def test_get_IENetwork_from_nGraph():
func = create_relu([1, 3, 22, 22])
caps = Function.to_capsule(func)
cnnNetwork = IENetwork(caps)
assert cnnNetwork != None
assert ng.function_from_cnn(cnnNetwork) != None
func2 = ng.function_from_cnn(cnnNetwork)
assert func2 != None
def test_get_IENetwork_from_nGraph():
element_type = Type.f32
param = Parameter(element_type, Shape([1, 3, 22, 22]))
relu = ng.relu(param)
func = Function([relu], [param], 'test')
caps = Function.to_capsule(func)
cnnNetwork = IENetwork(caps)
assert cnnNetwork != None
assert ng.function_from_cnn(cnnNetwork) != None
func2 = ng.function_from_cnn(cnnNetwork)
assert func2 != None
def test_create_two_exec_net():
function = create_relu([
ng.Dimension(0, 5),
ng.Dimension(4),
ng.Dimension(20),
ng.Dimension(20)
])
net = ng.function_to_cnn(function)
ie_core = IECore()
ie_core.register_plugin("ov_template_plugin", "TEMPLATE")
exec_net1 = ie_core.load_network(net, "TEMPLATE", num_requests=2)
assert ng.function_from_cnn(net) != None
exec_net2 = ie_core.load_network(net, "TEMPLATE", num_requests=2)
assert ng.function_from_cnn(net) != None
def test_serialize():
ie = IECore()
net = ie.read_network(model=test_net_xml, weights=test_net_bin)
net.serialize("./serialized_net.xml", "./serialized_net.bin")
serialized_net = ie.read_network(model="./serialized_net.xml",
weights="./serialized_net.bin")
func_net = ng.function_from_cnn(net)
ops_net = func_net.get_ordered_ops()
ops_net_names = [op.friendly_name for op in ops_net]
func_serialized_net = ng.function_from_cnn(serialized_net)
ops_serialized_net = func_serialized_net.get_ordered_ops()
ops_serialized_net_names = [op.friendly_name for op in ops_serialized_net]
assert ops_serialized_net_names == ops_net_names
os.remove("./serialized_net.xml")
os.remove("./serialized_net.bin")
def test_getting_shapes():
ie = IECore()
net = ie.read_network(model=test_net_xml, weights=test_net_bin)
func = ng.function_from_cnn(net)
ops = func.get_ordered_ops()
shapes = [sh for sh in ops[2].shape]
assert shapes == [1, 16, 32, 32]
def test_dlsdk_launcher_valid_affinity_map(self, mocker, models_dir):
affinity_map = {'conv1': 'GPU'}
if not has_layers():
affinity_map.update({'conv1/Dims294/copy_const': 'GPU'})
mocker.patch(
'openvino.tools.accuracy_checker.launcher.dlsdk_launcher.read_yaml',
return_value=affinity_map)
dlsdk_test_model = get_dlsdk_test_model(
models_dir, {
'device': 'HETERO:CPU,GPU',
'affinity_map': './affinity_map.yml'
})
if has_layers():
layers = dlsdk_test_model.network.layers
for key, value in affinity_map.items():
assert layers[key].affinity == value
else:
ng_function = ng.function_from_cnn(dlsdk_test_model.network)
for node in ng_function.get_ordered_ops():
if node.get_friendly_name() != 'conv1':
continue
assert node.get_friendly_name() in affinity_map
assert node.get_rt_info()['affinity'] == affinity_map[
node.get_friendly_name()]
def test_moc_transformations():
net = get_test_cnnnetwork()
ApplyMOCTransformations(net, False)
f = ng.function_from_cnn(net)
assert f != None
assert len(f.get_ops()) == 3
def test_low_latency_transformations():
net = get_test_cnnnetwork()
ApplyLowLatencyTransformation(net)
f = ng.function_from_cnn(net)
assert f != None
assert len(f.get_ops()) == 3
def test_pruning_transformations():
net = get_test_cnnnetwork()
ApplyPruningTransformation(net)
f = ng.function_from_cnn(net)
assert f != None
assert len(f.get_ops()) == 3
def init_network(self):
"""Function which initializes Intel Inference Engine.
