def apply_offline_transformations(input_model: str, argv: argparse.Namespace):
# This variable is only needed by GenerateMappingFile transformation
# to produce correct mapping
extract_names = argv.framework in ['tf', 'mxnet', 'kaldi']
from openvino.offline_transformations import generate_mapping_file, serialize # pylint: disable=import-error,no-name-in-module
from openvino.frontend import FrontEndManager # pylint: disable=no-name-in-module,import-error
from openvino.tools.mo.back.preprocessing import apply_preprocessing # pylint: disable=no-name-in-module,import-error
fem = FrontEndManager()
# We have to separate fe object lifetime from fem to
# avoid segfault during object destruction. So fe must
# be destructed before fem object explicitly.
def read_model(path_to_xml):
fe = fem.load_by_framework(framework="ir")
function = fe.convert(fe.load(path_to_xml))
return function
func = read_model(input_model + "_tmp.xml")
# TODO: use ngraph preprocessing (Mean/Scale/ReverseInputChannels) for legacy frontends
reverse_input_channels = False
if 'reverse_input_channels' in argv:
reverse_input_channels = argv.reverse_input_channels
argv.reverse_input_channels = False
mean_scale_values = {}
if 'mean_scale_values' in argv:
mean_scale_values = argv.mean_scale_values
argv.mean_scale_values = {}
scale = None
if 'scale' in argv:
scale = argv.scale
argv.scale = None
# Apply preprocessing for layouts only
apply_preprocessing(ov_function=func, argv=argv)
if 'reverse_input_channels' in argv:
argv.reverse_input_channels = reverse_input_channels
if 'mean_scale_values' in argv:
argv.mean_scale_values = mean_scale_values
if 'scale' in argv:
argv.scale = scale
apply_moc_transformations(func)
params_with_custom_types = create_params_with_custom_types(
argv.packed_user_shapes)
apply_moc_legacy_transformations(func, params_with_custom_types)
apply_user_transformations(func, parse_transform(argv.transform))
if "compress_fp16" in argv and argv.compress_fp16:
compress_model(func)
serialize(func,
str(input_model + ".xml").encode('utf-8'),
(input_model + ".bin").encode('utf-8'))
path_to_mapping = input_model + ".mapping"
generate_mapping_file(func, path_to_mapping.encode('utf-8'), extract_names)
def moc_emit_ir(ngraph_function: Model, argv: argparse.Namespace):
output_dir = argv.output_dir if argv.output_dir != '.' else os.getcwd()
# Apply preprocessing (mean/scale/reverse_channels/convert_layout/etc)
apply_preprocessing(ov_function=ngraph_function, argv=argv)
# Apply transformations
from openvino.tools.mo.back.offline_transformations import apply_user_transformations, apply_moc_transformations
apply_user_transformations(ngraph_function, parse_transform(argv.transform))
apply_moc_transformations(ngraph_function)
if argv.compress_fp16:
from openvino.tools.mo.back.offline_transformations import compress_model
compress_model(ngraph_function)
orig_model_name = os.path.normpath(os.path.join(output_dir, argv.model_name))
from openvino.offline_transformations_pybind import serialize # pylint: disable=import-error,no-name-in-module
serialize(ngraph_function, (orig_model_name + ".xml").encode('utf-8'), (orig_model_name + ".bin").encode('utf-8'))
del argv.feManager
# add meta information to IR
append_ir_info(file=orig_model_name,
meta_info=get_meta_info(argv),
mean_data=None,
input_names=None)
print('[ SUCCESS ] Generated IR version {} model.'.format(get_ir_version(argv)))
print('[ SUCCESS ] XML file: {}.xml'.format(orig_model_name))
print('[ SUCCESS ] BIN file: {}.bin'.format(orig_model_name))
return 0
def moc_emit_ir(ngraph_function: Model, argv: argparse.Namespace):
output_dir = argv.output_dir if argv.output_dir != '.' else os.getcwd()
# Apply preprocessing (mean/scale/reverse_channels/convert_layout/etc)
apply_preprocessing(ov_function=ngraph_function, argv=argv)
# Apply transformations
from openvino.tools.mo.back.offline_transformations import apply_user_transformations, apply_moc_transformations, \
apply_moc_legacy_transformations
apply_moc_transformations(ngraph_function)
from openvino.offline_transformations import compress_quantize_weights_transformation
compress_quantize_weights_transformation(ngraph_function)
if argv.framework == "onnx":
# set OldApi map in IR to be executed via OV API 1.x and for parity with legacy MO
params_with_custom_types = [] if argv.placeholder_data_types is None \
else list(argv.placeholder_data_types.keys())
apply_moc_legacy_transformations(ngraph_function,
params_with_custom_types)
apply_user_transformations(ngraph_function,
parse_transform(argv.transform))
if argv.compress_fp16:
from openvino.tools.mo.back.offline_transformations import compress_model
compress_model(ngraph_function)
orig_model_name = os.path.normpath(
os.path.join(output_dir, argv.model_name))
from openvino.runtime import serialize # pylint: disable=import-error,no-name-in-module
from openvino.offline_transformations import generate_mapping_file # pylint: disable=import-error,no-name-in-module
serialize(ngraph_function, (orig_model_name + ".xml").encode('utf-8'),
(orig_model_name + ".bin").encode('utf-8'))
del argv.feManager
path_to_mapping = orig_model_name + ".mapping"
extract_names = argv.framework in ['tf', 'mxnet', 'kaldi']
generate_mapping_file(ngraph_function, path_to_mapping.encode('utf-8'),
extract_names)
# add meta information to IR
append_ir_info(file=orig_model_name,
meta_info=get_meta_info(argv),
mean_data=None,
input_names=None,
legacy_path=False)
print('[ SUCCESS ] Generated IR version {} model.'.format(
get_ir_version(argv)))
print('[ SUCCESS ] XML file: {}.xml'.format(orig_model_name))
print('[ SUCCESS ] BIN file: {}.bin'.format(orig_model_name))
return 0
def process_function(ov_function: Model, argv: Namespace):
apply_preprocessing(ov_function=ov_function, argv=argv)
