def driver(argv: argparse.Namespace):
model_proto = load_onnx_model(argv.input_model)
model_graph = model_proto.graph # pylint: disable=no-member
# print(model_graph)
# assert len(model_graph) == 1, "An ONNX model contains more than 1 graph: unsupported"
log.debug("Number of nodes in graph_def: {}".format(len(model_graph.node)))
log.debug(
"Number of all input ports (not true inputs) in graph_def: {}".format(
len(model_graph.input)))
log.debug("Number of initializers in graph_def: {}".format(
len(model_graph.initializer)))
log.debug("Number of real inputs in graph_def: {}".format(
len(model_graph.input) - len(model_graph.initializer)))
update_extractors_with_extensions(onnx_op_extractors)
try:
graph = protobuf2nx(model_proto)
log.debug("Number of nodes in NX graph: {}".format(
graph.number_of_nodes()))
graph.__setattr__(
'name',
argv.model_name if argv.model_name else model_proto.graph.name) # pylint: disable=no-member
graph.graph['layout'] = 'NCHW'
graph.graph['cmd_params'] = argv
graph.graph['fw'] = 'onnx'
graph.graph[
'feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3
graph.graph['ir_version'] = get_ir_version(argv)
except Exception as e:
raise Error(
'Cannot pre-process ONNX graph after reading from model file "{}". ' \
'File is corrupt or has unsupported format. Details: {}. ' +
refer_to_faq_msg(44),
argv.input_model,
str(e)
) from e
graph.check_empty_graph(
'protobuf2nx. It may happen due to problems with loaded model')
extract_node_attrs(
graph, lambda node: onnx_op_extractor(
node, check_for_duplicates(onnx_op_extractors)))
# --------------------------------- LOAD END ------------------------------------------------------
class_registration.apply_replacements(graph, [
class_registration.ClassType.FRONT_REPLACER,
class_registration.ClassType.MIDDLE_REPLACER,
class_registration.ClassType.BACK_REPLACER
])
return graph
def load(self, graph: Graph):
argv = graph.graph['cmd_params']
model_proto = load_onnx_model(argv.input_model)
model_graph = model_proto.graph # pylint: disable=no-member
# print(model_graph)
# assert len(model_graph) == 1, "An ONNX model contains more than 1 graph: unsupported"
log.debug("Number of nodes in graph_def: {}".format(
len(model_graph.node)))
log.debug(
"Number of all input ports (not true inputs) in graph_def: {}".
format(len(model_graph.input)))
log.debug("Number of initializers in graph_def: {}".format(
len(model_graph.initializer)))
log.debug("Number of real inputs in graph_def: {}".format(
len(model_graph.input) - len(model_graph.initializer)))
update_extractors_with_extensions(onnx_op_extractors)
try:
protobuf2nx(graph, model_proto)
except Exception as e:
raise Error(
'Cannot pre-process ONNX graph after reading from model file "{}". ' \
'File is corrupt or has unsupported format. Details: {}. ' +
refer_to_faq_msg(44),
argv.input_model,
str(e)
) from e
log.debug("Number of nodes in NX graph: {}".format(
graph.number_of_nodes()))
graph.__setattr__(
'name',
argv.model_name if argv.model_name else model_proto.graph.name) # pylint: disable=no-member
graph.graph['layout'] = 'NCHW'
graph.graph['fw'] = 'onnx'
graph.graph['feature_dim'] = 1
if hasattr(model_proto, 'opset_import'):
graph.graph['fw_opset_version'] = model_proto.opset_import[
0].version # pylint: disable=no-member
else:
graph.graph['fw_opset_version'] = None
graph.check_empty_graph(
'protobuf2nx. It may happen due to problems with loaded model')
extract_node_attrs(
graph, lambda node: onnx_op_extractor(
node, check_for_duplicates(onnx_op_extractors)))
send_op_names_info('onnx', graph)
send_shapes_info('onnx', graph)
def driver_R1(onnx_modelproto_bytes,
precision: str,
output_model_name: str,
outputs: list,
output_dir: str,
scale: float,
user_shapes: [None, list, np.array] = None,
mean_scale_values: [dict, list] = ()):
try:
model_proto = onnx.load_from_string(bytes(onnx_modelproto_bytes))
except Exception as e:
print("[python] onnx exception: ", str(e))
model_graph = model_proto.graph # pylint: disable=no-member
update_extractors_with_extensions(onnx_op_extractors)
try:
graph = protobuf2nx(model_proto)
log.debug("Number of nodes in NX graph: {}".format(
graph.number_of_nodes()))
graph.__setattr__(
'name',
output_model_name if output_model_name else model_proto.graph.name) # pylint: disable=no-member
graph.graph['layout'] = 'NCHW'
graph.graph['cmd_params'] = argparse.Namespace(
batch=None,
data_type='float',
disable_fusing=False,
disable_gfusing=False,