"""
# Load OpenVINO Inference Engine.
self.get_logger().info(f"Loading Inference Engine on {self.device}")
self.ie = IECore()
# Read and load the network.
self.net = self.ie.read_network(model=constants.MODEL_XML,
weights=constants.MODEL_BIN)
self.func = ng.function_from_cnn(self.net)
self.ops = self.func.get_ordered_ops()
self.exec_net = self.ie.load_network(network=self.net,
device_name=self.device)
# Read expected input image info from network and prepare input blobs.
# n: batch size, c: no. of channels, h: input height, w: input width
for self.input_key in self.net.input_info:
self.input_name = self.input_key
self.n, self.c, self.h, self.w = self.net.input_info[
self.input_key].input_data.shape
# Initializing to float for optimizing in later functions
self.h = float(self.h)
self.w = float(self.w)
# Prepare output blobs
self.out_blob = next(iter(self.net.outputs))
def __init__(self, model, confidence_threshold=0.5, iou_threshold=0.4):
super(YoloDetector, self).__init__(model)
# パラメータチェック
assert 0.0 <= confidence_threshold and confidence_threshold <= 1.0, \
"Confidence threshold is expected to be in range [0; 1]"
assert 0.0 <= iou_threshold and iou_threshold <= 1.0, \
"Intersection over union threshold is expected to be in range [0; 1]"
self.confidence_threshold = confidence_threshold
self.iou_threshold = iou_threshold
if hasattr(model, "input_info") :
# 2021.1 以降用
# このプログラムは1入力のモデルのみサポートしているので、チェック
assert len(model.input_info) == 1, "Expected 1 input blob"
self.input_blob = next(iter(model.input_info))
self.input_shape = model.input_info[self.input_blob].input_data.shape
else :
# 2020.4 以前用
# このプログラムは1入力のモデルのみサポートしているので、チェック
assert len(model.inputs) == 1, "Expected 1 input blob"
self.input_blob = next(iter(model.inputs))
self.input_shape = model.inputs[self.input_blob].shape
self.layer_params = {}
'''
次の行で2021.1で以下のwarningが出る対策
DeprecationWarning: 'layers' property of IENetwork class is deprecated.
For iteration over network please use get_ops()/get_ordered_ops() methods from nGraph Python API
'''
if ng :
# ng が定義されてたら → 2021.1以降
function = ng.function_from_cnn(model)
nodes = { n.friendly_name : n for n in function.get_ops() }
self.layer_params_aaa = {}
for layer_name in model.outputs.keys() :
# print(layer_name)
output_shapes = list(nodes[layer_name].inputs()[0].get_source_output().get_node().shape)
output_params = nodes[layer_name]._get_attributes()
assert output_shapes[2] == output_shapes[3], \
f"Invalid size of output blob. It sould be in NCHW layout and height should be equal to width. " \
"Current height = {output_shapes[2]}, " \
"current width = {output_shapes[3]}"
self.layer_params[layer_name] = self.YoloParams(output_params, output_shapes[2])
else :
layers = model.layers
for layer_name in model.outputs.keys() :
# print(layer_name)
output_shapes = layers[layers[layer_name].parents[0]].out_data[0].shape
output_params = layers[layer_name].params
assert output_shapes[2] == output_shapes[3], \
f"Invalid size of output blob. It sould be in NCHW layout and height should be equal to width. " \
"Current height = {output_shapes[2]}, " \
"current width = {output_shapes[3]}"
self.layer_params[layer_name] = self.YoloParams(output_params, output_shapes[2])
def test_create_IENetwork_from_nGraph():
func = create_ngraph_function([1, 3, 22, 22])
caps = Function.to_capsule(func)
cnnNetwork = IENetwork(caps)
assert cnnNetwork != None
func2 = ng.function_from_cnn(cnnNetwork)
assert func2 != None
assert len(func2.get_ops()) == 3
def test_make_stateful_transformations():
net = get_test_cnnnetwork()
ApplyMakeStatefulTransformation(net, {"parameter": "result"})
f = ng.function_from_cnn(net)
assert f != None
assert len(f.get_parameters()) == 0
assert len(f.get_results()) == 0
