[converter] Fix mobilenet conversion problem

tfjs-converter/python/tensorflowjs/converters/fuse_depthwise_conv2d.py

@@ -21,12 +21,6 @@
 from __future__ import division
 from __future__ import print_function
 
-import tensorflow as tf
-
-from tensorflow.core.framework import graph_pb2
-from tensorflow.core.framework import node_def_pb2
-from tensorflow.python.framework import function
-
 from tensorflowjs.converters import graph_rewrite_util
 
 def _is_supported_activation(node):
@@ -149,41 +143,3 @@ def _fuse_depthwise_conv2d_with_match_function(input_graph_def, match_function):
 
   # No pattern detected
   return input_graph_def
-
-def extract_op_attributes(input_graph_def):
-  """Since TF does not allow function defined custom op to have any attributes,
-     we need to clean up the attributes for FusedDepthwiseConv2dNative op.
-  Args:
-    input_graph_def: Modified tf.Graph object.
-  """
-  result_graph_def = graph_pb2.GraphDef()
-  for node in input_graph_def.node:
-    new_node = node_def_pb2.NodeDef()
-    new_node.CopyFrom(node)
-    if new_node.op == graph_rewrite_util.FUSED_DEPTHWISE_CONV2D:
-      new_node.ClearField('attr')
-      if len(new_node.input) > 3:
-        new_node.input[:] = new_node.input[0:3]
-    result_graph_def.node.extend([new_node])
-  result_graph_def.versions.CopyFrom(input_graph_def.versions)
-  return result_graph_def
-
-def register_fused_depthwise_conv2d_func(graph):
-  """Register FusedDepthwiseConv2dNative op with function def, this is needed
-  for importing graph_def with unregistered op.
-  Args:
-    graph: A tf.Graph object to insert FusedDepthwiseConv2d function into.
-  """
-
-  # Create a function for FusedDepthwiseConv2dNative op
-  @function.Defun(tf.float32, tf.float32, tf.float32,
-                  func_name=graph_rewrite_util.FUSED_DEPTHWISE_CONV2D)
-  def fused_depthwise_conv2d_fn(*args):
-  # This is a placeholder for the Op definition, the exact implemenation of the
-  # function does not matter.
-    return tf.nn.depthwise_conv2d(
-        args[0], args[1], strides=[1, 1, 1, 1], padding='SAME')
-  # Insert the function into graph
-  with graph.as_default():
-    fused_depthwise_conv2d_fn(
-        tf.ones([1, 1, 1]), tf.ones([1, 1, 1]), tf.ones([1]))

tfjs-converter/python/tensorflowjs/converters/fuse_prelu.py

@@ -21,30 +21,11 @@
 from __future__ import division
 from __future__ import print_function
 
-import tensorflow as tf
-
 from tensorflow.core.framework import attr_value_pb2
-from tensorflow.core.framework import op_def_pb2
-from tensorflow.python.framework import function
-from tensorflow.python.framework import op_def_registry
 from tensorflow.python.framework import tensor_util
 
 from tensorflowjs.converters import graph_rewrite_util
 
-# The function is only needed for TensorFlow 1.X and 2.0. Remove once tfjs
-# no longer depends on these versions.
-def register_prelu_op():
-  """Register a virtual PReLU OpDef.
-
-  This allows to bypass MetaGraph validity checks on TensorFlow 1.X and 2.0.
-  """
-
-  prelu_op_def = op_def_pb2.OpDef()
-  prelu_op_def.name = 'Prelu'
-  missing_op_list = op_def_pb2.OpList()
-  missing_op_list.op.extend([prelu_op_def])
-  op_def_registry.register_op_list(missing_op_list)
-
 def fuse_ops_for_prelu(input_graph_def):
   """Modifies the provided graph by fusing a set of ops into a single Prelu op.
   The formula of PReLU is:
@@ -208,22 +189,4 @@ def fuse_prelu_with_fused_conv2d_or_matmul(input_graph_def):
 
   return graph_rewrite_util.cleanup_graph_def(
       input_graph_def, nodes_to_skip, inputs_to_remove)
-
-def register_prelu_func(graph):
-  """Register Prelu op with function def, this is needed for importing graph_def
-  with unregistered Prelu op.
-  Args:
-    graph: A tf.Graph object to insert prelu function into.
-  """
-
-  # Create a function for Prelu op
-  @function.Defun(tf.float32, tf.float32, func_name='Prelu')
-  def prelu_fn(*args):
-    return tf.add(args[0], args[1])
-  # Insert the function into graph
-  with graph.as_default():
-    prelu_fn(tf.constant(1.0), tf.constant(1.0))
-
-
-if hasattr(op_def_registry, 'register_op_list'):
-  register_prelu_op()
+

tfjs-converter/python/tensorflowjs/converters/tf_saved_model_conversion_v2.py

@@ -21,7 +21,6 @@
 import json
 import os
 
-import numpy as np
 import tensorflow as tf
 from tensorflow.core.framework import graph_pb2
 from tensorflow.core.framework import node_def_pb2
@@ -42,6 +41,7 @@
 from tensorflowjs.converters import fold_batch_norms
 from tensorflowjs.converters import fuse_prelu
 from tensorflowjs.converters import fuse_depthwise_conv2d
+from tensorflowjs.converters import graph_rewrite_util
 from tensorflowjs import resource_loader
 
 # enable eager execution for v2 APIs
@@ -126,7 +126,6 @@ def optimize_graph(graph, signature_def, output_graph,
     skip_op_check: Bool whether to skip the op check.
     strip_debug_ops: Bool whether to strip debug ops.
   """
-  fuse_prelu.register_prelu_func(graph)
 
   # Add a collection 'train_op' so that Grappler knows the outputs.
   for _, output in signature_def.outputs.items():
@@ -221,27 +220,17 @@ def extract_weights(graph_def,
   print('Writing weight file ' + output_graph + '...')
   const_manifest = []
 
-  graph = tf.Graph()
-  fuse_prelu.register_prelu_func(graph)
-  fuse_depthwise_conv2d.register_fused_depthwise_conv2d_func(graph)
-
-  extracted_graph = fuse_depthwise_conv2d.extract_op_attributes(graph_def)
-  with tf.compat.v1.Session(graph=graph) as sess:
-    tf.import_graph_def(extracted_graph, name='')
-    for const in constants:
-      tensor = graph.get_tensor_by_name(const.name + ':0')
-      value = tensor.eval(session=sess)
-      if not isinstance(value, np.ndarray):
-        value = np.array(value)
-
-      const_manifest.append({'name': const.name, 'data': value})
-
-      # Restore the conditional inputs
-      const.input[:] = const_inputs[const.name]
-
-      # Remove the binary array from tensor and save it to the external file.
-      for field_name in CLEARED_TENSOR_FIELDS:
-        const.attr["value"].tensor.ClearField(field_name)
+  for const in constants:
+    const_manifest.append({
+        'name': const.name,
+        'data': graph_rewrite_util.values_from_const(const)
+    })
+    # Restore the conditional inputs
+    const.input[:] = const_inputs[const.name]
+
+    # Remove the binary array from tensor and save it to the external file.
+    for field_name in CLEARED_TENSOR_FIELDS:
+      const.attr["value"].tensor.ClearField(field_name)
 
   write_artifacts(MessageToDict(graph_def), [const_manifest], output_graph,
                   tf_version, signature_def,