disable_resnet_optimization=False,
enable_concat_optimization=False,
extensions=mo_extensions,
finegrain_fusing=None,
framework='onnx',
freeze_placeholder_with_value=None,
generate_deprecated_IR_V2=False,
input=None,
input_model=None,
input_shape=None,
keep_shape_ops=False,
log_level='ERROR',
mean_scale_values={},
mean_values=(),
model_name=None,
move_to_preprocess=False,
output=None,
output_dir='.',
placeholder_shapes=None,
reverse_input_channels=False,
scale=None,
scale_values=(),
silent=False,
version=False)
graph.graph['fw'] = 'onnx'
graph.graph[
'feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3
graph.graph['ir_version'] = 5
except Exception as e:
raise Error(
'Cannot pre-process ONNX 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
graph.check_empty_graph(
'protobuf2nx. It may happen due to problems with loaded model')
extract_node_attrs(
graph, lambda node: onnx_op_extractor(
node, check_for_duplicates(onnx_op_extractors)))
# --------------------------------- LOAD END ------------------------------------------------------
class_registration.apply_replacements(
graph, class_registration.ClassType.FRONT_REPLACER)
class_registration.apply_replacements(
graph, class_registration.ClassType.MIDDLE_REPLACER)
fuse_pad(graph)
graph_clean_up_onnx(graph)
mark_unfused_nodes(graph, 'False')
convert_batch_norm(graph)
graph_clean_up_onnx(graph)
convert_scale_shift_to_mul_add(graph)
graph_clean_up_onnx(graph)
fuse_mul_add_sequence(graph)
graph_clean_up_onnx(graph)
fuse_linear_ops(graph)
graph_clean_up_onnx(graph)
grouped_convolutions_fusing(graph)
graph_clean_up_onnx(graph)
fuse_linear_ops(graph)
graph_clean_up_onnx(graph)
convert_muladd_to_scaleshift_or_power(graph)
graph_clean_up_onnx(graph)
convert_mul_add_to_power(graph)
graph_clean_up_onnx(graph)
convert_reshape(graph)
graph_clean_up_onnx(graph)
convert_add_or_mul_to_scaleshift(graph) # scale = 1
graph_clean_up_onnx(graph)
fuse_pad(graph)
graph_clean_up_onnx(graph)
fuse_sequence_of_reshapes(graph)
graph_clean_up_onnx(graph)
pattern = EltwiseInputNormalize()
pattern.find_and_replace_pattern(graph)
merge_nodes_permutations(graph)
permute_data_nodes_attrs(graph)
permute_op_nodes_attrs(graph)
class_registration.apply_replacements(
graph, class_registration.ClassType.BACK_REPLACER)
for_graph_and_each_sub_graph_recursively(graph, remove_const_ops)
CreateConstNodesReplacement().find_and_replace_pattern(graph)
for_graph_and_each_sub_graph_recursively(graph, remove_output_ops)
weights, xml_string = prepare_emit_ir(graph=graph,
data_type=precision,
output_dir=output_dir,
output_model_name=output_model_name,
meta_info={'unset': []})
return weights, xml_string
def driver_R5(onnx_modelproto_bytes,
precision: str,
output_model_name: str,
outputs: list,
output_dir: str,
scale: float,
user_shapes: [None, list, np.array] = None,
mean_scale_values: [dict, list] = ()):
try:
model_proto = onnx.load_from_string(bytes(onnx_modelproto_bytes))
except Exception as e:
print("[python] onnx exception: ", str(e))
model_graph = model_proto.graph # pylint: disable=no-member
log.debug("Number of nodes in graph_def: {}".format(len(model_graph.node)))
log.debug(
"Number of all input ports (not true inputs) in graph_def: {}".format(
len(model_graph.input)))
log.debug("Number of initializers in graph_def: {}".format(
len(model_graph.initializer)))
log.debug("Number of real inputs in graph_def: {}".format(
len(model_graph.input) - len(model_graph.initializer)))
update_extractors_with_extensions(onnx_op_extractors)
try:
graph = protobuf2nx(model_proto)
log.debug("Number of nodes in NX graph: {}".format(
graph.number_of_nodes()))
graph.__setattr__(
'name',
output_model_name if output_model_name else model_proto.graph.name) # pylint: disable=no-member
graph.graph['layout'] = 'NCHW'
graph.graph['fw'] = 'onnx'
graph.graph[
'feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3
graph.graph['ir_version'] = 4
extract_node_attrs(graph, lambda node:
(True, common_onnx_fields(node)))
except Exception as e:
raise Error(
'Cannot pre-process ONNX 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, _ = user_data_repack(
graph, user_shapes, outputs, None)
output_op_nodes = add_output_ops(graph, packed_outputs)
input_op_nodes = add_input_ops(graph, packed_user_shapes, True)
graph_clean_up(graph)
check_empty_graph(graph, 'add_output_ops and add_input_ops')
extract_node_attrs(
graph, lambda node: onnx_op_extractor(
