Only run Grappler optimizations when necessary in TFLiteConverter.

PiperOrigin-RevId: 246023134

tensorflow/lite/python/lite.py

@@ -157,7 +157,12 @@ class TFLiteConverterBase(object):
 
   def _grappler_config(self, target_ops):
     is_only_flex_enabled = set([OpsSet.SELECT_TF_OPS]) == target_ops
-    return _get_grappler_config(enable_layout_optimizer=is_only_flex_enabled)
+    if is_only_flex_enabled:
+      # The layout optimizer turns NHCW to NCHW. This provides performance
+      # optimizations when Flex mode is enabled. However, this is not compatible
+      # with builtin ops.
+      return _get_grappler_config(["layout"])
+    return None
 
   def _validate_representative_dataset(self):
     if self.representative_dataset:
@@ -339,12 +344,15 @@ class TFLiteConverterV2(TFLiteConverterBase):
     output_tensors = frozen_func.outputs
 
     # Run a Grappler pass.
-    graph_def = _run_graph_optimizations(
-        frozen_func.graph.as_graph_def(),
-        input_tensors,
-        output_tensors,
-        self._grappler_config(self.target_spec.supported_ops),
-        graph=frozen_func.graph)
+    graph_def = frozen_func.graph.as_graph_def()
+    config = self._grappler_config(self.target_spec.supported_ops)
+    if config:
+      graph_def = _run_graph_optimizations(
+          graph_def,
+          input_tensors,
+          output_tensors,
+          config,
+          graph=frozen_func.graph)
 
     # Checks dimensions in input tensor.
     for tensor in input_tensors:
@@ -862,14 +870,15 @@ class TFLiteConverter(TFLiteConverterBase):
         "dump_graphviz_video": self.dump_graphviz_video
     }
 
-    optimized_graph = None
-    if self.inference_type == constants.QUANTIZED_UINT8:
-      optimized_graph = self._graph_def
-    else:
+    optimized_graph = self._graph_def
+    if self.inference_type != constants.QUANTIZED_UINT8:
       try:
-        optimized_graph = _run_graph_optimizations(
-            self._graph_def, self._input_tensors, self._output_tensors,
-            self._grappler_config(self.target_ops))
+        config = self._grappler_config(self.target_ops)
+        if config:
+          optimized_graph = _run_graph_optimizations(self._graph_def,
+                                                     self._input_tensors,
+                                                     self._output_tensors,
+                                                     config)
       except Exception:
         optimized_graph = self._graph_def
 

tensorflow/lite/python/util.py

@@ -20,7 +20,6 @@ from __future__ import print_function
 
 from tensorflow.core.protobuf import config_pb2 as _config_pb2
 from tensorflow.core.protobuf import meta_graph_pb2 as _meta_graph_pb2
-from tensorflow.core.protobuf import rewriter_config_pb2 as _rewriter_config_pb2
 from tensorflow.lite.python.op_hint import convert_op_hints_to_stubs
 from tensorflow.lite.python.op_hint import find_all_hinted_output_nodes
 from tensorflow.lite.toco import types_pb2 as _types_pb2
@@ -148,33 +147,19 @@ def set_tensor_shapes(tensors, shapes):
           raise ValueError(message)
 
 
-def get_grappler_config(enable_layout_optimizer=False, function_only=False):
+def get_grappler_config(optimizers_list):
   """Creates a tf.compat.v1.ConfigProto for configuring Grappler.
 
   Args:
-    enable_layout_optimizer: Bool indicating whether to run the layout
-      optimizer. This turns NHCW to NCHW. This provides performance
-      optimizations when Flex mode is enabled. (default False)
-    function_only: Bool indiciating whether to only run the function optimizer.
-      This inlines functions and is required for freezing models with functions.
-      (default False)
+    optimizers_list: List of strings that represents the list of optimizers.
 
   Returns:
     tf.ConfigProto.
   """
   config = _config_pb2.ConfigProto()
   rewrite_options = config.graph_options.rewrite_options
-  if function_only:
-    rewrite_options.optimizers.append("function")
-  else:
-    if enable_layout_optimizer:
-      rewrite_options.layout_optimizer = _rewriter_config_pb2.RewriterConfig.ON
-    else:
-      rewrite_options.layout_optimizer = _rewriter_config_pb2.RewriterConfig.OFF
-
-    # Avoid remapping as it creates ops like _FusedConv2D, which are not
-    # supported by TFLite.
-    rewrite_options.remapping = _rewriter_config_pb2.RewriterConfig.OFF
+  for optimizer in optimizers_list:
+    rewrite_options.optimizers.append(optimizer)
   return config
 
 
@@ -242,7 +227,7 @@ def freeze_graph(sess, input_tensors, output_tensors):
     return _convert_op_hints_if_present(sess, output_tensors)
 
   # Runs a Grappler pass in order to inline any functions in the graph.
-  config = get_grappler_config(function_only=True)
+  config = get_grappler_config(["function"])
   graph_def = run_graph_optimizations(
       sess.graph_def, input_tensors, output_tensors, config, graph=sess.graph)