def tf2nx(argv: argparse.Namespace, model_file_name: str, output_model_name: str, outputs: list, output_dir: str,
scale: float, is_binary: bool,
user_shapes: [None, list, np.array] = None,
mean_scale_values: [dict, list] = ()):
"""
Convert TF GraphDef object to NetworkX representation.
The resulting graph is still TF-specific and needs normalization passes to be applied.
The specific TF structure assumes each GraphDef node is converted to a single
NetworkX node, node id is an original TF node name, and edges go directly from one op to another op.
"""
meta_info = get_meta_info(argv)
if argv.tensorflow_custom_layer_libraries:
libraries = argv.tensorflow_custom_layer_libraries.split(',')
for library in libraries:
log.info('Loading library "{}" with custom operations'.format(library))
tf.load_op_library(library)
graph_def, variables_values = load_tf_graph_def(graph_file_name=model_file_name, is_binary=is_binary,
checkpoint=argv.input_checkpoint,
user_output_node_names_list=outputs,
model_dir=argv.saved_model_dir,
meta_graph_file=argv.input_meta_graph,
saved_model_tags=argv.saved_model_tags)
try:
tf.import_graph_def(graph_def, name='')
except:
log.warning("TensorFlow post-processing of loaded model was unsuccessful. "
"This is an optional step that Model Optimizer performs for any input model but it is not usually "
"required for all models."
"It likely means that the original model is ill-formed. "
"Model Optimizer will continue converting this model.")
log.debug("Number of nodes in graph_def: {}".format(len(graph_def.node))) # pylint: disable=no-member
if argv.tensorboard_logdir:
tensorboard.dump_for_tensorboard(graph_def, argv.tensorboard_logdir)
update_extractors_with_extensions(tf_op_extractors)
try:
graph = protobuf2nx(graph_def)
graph.__setattr__('name', output_model_name)
# 'layout' parameter change may cause an issue in EltwiseInputReshape replacer
# and convert_nhwc_to_nchw(graph)
graph.graph['layout'] = 'NCHW' if argv.disable_nhwc_to_nchw else 'NHWC'
graph.graph['cmd_params'] = argv
graph.graph['fw'] = 'tf'
graph.graph['ir_version'] = 2 if argv.generate_deprecated_IR_V2 else 4
if graph.graph['ir_version'] == 2:
# When the deprecated IR version was requested,
# we configure only those phases that can lead to
# functional regressions in the version 2.
# BasicLSTMCell is one such transformation; when it is turned off,
# the body of TF basic_lstm_cell is converted as-is in a decomposed form,
# and should work in version 2.
BasicLSTMCell.enabled = False
# placeholder for request from a transformation pass to repeat the entire conversion
graph.graph['repeat_conversion'] = False
graph = restore_edges(graph, get_tf_edges)
graph = remove_control_dependency_inputs(graph)
# extract basic attributes earlier to enable some passes that relies on them before full attribute
# extractor is called
extract_node_attrs(graph, lambda node: (True, common_tf_fields(node)))
except Exception as e:
raise Error(
'Cannot pre-process TensorFlow graph after reading from model file "{}". ' \
'File is corrupt or has unsupported format. Details: {}. ' +
refer_to_faq_msg(44),
model_file_name,
str(e)
) from e
check_empty_graph(graph, 'protobuf2nx. It may happen due to problems with loaded model')
packed_user_shapes, packed_outputs, freeze_placeholder = user_data_repack(graph, user_shapes, outputs,
argv.freeze_placeholder_with_value)
if freeze_placeholder is not None:
FreezePlaceholderValue.enabled = True
FreezePlaceholderValue.replacement_dict = freeze_placeholder
update_registration()
GemmResolver.enabled = False
inputs = list(packed_user_shapes.keys()) if packed_user_shapes is not None and isinstance(packed_user_shapes,
dict) else None
graph.graph['inputs'] = inputs # save user defined inputs for other extensions
output_op_nodes = add_output_ops(graph, packed_outputs, inputs=packed_user_shapes)
input_op_nodes = add_input_ops(graph, packed_user_shapes, True)
# this call of 'graph_clean_up' removes child nodes of outputs which is useful when custom output is specified
graph_clean_up_tf(graph)