node, check_for_duplicates(onnx_op_extractors)))
class_registration.apply_replacements(
graph, class_registration.ClassType.FRONT_REPLACER)
create_tensor_nodes(graph)
graph_clean_up(graph)
override_placeholder_shapes(graph, packed_user_shapes)
graph_clean_up(graph)
remove_op_nodes(graph, {'op': 'Identity'})
graph_clean_up(graph)
remove_output_ops(graph)
partial_infer(graph)
graph_clean_up(graph)
check_empty_graph(graph, 'partial_infer')
input_op_nodes = add_input_ops(graph, packed_user_shapes, False)
graph_clean_up(graph)
check_empty_graph(graph, 'add_input_ops')
scale_input(graph, scale)
add_mean_scale_values(graph, mean_scale_values)
convert_dilated_convolution(graph)
graph_clean_up(graph)
graph_clean_up(graph)
remove_op_nodes(graph, {'op': 'Identity'})
remove_useless_split(graph)
class_registration.apply_replacements(
graph, class_registration.ClassType.MIDDLE_REPLACER)
convert_gemm_to_fully_connected(graph)
NormalizeFullyConnected().find_and_replace_pattern(graph)
fuse_pad(graph)
graph_clean_up(graph)
convert_batch_norm(graph)
graph_clean_up(graph)
convert_scale_shift_to_mul_add(graph)
graph_clean_up(graph)
fuse_mul_add_sequence(graph)
graph_clean_up(graph)
fuse_linear_ops(graph)
graph_clean_up(graph)
grouped_convolutions_fusing(graph)
graph_clean_up(graph)
fuse_linear_ops(graph)
graph_clean_up(graph)
convert_muladd_to_scaleshift_or_power(graph)
graph_clean_up(graph)
convert_mul_add_to_power(graph)
graph_clean_up(graph)
convert_reshape(graph)
convert_add_to_scaleshift(graph) # scale = 1
convert_mul_to_scaleshift(graph) # biases = 0
fuse_pad(graph)
graph_clean_up(graph)
fuse_sequence_of_reshapes(graph)
graph_clean_up(graph)
pattern = EltwiseInputNormalize()
pattern.find_and_replace_pattern(graph)
merge_nodes_permutations(graph)
permute_data_nodes_attrs(graph)
permute_op_nodes_attrs(graph)
class_registration.apply_replacements(
graph, class_registration.ClassType.BACK_REPLACER)
weights, xml_string = prepare_emit_ir(graph=graph,
data_type=precision,
output_dir=output_dir,
output_model_name=output_model_name,
meta_info={'unset': []})
return weights, xml_string
def driver(argv: argparse.Namespace, model_file_name: str, output_model_name: str, output_dir: str):
meta_info = get_meta_info(argv)
model_proto = load_onnx_model(model_file_name)
model_graph = model_proto.graph # pylint: disable=no-member
# print(model_graph)
# assert len(model_graph) == 1, "An ONNX model contains more than 1 graph: unsupported"
log.debug("Number of nodes in graph_def: {}".format(len(model_graph.node)))
log.debug("Number of all input ports (not true inputs) in graph_def: {}".format(len(model_graph.input)))
log.debug("Number of initializers in graph_def: {}".format(len(model_graph.initializer)))
log.debug("Number of real inputs in graph_def: {}".format(len(model_graph.input) - len(model_graph.initializer)))
update_extractors_with_extensions(onnx_op_extractors)
try:
graph = protobuf2nx(model_proto)
log.debug("Number of nodes in NX graph: {}".format(graph.number_of_nodes()))
graph.__setattr__('name',
output_model_name if output_model_name else model_proto.graph.name) # pylint: disable=no-member
graph.graph['layout'] = 'NCHW'
graph.graph['cmd_params'] = argv
graph.graph['fw'] = 'onnx'
graph.graph['feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3
graph.graph['ir_version'] = 2 if argv.generate_deprecated_IR_V2 else 5
except Exception as e:
raise Error(
'Cannot pre-process ONNX 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
graph.check_empty_graph('protobuf2nx. It may happen due to problems with loaded model')
extract_node_attrs(graph, lambda node: onnx_op_extractor(node, check_for_duplicates(onnx_op_extractors)))
# --------------------------------- LOAD END ------------------------------------------------------
class_registration.apply_replacements(graph, class_registration.ClassType.FRONT_REPLACER)
class_registration.apply_replacements(graph, class_registration.ClassType.MIDDLE_REPLACER)
fuse_pad(graph)
graph_clean_up_onnx(graph)
# Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes
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_onnx(graph)
if not argv.disable_fusing:
# Converting ScaleShift layer to Mul->Add
convert_scale_shift_to_mul_add(graph)
graph_clean_up_onnx(graph)
# Fusing the sequences of Mul/Add operations
fuse_mul_add_sequence(graph)
graph_clean_up_onnx(graph)
# Fusing linear operation to Convolution
fuse_linear_ops(graph)
graph_clean_up_onnx(graph)
