ -42,6 +42,7 @@
     conv2d=curvature_blocks.Conv2DTwoKroneckerFactored,
     generic=curvature_blocks.NaiveDiagonal,
     scale_and_shift=curvature_blocks.ScaleAndShiftDiagonal,
+    repeated_dense=curvature_blocks.RepeatedDenseKroneckerFactored,
 )
 
 
@@ -860,4 +861,3 @@ def to_diagonal_block_dense_matrix(self, state: State) -> tuple[Array, ...]:
   @utils.auto_scope_method
   def to_dense_matrix(self, state: State) -> Array:
     return scipy.linalg.block_diag(*self.to_diagonal_block_dense_matrix(state))
-

kfac_jax/_src/layers_and_loss_tags.py

@@ -13,11 +13,13 @@
 # limitations under the License.
 """K-FAC losses and layers tagging Jax primitives."""
 import dataclasses
+import functools
 from typing import Any, Generic, Sequence, TypeVar
 
 import jax
 from jax import core
 
+
 # Types for annotation
 T = TypeVar("T")
 Array = jax.Array
@@ -452,6 +454,12 @@ def register_scale_and_shift(
   )
 
 
+register_repeated_dense = functools.partial(
+    register_dense,
+    variant="repeated_dense",
+)
+
+
 class LossTagEqn(core.JaxprEqn):
   """A class used only for annotation purposes."""
   primitive: LossTag

kfac_jax/_src/tag_graph_matcher.py

@@ -1113,19 +1113,31 @@ def _dense_parameter_extractor(
   assert False
 
 
-def _make_dense_pattern(
+def _make_general_dense_pattern(
     with_bias: bool,
     reshape: bool,
+    num_repeated_axes: int,
     in_dim: int = 13,
     out_dim: int = 7,
 ) -> GraphPattern:
-  x_shape = [2, in_dim]
+  """Creates a pattern for a dense or repeated dense layer."""
+  in_axes = [0, [None, None]] if with_bias else [0, [None]]
+  f = _dense_with_reshape if reshape else _dense
+  for _ in range(num_repeated_axes):
+    f = jax.vmap(f, in_axes=in_axes)
+
+  x_shape = [9] * num_repeated_axes + [2, in_dim]
   p_shapes = ([[in_dim, out_dim], [out_dim]] if with_bias else
               [[in_dim, out_dim]])
+
+  name = "dense_with_bias" if with_bias else "dense_no_bias",
+  if num_repeated_axes > 0:
+    name = f"repeated[{num_repeated_axes}]_{name}"
+
   return GraphPattern(
-      name="dense_with_bias" if with_bias else "dense_no_bias",
+      name=name,
       tag_primitive=tags.layer_tag,
-      compute_func=_dense_with_reshape if reshape else _dense,
+      compute_func=f,
       parameters_extractor_func=_dense_parameter_extractor,
       example_args=[np.zeros(x_shape), [np.zeros(s) for s in p_shapes]],
   )
@@ -1338,9 +1350,15 @@ def _make_normalization_haiku_pattern(
 
 
 DEFAULT_GRAPH_PATTERNS = (
-    _make_dense_pattern(True, False),
-    _make_dense_pattern(True, True),
-    _make_dense_pattern(False, False),
+    _make_general_dense_pattern(True, False, 0),
+    _make_general_dense_pattern(True, False, 1),
+    _make_general_dense_pattern(True, False, 2),
+    _make_general_dense_pattern(True, True, 0),
+    _make_general_dense_pattern(True, True, 1),
+    _make_general_dense_pattern(True, True, 2),
+    _make_general_dense_pattern(False, False, 0),
+    _make_general_dense_pattern(False, False, 1),
+    _make_general_dense_pattern(False, False, 2),
     _make_conv2d_pattern(True),
     _make_conv2d_pattern(False),
     _make_scale_and_shift_pattern(1, True, True),