check_empty_graph(graph, 'add_output_ops and add_input_ops. It may happen due to absence of \'Placeholder\' layer '
'in the model')
variables_to_constants(graph, variables_values)
del variables_values
graph_clean_up_tf(graph)
if argv.tensorflow_custom_operations_config_update:
if update_custom_replacement_config_file(graph, argv.tensorflow_custom_operations_config_update):
return 0
else:
return 1
unsupported_ops_to_offload_to_tf = list()
MAX_ITERATIONS = 5
cur_iteration = 0
while cur_iteration < MAX_ITERATIONS:
graph_copy = copy.deepcopy(graph) # create a copy of graph for the case when some ops are unsupported
if argv.tensorflow_subgraph_patterns is not None:
csc.replace_subgraph_calls(graph, argv.tensorflow_subgraph_patterns)
if argv.tensorflow_operation_patterns is not None:
csc.offload_operations_to_tf(graph, argv.tensorflow_operation_patterns)
if argv.offload_unsupported_operations_to_tf and len(unsupported_ops_to_offload_to_tf):
csc.offload_unsupported_operations_to_tf(graph, unsupported_ops_to_offload_to_tf)
extract_node_attrs(graph, lambda node: tf_op_extractor(node, check_for_duplicates(tf_op_extractors)))
if argv.tensorflow_use_custom_operations_config is not None:
registry = CustomReplacementRegistry()
registry.add_custom_replacement_description_from_config(argv.tensorflow_use_custom_operations_config)
# automatically generate sub-classes for custom replacements that replace sub-graph with a single node
for replacement_desc in registry.get_all_replacements_descriptions():
if replacement_desc.has('op'):
type('FrontReplacementFromConfigFileOp' + replacement_desc.op, (FrontReplacementFromConfigFileOp,),
{'replacement_id': replacement_desc.id})
update_registration()
override_placeholder_shapes(graph, packed_user_shapes)
# the user shapes are used to convert TensorFlow Object Detection API models
graph.graph['user_shapes'] = packed_user_shapes
class_registration.apply_replacements(graph, class_registration.ClassType.FRONT_REPLACER)
override_batch(graph, argv.batch)
create_tensor_nodes(graph)
graph_clean_up_tf(graph)
remove_output_ops(graph)
partial_infer(graph)
delete_control_flow_edges(graph)
replacer = AddIsCyclicAttribute()
replacer.find_and_replace_pattern(graph)
# TENSOR ITERATOR CREATING BEGINS
if graph.graph['is_cyclic']:
replacer = DeleteSelect()
replacer.find_and_replace_pattern(graph)
replacer = SmartInputMatcher()
replacer.find_and_replace_pattern(graph)
replacer = SmartOutputMatcher()
replacer.find_and_replace_pattern(graph)
replacer = LoopConditionMatcher()
replacer.find_and_replace_pattern(graph)
replacer = SimpleConditionMather()
replacer.find_and_replace_pattern(graph)
replacer = BackEdgesMatching()
replacer.find_and_replace_pattern(graph)
replacer = ConditionChecks()
replacer.find_and_replace_pattern(graph)
delete_not_executable(graph)
graph_clean_up_tf(graph)
if graph.graph['is_cyclic']:
replacer = SimpleInputMatcher()
replacer.find_and_replace_pattern(graph)
replacer = BackEdgeSimpleInputMatcher()
replacer.find_and_replace_pattern(graph)
# Here will be optimizing path (ops after Enter and before body take out of body)
replacer = TensorIteratorMerge()
replacer.find_and_replace_pattern(graph)
# TENSOR ITERATOR CREATING ENDS
check_for_cycle(graph)
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
check_empty_graph(graph, 'partial_infer')
csc.prepare_tf_call_nodes(graph)
graph_clean_up_tf(graph)
duplicate_shared_weights(graph)
input_op_nodes = add_input_ops(graph, packed_user_shapes, False)
graph_clean_up_tf(graph)
check_empty_graph(graph, 'add_input_ops')
change_placeholders_types_to_FP32(graph)
scale_input(graph, scale)
add_mean_scale_values(graph, mean_scale_values)
convert_dilated_convolution(graph)
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
l2_norm_to_norm(graph)
graph_clean_up_tf(graph)
remove_op_nodes(graph, {'identity': True})
remove_useless_split(graph)
class_registration.apply_replacements(graph, class_registration.ClassType.MIDDLE_REPLACER)
mean_to_avgpool(graph)
convert_nasnet(graph)
fuse_pad(graph)