if not argv.disable_gfusing:
grouped_convolutions_fusing(graph)
graph_clean_up_onnx(graph)
if not argv.disable_fusing:
fuse_linear_ops(graph)
graph_clean_up_onnx(graph)
AddQuantizeFuse().find_and_replace_pattern(graph)
MulQuantizeFuse().find_and_replace_pattern(graph)
convert_muladd_to_scaleshift_or_power(graph)
graph_clean_up_onnx(graph)
convert_mul_add_to_power(graph)
graph_clean_up_onnx(graph)
convert_reshape(graph)
graph_clean_up_onnx(graph)
convert_add_or_mul_to_scaleshift(graph) # scale = 1
graph_clean_up_onnx(graph)
fuse_pad(graph)
graph_clean_up_onnx(graph)
if argv.reverse_input_channels:
reverse_input_channels(graph)
if argv.move_to_preprocess:
move_scaleshift_to_preprocess(graph)
graph_clean_up_onnx(graph)
fuse_sequence_of_reshapes(graph)
graph_clean_up_onnx(graph)
pattern = EltwiseInputNormalize()
pattern.find_and_replace_pattern(graph)
merge_nodes_permutations(graph)
permute_data_nodes_attrs(graph)
permute_op_nodes_attrs(graph)
class_registration.apply_replacements(graph, class_registration.ClassType.BACK_REPLACER)
for_graph_and_each_sub_graph_recursively(graph, remove_const_ops)
CreateConstNodesReplacement().find_and_replace_pattern(graph)
for_graph_and_each_sub_graph_recursively(graph, remove_output_ops)
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 driver(onnx_modelproto_bytes, precision: str, output_model_name: str,
output_dir: str):
try:
model_proto = onnx.load_from_string(bytes(onnx_modelproto_bytes))
except Exception as e:
print("[python] onnx exception: ", str(e))
model_graph = model_proto.graph # pylint: disable=no-member
log.debug("Number of nodes in graph_def: {}".format(len(model_graph.node)))
log.debug(
"Number of all input ports (not true inputs) in graph_def: {}".format(
len(model_graph.input)))
log.debug("Number of initializers in graph_def: {}".format(
len(model_graph.initializer)))
log.debug("Number of real inputs in graph_def: {}".format(
len(model_graph.input) - len(model_graph.initializer)))
update_extractors_with_extensions(onnx_op_extractors)
try:
graph = protobuf2nx(model_proto)
log.debug("Number of nodes in NX graph: {}".format(
graph.number_of_nodes()))
graph.__setattr__(
'name',
output_model_name if output_model_name else model_proto.graph.name) # pylint: disable=no-member
graph.graph['layout'] = 'NCHW'
graph.graph['fw'] = 'onnx'
graph.graph[
'feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3
graph.graph['cmd_params'] = argparse.Namespace(
batch=None,
data_type='float',
disable_fusing=False,
disable_gfusing=False,
disable_resnet_optimization=False,
enable_concat_optimization=False,
extensions=mo_extensions,
finegrain_fusing=None,
framework='onnx',
freeze_placeholder_with_value=None,
generate_deprecated_IR_V2=False,
input=None,
input_model=None,
input_shape=None,
keep_shape_ops=False,
log_level='ERROR',
mean_scale_values={},
mean_values=(),
model_name=None,
move_to_preprocess=False,
output=None,
output_dir='.',
placeholder_shapes=None,
reverse_input_channels=False,
scale=None,
scale_values=(),
silent=False,
version=False,
blobs_as_inputs=False,
keep_quantize_ops_in_IR=False,
generate_experimental_IR_V10=False)
graph.graph['ir_version'] = 6
except Exception as e:
raise Error(
'Cannot pre-process ONNX 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
graph.check_empty_graph(
'protobuf2nx. It may happen due to problems with loaded model')
extract_node_attrs(
graph, lambda node: onnx_op_extractor(
node, check_for_duplicates(onnx_op_extractors)))
# --------------------------------- LOAD END ------------------------------------------------------
class_registration.apply_replacements(
graph, class_registration.ClassType.FRONT_REPLACER)
class_registration.apply_replacements(
graph, class_registration.ClassType.MIDDLE_REPLACER)
fuse_pad(graph)
graph_clean_up_onnx(graph)
for_graph_and_each_sub_graph_recursively(
graph, convert_matmul_to_fully_connected)
# Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes
mark_unfused_nodes(graph, False)
# 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_onnx(graph)
# Converting ScaleShift layer to Mul->Add
convert_scale_shift_to_mul_add(graph)
graph_clean_up_onnx(graph)
# Fusing the sequences of Mul/Add operations
fuse_mul_add_sequence(graph)
graph_clean_up_onnx(graph)
# Fusing linear operation to Convolution
fuse_linear_ops(graph)
graph_clean_up_onnx(graph)
grouped_convolutions_fusing(graph)
graph_clean_up_onnx(graph)
fuse_linear_ops(graph)
