Add transforms to ensure no faked quantization between Add and ReLU.

Without the transform, the quantized model convert to tensorflow lite as:
Inputs		Outputs	builtin_options		      			opcode
[169, 170, 37]	[171]	{'fused_activation_function': 1, ...}		CONV_2D
[171, 172, 70]	[173]	{...}						DEPTHWISE_CONV_2D
[173, 174, 38]	[175]	{...}						CONV_2D
[169, 175]	[176]	{'fused_activation_function': 0}		ADD
[176]		[177]	None						RELU
[177]		[178]	None						QUANTIZE

With the transform, RELU fused to ADD as:
Inputs		Outputs	builtin_options					opcode
[136, 137, 33]	[138]	{“fused_activation_function': 1, ...}		CONV_2D
[138, 139, 68]	[140]	{...}						DEPTHWISE_CONV_2D
[140, 141, 34]	[142]	{...}						CONV_2D
[136, 142]	[143]	{'fused_activation_function': 1}		ADD

PiperOrigin-RevId: 315794544

Author: tensorflower-gardener

tensorflow_model_optimization/python/core/quantization/keras/default_8bit/default_8bit_quantize_layout_transform.py

@@ -60,8 +60,9 @@ def apply(self, model, layer_quantize_map):
         default_8bit_transforms.ConcatTransform4Inputs(),
         default_8bit_transforms.ConcatTransform3Inputs(),
         default_8bit_transforms.ConcatTransform(),
+        default_8bit_transforms.AddReLUQuantize(),
+        default_8bit_transforms.AddActivationQuantize(),
     ]
-
     return model_transformer.ModelTransformer(
         model, transforms,
         layer_quantize_map.keys(), layer_quantize_map).transform()

tensorflow_model_optimization/python/core/quantization/keras/default_8bit/default_8bit_transforms.py

@@ -228,6 +228,37 @@ def pattern(self):
         inputs=[Conv2DBatchNormQuantize.pattern(self)])
 
 
+class AddReLUQuantize(transforms.Transform):
+  """Ensure FQ does not get placed between Add and ReLU."""
+
+  def pattern(self):
+    return LayerPattern('ReLU', inputs=[LayerPattern('Add')])
+
+  def replacement(self, match_layer):
+    relu_layer_node = match_layer
+    add_layer_node = relu_layer_node.input_layers[0]
+
+    add_layer_node.metadata['quantize_config'] = \
+      default_8bit_quantize_configs.NoOpQuantizeConfig()
+
+    return match_layer
+
+  def custom_objects(self):
+    return {
+        'NoOpQuantizeConfig': default_8bit_quantize_configs.NoOpQuantizeConfig,
+    }
+
+
+class AddActivationQuantize(AddReLUQuantize):
+  """Ensure FQ does not get placed between Add and ReLU."""
+
+  def pattern(self):
+    return LayerPattern(
+        'Activation',
+        config={'activation': 'relu'},
+        inputs=[LayerPattern('Add')])
+
+
 class InputLayerQuantize(transforms.Transform):
   """Quantizes InputLayer, by adding QuantizeLayer after it.
 

tensorflow_model_optimization/python/core/quantization/keras/default_8bit/default_8bit_transforms_test.py

@@ -221,6 +221,33 @@ def testConv2DBatchNormReLUQuantize(
     self.assertAllClose(
         transformed_model.predict(inputs), model.predict(inputs))
 
+  @parameterized.parameters(
+      ('relu', default_8bit_transforms.AddReLUQuantize),
+      ('act_relu', default_8bit_transforms.AddActivationQuantize),
+  )
+  def testAddReLUQuantize(self, activation_type, transform_type):
+    add = keras.layers.Add()
+    if activation_type == 'relu':
+      activation = keras.layers.ReLU(6.0)
+    elif activation_type == 'act_relu':
+      activation = keras.layers.Activation('relu')
+
+    inp1 = keras.layers.Input((3,))
+    inp2 = keras.layers.Input((3,))
+    x = activation(add([inp1, inp2]))
+    model = keras.Model([inp1, inp2], x)
+
+    transformed_model, updated_metadata = ModelTransformer(
+        model,
+        [transform_type()],
+    ).transform()
+
+    add_layer = transformed_model.layers[2]
+
+    self.assertIsInstance(
+        updated_metadata.get(add_layer.name).get('quantize_config'),
+        default_8bit_quantize_configs.NoOpQuantizeConfig)
+
   def testAddsQuantizeLayerAfterInputLayer(self):
     inp1 = keras.layers.Input((3,))
     inp2 = keras.layers.Input((3,))