def test_get_ops_from_IENetwork():
ie = IECore()
net = ie.read_network(model=test_net_xml, weights=test_net_bin)
func = ng.function_from_cnn(net)
ops = func.get_ordered_ops()
ops_names = [op.friendly_name for op in ops]
assert len(ops_names) != 0
assert 'data' in ops_names
def test_get_set_rt_info():
ie = IECore()
net = ie.read_network(model=test_net_xml, weights=test_net_bin)
func = ng.function_from_cnn(net)
ops = func.get_ordered_ops()
rt_info = ops[14].get_rt_info()
rt_info["affinity"] = "test_affinity"
assert ops[14].get_rt_info()["affinity"] == "test_affinity"
def _get_meta_from_ngraph(self, layers_info):
ng_func = ngraph.function_from_cnn(self.net)
for node in ng_func.get_ordered_ops():
layer_name = node.get_friendly_name()
if layer_name not in layers_info.keys():
continue
layers_info[layer_name].meta = node._get_attributes()
layers_info[layer_name].type = node.get_type_name()
return layers_info
def import_onnx_model(model: onnx.ModelProto) -> Function:
onnx.checker.check_model(model)
model_byte_string = model.SerializeToString()
ie = IECore()
ie_network = ie.read_network(model=model_byte_string, weights=b"", init_from_buffer=True)
ng_function = ng.function_from_cnn(ie_network)
return ng_function
def test_CreateIENetworkFromNGraph():
element_type = Type.f32
param = Parameter(element_type, Shape([1, 3, 22, 22]))
relu = ng.relu(param)
func = Function([relu], [param], 'test')
caps = Function.to_capsule(func)
cnnNetwork = IENetwork(caps)
assert cnnNetwork != None
assert ng.function_from_cnn(cnnNetwork) != None
assert len(cnnNetwork.layers) == 2
def test_serialize(device):
ie = IECore()
if device == "CPU":
if ie.get_metric(device, "FULL_DEVICE_NAME") == "arm_compute::NEON":
pytest.skip("Can't run on ARM plugin due-to ngraph")
import ngraph as ng
net = ie.read_network(model=test_net_xml, weights=test_net_bin)
net.serialize("./serialized_net.xml", "./serialized_net.bin")
serialized_net = ie.read_network(model="./serialized_net.xml",
weights="./serialized_net.bin")
func_net = ng.function_from_cnn(net)
ops_net = func_net.get_ordered_ops()
ops_net_names = [op.friendly_name for op in ops_net]
func_serialized_net = ng.function_from_cnn(serialized_net)
ops_serialized_net = func_serialized_net.get_ordered_ops()
ops_serialized_net_names = [op.friendly_name for op in ops_serialized_net]
assert ops_serialized_net_names == ops_net_names
os.remove("./serialized_net.xml")
os.remove("./serialized_net.bin")
def test_query_network(device):
ie = IECore()
net = ie.read_network(model=test_net_xml, weights=test_net_bin)
query_res = ie.query_network(net, device)
func_net = ng.function_from_cnn(net)
ops_net = func_net.get_ordered_ops()
ops_net_names = [op.friendly_name for op in ops_net]
assert [key for key in query_res.keys() if key not in ops_net_names] == [], \
"Not all network layers present in query_network results"
assert next(iter(set(query_res.values()))) == device, "Wrong device for some layers"
def load(self):
with open(os.path.join(self.model_path, 'config.json')) as f:
data = json.load(f)
try:
self.validate_json_configuration(data)
self.set_model_configuration(data)
except ApplicationError as e:
raise e
with open(os.path.join(self.model_path, 'classes.txt'), 'r') as f:
self.classes = [line.strip() for line in f.readlines()]
self.ie = IECore()
# Load the model
models_path = Path(self.model_path)
model_xml = next(models_path.glob("*.xml"))
model_bin = model_xml.parent / (model_xml.stem + ".bin")
self.net = self.ie.read_network(model=str(model_xml),
weights=str(model_bin))
func = ng.function_from_cnn(self.net)
ops = func.get_ordered_ops()
assert (len(self.net.input_info.keys()) in {1, 2}), \
"Network must have one or two inputs"
self.input_name, self.input_info_name = None, None
for k, v in self.net.input_info.items():