graph_clean_up_onnx(graph)
MarkNodesToFuseUpToFakeQuantize().find_and_replace_pattern(graph)
FakeQuantizeFuse().find_and_replace_pattern(graph)
AddFakeQuantizeFuse().find_and_replace_pattern(graph)
MulFakeQuantizeFuse().find_and_replace_pattern(graph)
convert_muladd_to_scaleshift(graph)
graph_clean_up_onnx(graph)
graph_clean_up_onnx(graph)
convert_add_or_mul_to_scaleshift(graph) # scale = 1
graph_clean_up_onnx(graph)
fuse_pad(graph)
graph_clean_up_onnx(graph)
FuseReshapesSequence().find_and_replace_pattern(graph)
RemoveRedundantReshapes().find_and_replace_pattern(graph)
graph_clean_up_onnx(graph)
pattern = EltwiseInputNormalize()
pattern.find_and_replace_pattern(graph)
merge_nodes_permutations(graph)
permute_data_nodes_attrs(graph)
permute_op_nodes_attrs(graph)
graph_clean_up_onnx(graph)
class_registration.apply_replacements(
graph, class_registration.ClassType.BACK_REPLACER)
for_graph_and_each_sub_graph_recursively(graph, remove_const_ops)
CreateConstNodesReplacement().find_and_replace_pattern(graph)
for_graph_and_each_sub_graph_recursively(graph, remove_output_ops)
weights, xml_string = prepare_emit_ir(graph=graph,
data_type=precision,
output_dir=output_dir,
output_model_name=output_model_name,
meta_info={'unset': []})
return weights, xml_string
def extract(cls, loop_node):
Loop.update_node_stat(loop_node, {})
body_graph_proto = onnx_attr(loop_node, 'body', 'g', None)
main_graph = loop_node.graph
# create a Graph object for the body and take graph attributes from the main graph
body_graph = Graph()
main_graph_attrs_copy = copy.deepcopy(main_graph.graph)
del main_graph_attrs_copy['tensor_mapping']
body_graph.graph.update(main_graph_attrs_copy)
loop_node['body'] = body_graph
# maps a tensor name to a node produced it and the node port: str -> (node_id, node_port)
data_nodes_map = {}
body_graph.graph[
'tensor_mapping'] = data_nodes_map # save mapping for possible Loop inside the Loop
body_parameters = add_initializers_and_inputs_to_graph(
body_graph, body_graph_proto, data_nodes_map)
external_edges = [
] # (src_node, src_out_port), dest_body_parameter_node
additional_params = {
} # (src_node, src_out_port) -> parameter_node (for manually added Parameters)
# Go through all nodes in the original model order because data nodes are defined on-the-fly and order matters
for pb_node in body_graph_proto.node:
# create an NX node
id = body_graph.unique_id(node_id(pb_node))
body_graph.add_node(id, pb=pb_node, kind='op')
# add incoming edges based on data_nodes_map
for dst_port, inp in enumerate(pb_node.input):
# should add edge inp --> id
if inp not in data_nodes_map:
if inp == '':
# input is omitted; most likely it corresponds to an optional input for an operator
continue
elif inp in main_graph.graph['tensor_mapping']:
log.debug(
'The edge between outer and inner graphs detected: {} -> {}'
.format(inp, id))
if main_graph.graph['tensor_mapping'][
inp] not in additional_params:
# create new Parameter body node and connect the body node with the outer graph using it
param_id = str(inp)
body_graph.add_node(param_id,
kind='op',
op='Parameter',
name=param_id,
pb=None,
shape=None)
parameter_node = Node(body_graph, param_id)
# need to manually update necessary attrs for the node because extractor will not be called
# for it because the node does not have .pb attribute
Parameter.update_node_stat(parameter_node, {})
external_edges.append(
(main_graph.graph['tensor_mapping'][inp],
parameter_node))
src_id, src_port = param_id, 0
additional_params[main_graph.graph[
'tensor_mapping'][inp]] = parameter_node
else:
src_id, src_port = additional_params[
main_graph.graph['tensor_mapping'][inp]].id, 0
else:
raise Error(
'Reference to "{}" is not satisfied. A node refer not existing data tensor. ONNX '
'model is not consistent. Protobuf fragment: {}',
inp, pb_node)
else:
src_id, src_port = data_nodes_map[inp]
assert (body_graph.has_node(src_id))
edge_attrs = {
'out': src_port,
'in': dst_port,
'name': inp,
'fw_tensor_debug_info': [(inp, inp)],
'in_attrs': ['in', 'name'],
'out_attrs': ['out', 'name'],
'data_attrs': ['fw_tensor_debug_info']
}
body_graph.add_edge(src_id, id, **edge_attrs)
# add outgoing edges to data_nodes_map
for src_port, out in enumerate(pb_node.output):
if out in data_nodes_map:
log.debug("Detected reuse of blob {}.".format(out))
data_nodes_map[out] = (id, src_port)
body_results = []
for output in body_graph_proto.output:
tensor_name = str(output.name)