graph_clean_up_tf(graph)
convert_matmul_to_fully_connected(graph)
# Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes
for_graph_and_each_sub_graph_recursively(graph, lambda graph: mark_unfused_nodes(graph, argv.finegrain_fusing))
# Converting FusedBatchNorm layer to Mul->Add->Mul->Add sequence
# IE doesn't support BN with 4 inputs, so we have to split it to two ScaleShift
convert_batch_norm(graph)
graph_clean_up_tf(graph)
if not argv.disable_fusing:
# Converting ScaleShift layer to Mul->Add
for_graph_and_each_sub_graph_recursively(graph, convert_scale_shift_to_mul_add)
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
# Fusing the sequences of Mul/Add operations
for_graph_and_each_sub_graph_recursively(graph, fuse_mul_add_sequence)
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
# Fusing linear operation to Convolution
for_graph_and_each_sub_graph_recursively(graph, fuse_linear_ops)
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
if not argv.disable_gfusing:
grouped_convolutions_fusing(graph)
graph_clean_up_tf(graph)
if not argv.disable_fusing:
fuse_linear_ops(graph)
graph_clean_up_tf(graph)
# Converting Mul->Add to ScaleShift node
for_graph_and_each_sub_graph_recursively(graph, convert_muladd_to_scaleshift_or_power)
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
for_graph_and_each_sub_graph_recursively(graph, convert_mul_add_to_power)
# Need to eliminate dead nodes before doing update_fully_connected_shapes
# because update_fully_connected_shapes does partial inference and dead
# nodes will lead to sporadic failures.
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
for_graph_and_each_sub_graph_recursively(graph, update_fully_connected_shapes)
for_graph_and_each_sub_graph_recursively(graph, convert_mul_eltwise_to_leaky_relu)
graph_clean_up_tf(graph)
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
for_graph_and_each_sub_graph_recursively(graph, fuse_pad)
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
for_graph_and_each_sub_graph_recursively(graph, convert_reshape)
for_graph_and_each_sub_graph_recursively(graph, convert_squeeze)
for_graph_and_each_sub_graph_recursively(graph, convert_add_to_scaleshift) # scale = 1
for_graph_and_each_sub_graph_recursively(graph, convert_mul_to_scaleshift) # biases = 0
if argv.reverse_input_channels:
reverse_input_channels(graph)
if argv.move_to_preprocess:
move_scaleshift_to_preprocess(graph)
graph_clean_up_tf(graph)
for_graph_and_each_sub_graph_recursively(graph, fuse_sequence_of_reshapes)
pattern = EltwiseInputNormalize()
pattern.find_and_replace_pattern(graph)
conv_flatten_concat(graph)
for_graph_and_each_sub_graph_recursively(graph, apply_nhwc_to_nchw_permutation)
for_graph_and_each_sub_graph_recursively(graph, merge_nodes_permutations)
for_graph_and_each_sub_graph_recursively(graph, permute_data_nodes_attrs)
for_graph_and_each_sub_graph_recursively(graph, permute_op_nodes_attrs)
for_graph_and_each_sub_graph_recursively(graph, repack_fully_connected_weights_nhwc_to_nchw)
for_graph_and_each_sub_graph_recursively(graph, transpose_fully_connected_weights)
for_graph_and_each_sub_graph_recursively(graph, graph_clean_up_tf)
if argv.offload_unsupported_operations_to_tf:
unsupported_ops_to_offload_to_tf = find_unsupported_ops(graph)
if len(unsupported_ops_to_offload_to_tf) == 0:
log.info('All operations are supported! Exit from the loop.')
if not need_to_repeat_conversion(graph):
break
else:
print('After {} iteration there are {} unsupported ops'.format(cur_iteration + 1,
len(unsupported_ops_to_offload_to_tf)))
else:
if not need_to_repeat_conversion(graph):
break
graph = graph_copy
cur_iteration += 1
class_registration.apply_replacements(graph, class_registration.ClassType.BACK_REPLACER)
prepare_emit_ir(graph=graph, data_type=argv.data_type, output_dir=output_dir, output_model_name=output_model_name,
meta_info=meta_info)
return 0
def find_and_replace_pattern(self, graph: Graph):
variables_to_constants(graph, graph.graph['variables_values'])
del graph.graph['variables_values']