if len(v.layout) == 4:
self.input_name = k
v.precision = "U8"
elif len(v.layout) == 2:
self.input_info_name = k
v.precision = "FP32"
assert (v.input_data.shape[1] in {3, 6}
and v.input_data.shape[0]
== 1), "Input info should be 3 or 6 values length"
self.output_name, output_info = "", self.net.outputs[next(
iter(self.net.outputs.keys()))]
output_ops = {op.friendly_name : op for op in ops \
if op.friendly_name in self.net.outputs and op.get_type_name() == "DetectionOutput"}
if len(output_ops) != 0:
self.output_name, output_info = output_ops.popitem()
if self.output_name == "":
print("Can't find a DetectionOutput layer in the topology")
sys.exit(-1)
output_info.precision = "FP32"
self.exec_net = self.ie.load_network(network=self.net,
device_name="CPU")
def get_net_copy_with_output(model: str, output: str, core: IECore):
net_copy = get_net(model=model, core=core)
func = ng.function_from_cnn(net_copy)
if output not in ['None', None]:
# output with port_id in name is absent in ops list
founded_op = [op for op in func.get_ops() if op.friendly_name == output]
if founded_op:
net_copy.add_outputs(output)
else:
split = output.rsplit(".", 1)
net_copy.add_outputs((split[0], int(split[1])))
return net_copy
def _get_output_info(self):
def get_parent(node):
return node.inputs()[0].get_source_output().get_node()
ng_func = ngraph.function_from_cnn(self.net)
output_info = {}
for node in ng_func.get_ordered_ops():
layer_name = node.get_friendly_name()
if layer_name not in self.net.outputs:
continue
shape = list(get_parent(node).shape)
yolo_params = self.Params(node._get_attributes(), shape[2:4])
output_info[layer_name] = (shape, yolo_params)
return output_info
def get_objects(output, net, new_frame_height_width, source_height_width, prob_threshold, is_proportional):
objects = list()
function = ng.function_from_cnn(net)
for layer_name, out_blob in output.items():
#out_blob = out_blob.buffer.reshape(net.layers[net.layers[layer_name].parents[0]].out_data[0].shape)
#layer_params = YoloParams(net.layers[layer_name].params, out_blob.shape[2])
out_blob = out_blob.buffer.reshape(net.outputs[layer_name].shape)
params = [x._get_attributes() for x in function.get_ordered_ops() if x.get_friendly_name() == layer_name][0]
layer_params = YoloParams(params, out_blob.shape[2])
objects += parse_yolo_region(out_blob, new_frame_height_width, source_height_width, layer_params,
prob_threshold, is_proportional)
return objects
def test_get_ops_from_IENetwork():
ie = IECore()
net = ie.read_network(model=test_net_xml, weights=test_net_bin)
func = ng.function_from_cnn(net)
ops = func.get_ordered_ops()
ops_names = [op.friendly_name for op in ops]
assert ops_names == ['data', '20/mean/Fused_Mul_614616_const', '19/WithoutBiases', 'data_add_575/copy_const',
'19/Fused_Add_', '21', '22', 'onnx_initializer_node_8/Output_0/Data__const',
'23/WithoutBiases', '23/Dims357/copy_const', '23', '25/mean/Fused_Mul_618620_const',
'24/WithoutBiases', 'data_add_578583/copy_const', '24/Fused_Add_', '26', '27',
'28/Reshape/Cast_1955_const', '28/Reshape', 'onnx_initializer_node_17/Output_0/Data__const',
'29/WithoutBiases', 'onnx_initializer_node_18/Output_0/Data_/copy_const', '29', 'fc_out',
'fc_out/sink_port_0']
def __init__(self, *args, upsample_ratio=1, **kwargs):
super().__init__(*args, **kwargs)
self.pooled_heatmaps_blob_name = 'pooled_heatmaps'
self.heatmaps_blob_name = 'heatmaps'
self.pafs_blob_name = 'pafs'
function = ng.function_from_cnn(self.net)
paf = function.get_output_op(0)
paf = paf.inputs()[0].get_source_output().get_node()
paf.set_friendly_name(self.pafs_blob_name)
heatmap = function.get_output_op(1)
heatmap = heatmap.inputs()[0].get_source_output().get_node()
heatmap.set_friendly_name(self.heatmaps_blob_name)