node_name, output_port = data_nodes_map[tensor_name]
assert body_graph.has_node(
node_name
), 'The body graph does not contain output with name "{}"'.format(
node_name)
body_results.append(
Node(body_graph,
add_opoutput(body_graph, node_name, output_port, False)))
# add 'internal_layer_id' attribute which is a must have attribute for the loop body node
for idx, body_node in enumerate(body_graph.get_op_nodes()):
body_node['internal_layer_id'] = idx
loop_carried_dependencies_count = len(body_graph_proto.input) - 2
scan_outputs_count = len(
body_graph_proto.output) - 1 - loop_carried_dependencies_count
# Loop inputs:
# 0 - trip count
# 1 - execution condition
# 2 .. - loop carried dependencies
# Loop outputs:
# 0 .. loop_carried_dependencies_count - 1 - loop carried dependencies
# loop_carried_dependencies_count .. - scan outputs
# Body inputs:
# 0 - iteration number
# 1 - execution condition
# 2 .. - loop carried dependencies
# Body outputs:
# 0 - execution condition
# 1 .. loop_carried_dependencies_count - loop carried dependencies
# loop_carried_dependencies_count + 1 .. - scan outputs
body_graph.stage = 'front'
# some of the inputs/outputs may not be connected but the normalization transformation will take care of it
# connection Loop body nodes with external input edges
next_loop_input_port_idx = sorted(loop_node.in_edges().keys())[-1] + 1
for (src_node, src_port), body_node in external_edges:
main_graph.add_edge(
src_node, loop_node.id, **{
'out': src_port,
'in': next_loop_input_port_idx,
'name': src_node,
'fw_tensor_debug_info': [(src_node, src_node)],
'in_attrs': ['in', 'name'],
'out_attrs': ['out', 'name'],
'data_attrs': ['fw_tensor_debug_info']
})
connect_body_input(loop_node, next_loop_input_port_idx, body_node)
next_loop_input_port_idx += 1
# mark current iteration input Parameter node
Loop.mark_current_iteration_parameter_node(loop_node,
body_parameters[0])
# connect initial value for "execution condition" input of the loop
connect_body_input(loop_node, 1, body_parameters[1])
# add back edge with "execution condition"
Loop.add_back_edge(loop_node, body_parameters[1], body_results[0])
# mark "execution condition" Result node
Loop.mark_execution_condition_result_node(loop_node, body_results[0])
# connect initial value for "loop carried" dependencies variables
for idx in range(loop_carried_dependencies_count):
connect_body_input(loop_node, idx + 2, body_parameters[idx + 2])
# add back edge for "loop carried" dependencies variables
for idx in range(loop_carried_dependencies_count):
Loop.add_back_edge(loop_node, body_parameters[idx + 2],
body_results[idx + 1])
# connect final value for "loop carried" dependencies variables
for idx in range(loop_carried_dependencies_count):
connect_body_output(loop_node, idx, body_results[idx + 1])
# connect "scan outputs" and mark axis for concatenation
for idx in range(loop_carried_dependencies_count,
loop_carried_dependencies_count + scan_outputs_count):
connect_body_output(loop_node, idx, body_results[idx + 1], axis=0)
# run function to parse body nodes attributes similar to the main graph
extract_node_attrs(
body_graph, lambda node: onnx_op_extractor(
node, check_for_duplicates(onnx_op_extractors)))
return cls.enabled
def extract(cls, loop_node):
Loop.update_node_stat(loop_node, {})
body_graph_proto = onnx_attr(loop_node, 'body', 'g', None)
main_graph = loop_node.graph
# create a Graph object for the body and take graph attributes from the main graph
body_graph = Graph()
main_graph_attrs_copy = {}
for attr_key, attr_value in main_graph.graph.items():
if attr_key not in ['tensor_mapping', 'parent_node']:
main_graph_attrs_copy[attr_key] = copy.deepcopy(attr_value)
body_graph.graph.update(main_graph_attrs_copy)
loop_node['body'] = body_graph
# save parent node for nested loops to know which node contains body (and which graph is on upper level)
body_graph.graph['parent_node'] = loop_node
# maps a tensor name to a node produced it and the node port: str -> (node_id, node_port)
data_nodes_map = {}
body_graph.graph['tensor_mapping'] = data_nodes_map # save mapping for possible Loop inside the Loop
body_parameters = add_initializers_and_inputs_to_graph(body_graph, body_graph_proto, data_nodes_map)
external_edges = [] # (src_node, src_out_port), dest_body_parameter_node
# save additional edges information for graph on each level, the first one is the deepest
additional_params = [] # (src_node, src_out_port) -> parameter_node (for manually added Parameters)
