feat: Implement the clustering rewrite for fp32

Implements a clustering rewrite for fully connected layers for fp32 models

Resolves: MLIA-1079

Signed-off-by: Nathan Bailey <nathan.bailey@arm.com>
Change-Id: I4c12f0bf911219b4066f0760976e424ebe900a0b

2 files changed, 38 insertions, 5 deletions

diff --git a/src/mlia/nn/rewrite/core/rewrite.py b/src/mlia/nn/rewrite/core/rewrite.py
index a8084e8..6a3695a 100644
--- a/src/mlia/nn/rewrite/core/rewrite.py
+++ b/src/mlia/nn/rewrite/core/rewrite.py
@@ -23,6 +23,9 @@ from mlia.nn.common import Optimizer
 from mlia.nn.common import OptimizerConfiguration
 from mlia.nn.rewrite.core.train import train
 from mlia.nn.rewrite.core.train import TrainingParameters
+from mlia.nn.rewrite.library.fc_clustering_layer import (
+    get_keras_model_clus as fc_clustering_rewrite,
+)
 from mlia.nn.rewrite.library.fc_layer import get_keras_model as fc_rewrite
 from mlia.nn.rewrite.library.fc_sparsity24_layer import (
     get_keras_model as fc_rewrite_sparsity24,
@@ -63,6 +66,24 @@ class Rewrite(ABC):
         """Return default post-processing rewrite options."""
 
 
+class ClusteringRewrite(Rewrite):
+    """Graph clustering rewrite logic to be used by RewritingOptimizer."""
+
+    strip_pruning_wrapper = staticmethod(tfmot.clustering.keras.strip_clustering)
+
+    def quantize(self, model: keras.Model) -> keras.Model:
+        """Return a quantized model."""
+        return model
+
+    def post_process(self, model: keras.Model) -> keras.Model:
+        """Return the clustering stripped model."""
+        return self.strip_pruning_wrapper(model)
+
+    def training_callbacks(self) -> list:
+        """Return default rewrite callbacks."""
+        return []
+
+
 class QATRewrite(Rewrite):
     """Logic for rewrites requiring quantization-aware training."""
 
@@ -157,7 +178,7 @@ class RewritingOptimizer(Optimizer):
         [
             FullyConnectedRewrite("fully-connected", fc_rewrite),
             Sparsity24Rewrite("fully-connected-sparsity24", fc_rewrite_sparsity24),
-            FullyConnectedRewrite("fully-connected-clustering", fc_rewrite),
+            ClusteringRewrite("fully-connected-clustering", fc_clustering_rewrite),
         ]
     )
 
@@ -191,7 +212,6 @@ class RewritingOptimizer(Optimizer):
             raise ConfigurationError(
                 "Input and output tensor names need to be set for rewrite."
             )
-
         orig_vs_repl_stats, total_stats = train(
             source_model=tflite_model,
             unmodified_model=tflite_model if use_unmodified_model else None,
diff --git a/src/mlia/nn/rewrite/library/fc_clustering_layer.py b/src/mlia/nn/rewrite/library/fc_clustering_layer.py
index 07c07ac..72931c0 100644
--- a/src/mlia/nn/rewrite/library/fc_clustering_layer.py
+++ b/src/mlia/nn/rewrite/library/fc_clustering_layer.py
@@ -3,11 +3,24 @@
 """Example rewrite with one fully connected clustered layer."""
 from typing import Any
 
+import tensorflow_model_optimization as tfmot
 from keras.api._v2 import keras  # Temporary workaround for now: MLIA-1107
 
-from mlia.nn.rewrite.library.fc_layer import get_keras_model
-
 
 def get_keras_model_clus(input_shape: Any, output_shape: Any) -> keras.Model:
     """Generate TensorFlow Lite model for clustering rewrite."""
-    return get_keras_model(input_shape, output_shape)
+    clustering_params = {
+        "number_of_clusters": 32,
+        "cluster_centroids_init": tfmot.clustering.keras.CentroidInitialization.LINEAR,
+    }
+    model = tfmot.clustering.keras.cluster_weights(
+        to_cluster=keras.Sequential(
+            [
+                keras.layers.InputLayer(input_shape=input_shape),
+                keras.layers.Flatten(),
+                keras.layers.Dense(units=output_shape),
+            ]
+        ),
+        **clustering_params
+    )
+    return model