# Add keypoints NMS to the network.
# Heuristic NMS kernel size adjustment depending on the feature maps upsampling ratio.
p = int(np.round(6 / 7 * upsample_ratio))
k = 2 * p + 1
pooled_heatmap = ng.max_pool(heatmap,
kernel_shape=(k, k),
pads_begin=(p, p),
pads_end=(p, p),
strides=(1, 1),
name=self.pooled_heatmaps_blob_name)
f = ng.impl.Function([
ng.result(heatmap, name=self.heatmaps_blob_name),
ng.result(pooled_heatmap, name=self.pooled_heatmaps_blob_name),
ng.result(paf, name=self.pafs_blob_name)
], function.get_parameters(), 'hpe')
self.image_blob_name = self._get_inputs(self.net)
self.net = IENetwork(ng.impl.Function.to_capsule(f))
self.exec_net = self.ie.load_network(
network=self.net,
device_name=self.device,
num_requests=self.max_num_requests)
self.requests = self.exec_net.requests
self.empty_requests = deque(self.requests)
self.num_joints = self.net.outputs[self.heatmaps_blob_name].shape[
1] - 1 # The last channel is for background.
target_size = self.net.input_info[
self.image_blob_name].input_data.shape[-2]
self.output_scale = target_size / self.net.outputs[
self.heatmaps_blob_name].shape[-2]
if self.target_size is None:
self.target_size = target_size
self.decoder = OpenPoseDecoder(num_joints=self.num_joints)
def test_query_network(device):
ie = IECore()
if device == "CPU":
if ie.get_metric(device, "FULL_DEVICE_NAME") == "arm_compute::NEON":
pytest.skip("Can't run on ARM plugin due-to ngraph")
import ngraph as ng
net = ie.read_network(model=test_net_xml, weights=test_net_bin)
query_res = ie.query_network(net, device)
func_net = ng.function_from_cnn(net)
ops_net = func_net.get_ordered_ops()
ops_net_names = [op.friendly_name for op in ops_net]
assert [key for key in query_res.keys() if key not in ops_net_names] == [], \
"Not all network layers present in query_network results"
assert next(iter(set(query_res.values()))) == device, "Wrong device for some layers"
def test_offline_api():
element_type = Type.f32
param = Parameter(element_type, Shape([1, 3, 22, 22]))
relu = ng.relu(param)
func = Function([relu], [param], 'test')
caps = Function.to_capsule(func)
cnnNetwork = IENetwork(caps)
assert cnnNetwork != None
ApplyMOCTransformations(cnnNetwork, False)
func2 = ng.function_from_cnn(cnnNetwork)
assert func2 != None
assert len(func2.get_ops()) == 3
def test_import_onnx_with_external_data():
model_path = os.path.join(os.path.dirname(__file__), "models/external_data.onnx")
ie = IECore()
ie_network = ie.read_network(model=model_path)
ng_function = ng.function_from_cnn(ie_network)
dtype = np.float32
value_a = np.array([1.0, 3.0, 5.0], dtype=dtype)
value_b = np.array([3.0, 5.0, 1.0], dtype=dtype)
# third input [5.0, 1.0, 3.0] read from external file
runtime = get_runtime()
computation = runtime.computation(ng_function)
result = computation(value_a, value_b)
assert np.allclose(result, np.array([3.0, 3.0, 3.0], dtype=dtype))
def test_reshape_with_partial_shape(device, shape, p_shape):
function = create_relu(shape)
net = ng.function_to_cnn(function)
net.reshape({"data": p_shape})
changedFunction = ng.function_from_cnn(net)
p_shape = ng.impl.PartialShape(p_shape)
assert changedFunction.get_parameters()[0].get_partial_shape().is_dynamic
assert changedFunction.get_results()[0].get_output_partial_shape(
0).is_dynamic
assert function.get_parameters()[0].get_partial_shape().is_dynamic
assert function.get_results()[0].get_output_partial_shape(0).is_dynamic
assert changedFunction.get_parameters()[0].get_partial_shape() == p_shape
assert changedFunction.get_results()[0].get_output_partial_shape(
0) == p_shape
assert function.get_parameters()[0].get_partial_shape() == p_shape
assert function.get_results()[0].get_output_partial_shape(0) == p_shape