# Go through all nodes in the original model order because data nodes are defined on-the-fly and order matters
for pb_node in body_graph_proto.node:
# create an NX node
id = body_graph.unique_id(node_id(pb_node))
body_graph.add_node(id, pb=pb_node, kind='op')
if hasattr(body_graph, 'op_names_statistic') and hasattr(pb_node, 'op_type'):
body_graph.op_names_statistic[pb_node.op_type] += 1
# add incoming edges based on data_nodes_map
for dst_port, inp in enumerate(pb_node.input):
# should add edge src_internal_id --> dst_id
if inp not in data_nodes_map:
if inp == '':
# input is omitted; most likely it corresponds to an optional input for an operator
continue
else:
is_finished = create_cross_body_edge(body_graph, external_edges, additional_params,
inp, id, dst_port)
if not is_finished:
raise Error(
'Reference to "{}" is not satisfied. A node refer not existing data tensor. ONNX '
'model is not consistent. Protobuf fragment: {}', inp, pb_node)
else:
src_id, src_port = data_nodes_map[inp]
create_edge_with_attrs(body_graph, inp, src_id, src_port, id, dst_port)
# add outgoing edges to data_nodes_map
for src_port, out in enumerate(pb_node.output):
if out in data_nodes_map:
log.debug("Detected reuse of blob {}.".format(out))
data_nodes_map[out] = (id, src_port)
body_results = []
for output in body_graph_proto.output:
tensor_name = str(output.name)
node_name, output_port = data_nodes_map[tensor_name]
assert body_graph.has_node(node_name), 'The body graph does not contain output with name "{}"'.format(
node_name)
body_results.append(Node(body_graph, add_opoutput(body_graph, node_name, output_port, False)))
# add 'internal_layer_id' attribute which is a must have attribute for the loop body node
for idx, body_node in enumerate(body_graph.get_op_nodes()):
body_node['internal_layer_id'] = idx
loop_carried_dependencies_count = len(body_graph_proto.input) - 2
scan_outputs_count = len(body_graph_proto.output) - 1 - loop_carried_dependencies_count
# Loop inputs:
# 0 - trip count
# 1 - execution condition
# 2 .. - loop carried dependencies
# Loop outputs:
# 0 .. loop_carried_dependencies_count - 1 - loop carried dependencies
# loop_carried_dependencies_count .. - scan outputs
# Body inputs:
# 0 - iteration number
# 1 - execution condition
# 2 .. - loop carried dependencies
# Body outputs:
# 0 - execution condition
# 1 .. loop_carried_dependencies_count - loop carried dependencies
# loop_carried_dependencies_count + 1 .. - scan outputs
# some of the inputs/outputs may not be connected but the normalization transformation will take care of it
# connection Loop body nodes with external input edges
next_loop_input_port_idx = sorted(loop_node.in_edges().keys())[-1] + 1
cur_graph = body_graph
for external_edges_subg in external_edges:
if 'parent_node' not in cur_graph.graph:
continue
cur_loop_node = cur_graph.graph['parent_node']
parent_graph = cur_loop_node.graph
for (src_node, src_port), body_node, tensor_name in external_edges_subg:
create_edge_with_attrs(parent_graph, tensor_name, src_node, src_port,
cur_loop_node.id, next_loop_input_port_idx)
Loop.connect_body_input(cur_loop_node, next_loop_input_port_idx, body_node)
next_loop_input_port_idx += 1
cur_graph = parent_graph
# mark current iteration input Parameter node
Loop.mark_current_iteration_parameter_node(loop_node, body_parameters[0])
# connect initial value for "execution condition" input of the loop
Loop.connect_body_input(loop_node, 1, body_parameters[1])
# add back edge with "execution condition"
Loop.add_back_edge(loop_node, body_parameters[1], body_results[0])
# mark "execution condition" Result node
Loop.mark_execution_condition_result_node(loop_node, body_results[0])
# connect initial value for "loop carried" dependencies variables
for idx in range(loop_carried_dependencies_count):
Loop.connect_body_input(loop_node, idx + 2, body_parameters[idx + 2])
# add back edge for "loop carried" dependencies variables
for idx in range(loop_carried_dependencies_count):
Loop.add_back_edge(loop_node, body_parameters[idx + 2], body_results[idx + 1])
# connect final value for "loop carried" dependencies variables
for idx in range(loop_carried_dependencies_count):
Loop.connect_body_output(loop_node, idx, body_results[idx + 1])
# connect "scan outputs" and mark axis for concatenation
for idx in range(loop_carried_dependencies_count, loop_carried_dependencies_count + scan_outputs_count):
