def test_2(self):
graph = build_graph(nodes,
edges, {
'mul': {
'can_be_fused': True
},
'mul_const_data': {
'shape': np.array([1]),
'value': np.array([-1])
},
'quantize_data': {
'shape': np.array([2, 3, 4, 4])
},
'mi_o_data': {
'shape': np.array([1]),
'value': np.array([0])
},
'ma_o_data': {
'shape': np.array([1]),
'value': np.array([1])
},
},
nodes_with_edges_only=True)
graph.stage = 'middle'
graph_ref = build_graph(nodes,
edges_ref, {
'quantize_data': {
'shape': np.array([2, 3, 4, 4])
},
'mul_const_data': {
'shape': np.array([1]),
'value': np.array([-1])
},
'mi_o_data': {
'shape': np.array([1]),
'value': np.array([1])
},
'ma_o_data': {
'shape': np.array([1]),
'value': np.array([0])
},
'mi_i_data': {
'shape': np.array([1]),
'value': np.array([10])
},
'ma_i_data': {
'shape': np.array([1]),
'value': np.array([-10])
},
},
nodes_with_edges_only=True)
MulFakeQuantizeFuse().find_and_replace_pattern(graph)
(flag, resp) = compare_graphs(graph,
graph_ref,
'output',
check_op_attrs=True)
self.assertTrue(flag, resp)
def test_negative_5(self):
graph = build_graph(
nodes,
edges, {
'mul': {
'can_be_fused': True
},
'mul_const_data': {
'shape': np.array([3, 1, 1]),
'value': np.array([[[0]], [[1]], [[2]]])
},
'quantize_data': {
'shape': np.array([2, 3, 4, 4])
},
'mi_o_data': {
'shape': np.array([1, 1, 1, 1]),
'value': np.broadcast_to(np.array([0]), (1, 1, 1, 1))
},
'ma_o_data': {
'shape': np.array([1, 1, 1, 1]),
'value': np.broadcast_to(np.array([1]), (1, 1, 1, 1))
},
},
nodes_with_edges_only=True)
graph.stage = 'middle'
graph_ref = graph.copy()
MulFakeQuantizeFuse().find_and_replace_pattern(graph)
(flag, resp) = compare_graphs(graph,
graph_ref,
'output',
check_op_attrs=True)
self.assertTrue(flag, resp)
def negative_test_1(self):
graph = build_graph(nodes, edges, nodes_with_edges_only=True)
graph.stage = 'middle'
graph_ref = build_graph(nodes, edges, nodes_with_edges_only=True)
MulFakeQuantizeFuse().find_and_replace_pattern(graph)
(flag, resp) = compare_graphs(graph,
graph_ref,
'output',
check_op_attrs=True)
self.assertTrue(flag, resp)
def find_and_replace_pattern(self, graph: Graph):
fw = graph.graph['fw']
argv = graph.graph['cmd_params']
layout = graph.graph['layout']
for_graph_and_each_sub_graph_recursively(graph, fuse_pad)
for_graph_and_each_sub_graph_recursively(graph, lambda G: G.clean_up())
# 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 batchNormInference with 4 inputs, so we have to split it to two ScaleShift
for_graph_and_each_sub_graph_recursively(graph, convert_batch_norm)
if fw == 'caffe':
# 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, Div().find_and_replace_pattern)
for_graph_and_each_sub_graph_recursively(graph, Sub().find_and_replace_pattern)
for_graph_and_each_sub_graph_recursively(graph, lambda G: G.clean_up())
if not argv.disable_fusing:
if fw != 'caffe':
# 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, lambda G: G.clean_up())
# 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, lambda G: G.clean_up())
normalize_eltwise_inputs(graph)
for_graph_and_each_sub_graph_recursively(graph, lambda G: G.clean_up())
# Fusing linear operation to Convolution
for_graph_and_each_sub_graph_recursively(graph, fuse_linear_ops)
for_graph_and_each_sub_graph_recursively(graph, lambda G: G.clean_up())
if not argv.disable_gfusing:
for_graph_and_each_sub_graph_recursively(graph, grouped_convolutions_fusing)
for_graph_and_each_sub_graph_recursively(graph, lambda G: G.clean_up())
if not argv.disable_fusing:
for_graph_and_each_sub_graph_recursively(graph, fuse_linear_ops)
for_graph_and_each_sub_graph_recursively(graph, lambda G: G.clean_up())
for_graph_and_each_sub_graph_recursively(graph, normalize_eltwise_inputs)
for_graph_and_each_sub_graph_recursively(graph, lambda G: G.clean_up())
MarkNodesToFuseUpToFakeQuantize().find_and_replace_pattern(graph)
FakeQuantizeFuse().find_and_replace_pattern(graph)
AddFakeQuantizeFuse().find_and_replace_pattern(graph)
MulFakeQuantizeFuse().find_and_replace_pattern(graph)
for_graph_and_each_sub_graph_recursively(graph, lambda G: G.clean_up())
for_graph_and_each_sub_graph_recursively(graph, fuse_pad)
for_graph_and_each_sub_graph_recursively(graph, lambda G: G.clean_up())
if layout != 'NHWC' and not argv.disable_resnet_optimization:
stride_optimization(graph)
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 driver(argv: argparse.Namespace, model_file_name: str,
output_model_name: str, output_dir: str):
log_step(argv.steps, 'LOAD')
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
if graph.graph['cmd_params'].generate_experimental_IR_V10:
version = 10
else:
version = 6
graph.graph[
'ir_version'] = 2 if argv.generate_deprecated_IR_V2 else version
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 ------------------------------------------------------
log_step(argv.steps, 'FRONT')
class_registration.apply_replacements(
graph, class_registration.ClassType.FRONT_REPLACER)
log_step(argv.steps, 'MIDDLE')
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, 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)
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)
if argv.reverse_input_channels:
reverse_input_channels(graph)
if argv.move_to_preprocess:
move_scaleshift_to_preprocess(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)
log_step(argv.steps, 'BACK')
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)
log_step(argv.steps, 'EMIT')
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