Include SepConv layer transforms and tests

Full support for SepConv layers 1D and 2D. Also adds
numerical tests to ensure validity.

PiperOrigin-RevId: 324901716

Author: nutsiepully

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

@@ -52,9 +52,13 @@ def apply(self, model, layer_quantize_map):
 
     transforms = [
         default_8bit_transforms.InputLayerQuantize(),
+        default_8bit_transforms.SeparableConv1DQuantize(),
         default_8bit_transforms.SeparableConvQuantize(),
+        default_8bit_transforms.Conv2DReshapeBatchNormReLUQuantize(),
+        default_8bit_transforms.Conv2DReshapeBatchNormActivationQuantize(),
         default_8bit_transforms.Conv2DBatchNormReLUQuantize(),
         default_8bit_transforms.Conv2DBatchNormActivationQuantize(),
+        default_8bit_transforms.Conv2DReshapeBatchNormQuantize(),
         default_8bit_transforms.Conv2DBatchNormQuantize(),
         default_8bit_transforms.ConcatTransform6Inputs(),
         default_8bit_transforms.ConcatTransform5Inputs(),

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

@@ -117,11 +117,39 @@ def _get_separable_conv2d_model(self):
     x = tf.keras.layers.ReLU()(x)
     return tf.keras.Model(i, x)
 
+  def _get_sepconv1d_bn_relu_model(self):
+    i = tf.keras.Input(shape=(8, 3))
+    x = tf.keras.layers.SeparableConv1D(
+        filters=5, kernel_size=3, strides=2)(i)
+    x = tf.keras.layers.BatchNormalization()(x)
+    x = tf.keras.layers.ReLU()(x)
+    return tf.keras.Model(i, x)
+
+  def _get_sepconv1d_bn_model(self):
+    i = tf.keras.Input(shape=(8, 3))
+    x = tf.keras.layers.SeparableConv1D(
+        filters=5, kernel_size=3, strides=2)(i)
+    x = tf.keras.layers.BatchNormalization()(x)
+    return tf.keras.Model(i, x)
+
+  def _get_sepconv1d_stacked_model(self):
+    i = tf.keras.Input(shape=(8, 3))
+    x = tf.keras.layers.SeparableConv1D(
+        filters=5, kernel_size=3, strides=2)(i)
+    x = tf.keras.layers.BatchNormalization()(x)
+    x = tf.keras.layers.SeparableConv1D(
+        filters=5, kernel_size=3, strides=2)(x)
+    x = tf.keras.layers.BatchNormalization()(x)
+    x = tf.keras.layers.ReLU()(x)
+    return tf.keras.Model(i, x)
+
   @parameterized.parameters([
       _get_single_conv_model, _get_single_dense_model,
       _get_single_conv_relu_model, _get_stacked_convs_model,
       _get_conv_bn_relu_model, _get_depthconv_bn_relu_model,
-      _get_separable_conv2d_model
+      _get_separable_conv2d_model,
+      _get_sepconv1d_bn_model, _get_sepconv1d_bn_relu_model,
+      _get_sepconv1d_stacked_model
   ])
   def testModelEndToEnd(self, model_fn):
     # 1. Check whether quantized model graph can be constructed.