Loop.connect_body_output(loop_node, idx, body_results[idx + 1], axis=0)
# run function to parse body nodes attributes similar to the main graph
extract_node_attrs(body_graph, lambda node: onnx_op_extractor(node, check_for_duplicates(onnx_op_extractors)))
return cls.enabled
def driver(argv: argparse.Namespace,
model_file_name: str,
output_model_name: str,
outputs: list,
output_dir: str,
scale: float,
user_shapes: [None, list, np.array] = None,
mean_scale_values: [dict, list] = ()):
meta_info = get_meta_info(argv)
model_proto = load_onnx_model(model_file_name)
model_graph = model_proto.graph # pylint: disable=no-member
#print(model_graph)
#assert len(model_graph) == 1, "An ONNX model contains more than 1 graph: unsupported"
log.debug("Number of nodes in graph_def: {}".format(len(model_graph.node)))
log.debug(
"Number of all input ports (not true inputs) in graph_def: {}".format(
len(model_graph.input)))
log.debug("Number of initializers in graph_def: {}".format(
len(model_graph.initializer)))
log.debug("Number of real inputs in graph_def: {}".format(
len(model_graph.input) - len(model_graph.initializer)))
update_extractors_with_extensions(onnx_op_extractors)
try:
graph = protobuf2nx(model_proto)
log.debug("Number of nodes in NX graph: {}".format(
graph.number_of_nodes()))
graph.__setattr__(
'name',
output_model_name if output_model_name else model_proto.graph.name) # pylint: disable=no-member
graph.graph['layout'] = 'NCHW'
graph.graph['cmd_params'] = argv
graph.graph['fw'] = 'onnx'
graph.graph[
'feature_dim'] = 1 if graph.graph['layout'] == 'NCHW' else 3
graph.graph['ir_version'] = 2 if argv.generate_deprecated_IR_V2 else 4
# 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_onnx_fields(node)))
except Exception as e:
raise Error(
'Cannot pre-process ONNX 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, _ = user_data_repack(
graph, user_shapes, outputs, None)
output_op_nodes = add_output_ops(graph, packed_outputs)
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(graph)
check_empty_graph(graph, 'add_output_ops and add_input_ops')
extract_node_attrs(
graph, lambda node: onnx_op_extractor(
node, check_for_duplicates(onnx_op_extractors)))
class_registration.apply_replacements(
graph, class_registration.ClassType.FRONT_REPLACER)
create_tensor_nodes(graph)
graph_clean_up(graph)
override_placeholder_shapes(graph, packed_user_shapes)
override_batch(graph, argv.batch)
graph_clean_up(graph)
remove_op_nodes(graph, {'op': 'Identity'})
graph_clean_up(graph)
remove_output_ops(graph)
partial_infer(graph)
graph_clean_up(graph)
check_empty_graph(graph, 'partial_infer')
input_op_nodes = add_input_ops(graph, packed_user_shapes, False)
graph_clean_up(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)
graph_clean_up(graph)
graph_clean_up(graph)
remove_op_nodes(graph, {'op': 'Identity'})
remove_useless_split(graph)
class_registration.apply_replacements(
graph, class_registration.ClassType.MIDDLE_REPLACER)
convert_gemm_to_fully_connected(graph)
NormalizeFullyConnected().find_and_replace_pattern(graph)
fuse_pad(graph)
graph_clean_up(graph)
# Mark nodes with attr 'can_be_fused': False to disable fusing for specified nodes
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(graph)
if not argv.disable_fusing:
# Converting ScaleShift layer to Mul->Add
convert_scale_shift_to_mul_add(graph)
graph_clean_up(graph)
# Fusing the sequences of Mul/Add operations
fuse_mul_add_sequence(graph)
graph_clean_up(graph)
# Fusing linear operation to Convolution
fuse_linear_ops(graph)
graph_clean_up(graph)
if not argv.disable_gfusing:
grouped_convolutions_fusing(graph)
graph_clean_up(graph)
if not argv.disable_fusing:
fuse_linear_ops(graph)
graph_clean_up(graph)
convert_muladd_to_scaleshift_or_power(graph)
graph_clean_up(graph)
convert_mul_add_to_power(graph)
graph_clean_up(graph)
convert_reshape(graph)
convert_add_to_scaleshift(graph) # scale = 1
convert_mul_to_scaleshift(graph) # biases = 0
fuse_pad(graph)
graph_clean_up(graph)
if argv.reverse_input_channels:
reverse_input_channels(graph)
if argv.move_to_preprocess:
move_scaleshift_to_preprocess(graph)
graph_clean_up(graph)
fuse_sequence_of_reshapes(graph)
graph_clean_up(graph)
pattern = EltwiseInputNormalize()
pattern.find_and_replace_pattern(graph)
merge_nodes_permutations(graph)
permute_data_nodes_attrs(graph)
permute_op_nodes_attrs(graph)
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
