Include InputLayer (when available) in Sequential config, so as not to lose a custom input name or dtype upon deserialization, if the user-defined model starts with an Input object.

PiperOrigin-RevId: 304062384

Author: fchollet

tensorflow_model_optimization/python/core/quantization/keras/quantize_test.py

@@ -305,7 +305,7 @@ def testQuantize_UsesBuiltinQuantizeConfig(self):
         quantize_annotate_layer(keras.layers.Dense(3, input_shape=(2,)))])
 
     quantized_model = quantize_apply(annotated_model)
-    quantized_layer = quantized_model.layers[0]
+    quantized_layer = quantized_model.layers[1]
 
     # 'activation' gets replaced while quantizing the model. Hence excluded
     # from equality checks.
@@ -327,7 +327,7 @@ def testQuantize_UsesQuantizeConfigFromUser_NoBuiltIn(self):
         '_TestQuantizeConfig': _TestQuantizeConfig
     }):
       quantized_model = quantize_apply(annotated_model)
-    quantized_layer = quantized_model.layers[0]
+    quantized_layer = quantized_model.layers[1]
 
     self._assert_layer_quantized(annotated_model.layers[0], quantized_layer)
     self.assertIsInstance(quantized_layer.quantize_config, _TestQuantizeConfig)
@@ -341,7 +341,7 @@ def testQuantize_PreferenceToUserSpecifiedQuantizeConfig(self):
 
     with custom_object_scope({'_TestQuantizeConfig': _TestQuantizeConfig}):
       quantized_model = quantize_apply(annotated_model)
-    quantized_layer = quantized_model.layers[0]
+    quantized_layer = quantized_model.layers[1]
 
     self._assert_layer_quantized(annotated_model.layers[0], quantized_layer)
     self.assertIsInstance(quantized_layer.quantize_config, _TestQuantizeConfig)