def test_tflite_parser_with_optional_options_out_of_scope(shared_data_folder):
parser = create_with_opt()
network = parser.CreateNetworkFromBinaryFile(
os.path.join(shared_data_folder, "mock_model.tflite"))
graphs_count = parser.GetSubgraphCount()
graph_id = graphs_count - 1
input_names = parser.GetSubgraphInputTensorNames(graph_id)
input_binding_info = parser.GetNetworkInputBindingInfo(
graph_id, input_names[0])
output_names = parser.GetSubgraphOutputTensorNames(graph_id)
preferred_backends = [ann.BackendId('CpuAcc'), ann.BackendId('CpuRef')]
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
opt_network, messages = ann.Optimize(network, preferred_backends,
runtime.GetDeviceSpec(),
ann.OptimizerOptions())
assert 0 == len(messages)
net_id, messages = runtime.LoadNetwork(opt_network)
assert "" == messages
def get_runtime(shared_data_folder, network_file):
parser= ann.ITfLiteParser()
preferred_backends = [ann.BackendId('CpuAcc'), ann.BackendId('CpuRef')]
network = parser.CreateNetworkFromBinaryFile(os.path.join(shared_data_folder, network_file))
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
yield preferred_backends, network, runtime
def test_caffe_parser_end_to_end(shared_data_folder):
parser = ann.ICaffeParser = ann.ICaffeParser()
# Load the network specifying the inputs and outputs
input_name = "Placeholder"
tensor_shape = {input_name: ann.TensorShape((1, 1, 28, 28))}
requested_outputs = ["output"]
network = parser.CreateNetworkFromBinaryFile(
os.path.join(shared_data_folder, 'mock_model.caffemodel'),
tensor_shape, requested_outputs)
# Specify preferred backend
preferred_backends = [ann.BackendId('CpuAcc'), ann.BackendId('CpuRef')]
input_binding_info = parser.GetNetworkInputBindingInfo(input_name)
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
opt_network, messages = ann.Optimize(network, preferred_backends,
runtime.GetDeviceSpec(),
ann.OptimizerOptions())
assert 0 == len(messages)
net_id, messages = runtime.LoadNetwork(opt_network)
assert "" == messages
# Load test image data stored in input_caffe.npy
input_tensor_data = np.load(
os.path.join(shared_data_folder,
'caffe_parser/input_caffe.npy')).astype(np.float32)
input_tensors = ann.make_input_tensors([input_binding_info],
[input_tensor_data])
# Load output binding info and
outputs_binding_info = []
for output_name in requested_outputs:
outputs_binding_info.append(
parser.GetNetworkOutputBindingInfo(output_name))
output_tensors = ann.make_output_tensors(outputs_binding_info)
runtime.EnqueueWorkload(net_id, input_tensors, output_tensors)
output_vectors = ann.workload_tensors_to_ndarray(output_tensors)
# Load golden output file for result comparison.
expected_output = np.load(
os.path.join(shared_data_folder,
'caffe_parser/golden_output_caffe.npy'))
# Check that output matches golden output to 4 decimal places (there are slight rounding differences after this)
np.testing.assert_almost_equal(output_vectors[0], expected_output, 4)
def test_tflite_parser_end_to_end(shared_data_folder):
parser = ann.ITfLiteParser()
network = parser.CreateNetworkFromBinaryFile(
os.path.join(shared_data_folder, "mock_model.tflite"))
graphs_count = parser.GetSubgraphCount()
graph_id = graphs_count - 1
input_names = parser.GetSubgraphInputTensorNames(graph_id)
input_binding_info = parser.GetNetworkInputBindingInfo(
graph_id, input_names[0])
output_names = parser.GetSubgraphOutputTensorNames(graph_id)
preferred_backends = [ann.BackendId('CpuAcc'), ann.BackendId('CpuRef')]
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
opt_network, messages = ann.Optimize(network, preferred_backends,
runtime.GetDeviceSpec(),
ann.OptimizerOptions())
assert 0 == len(messages)
net_id, messages = runtime.LoadNetwork(opt_network)
assert "" == messages
# Load test image data stored in input_lite.npy
input_tensor_data = np.load(
os.path.join(shared_data_folder, 'tflite_parser/input_lite.npy'))
input_tensors = ann.make_input_tensors([input_binding_info],
[input_tensor_data])
output_tensors = []
for index, output_name in enumerate(output_names):
out_bind_info = parser.GetNetworkOutputBindingInfo(
graph_id, output_name)
out_tensor_info = out_bind_info[1]
out_tensor_id = out_bind_info[0]
output_tensors.append((out_tensor_id, ann.Tensor(out_tensor_info)))
runtime.EnqueueWorkload(net_id, input_tensors, output_tensors)
output_vectors = []
for index, out_tensor in enumerate(output_tensors):
output_vectors.append(out_tensor[1].get_memory_area())
# Load golden output file for result comparison.
expected_outputs = np.load(
os.path.join(shared_data_folder,
'tflite_parser/golden_output_lite.npy'))
# Check that output matches golden output
assert (expected_outputs == output_vectors[0]).all()
def run(self):
self.start()
image = cv2.imread(self.image)
image = cv2.resize(image, (128, 128))
image = np.array(image, dtype=np.float32) / 255.0
# ONNX, Caffe and TF parsers also exist.
parser = ann.ITfLiteParser()
network = parser.CreateNetworkFromBinaryFile(self.model)
graph_id = 0
input_names = parser.GetSubgraphInputTensorNames(graph_id)
input_binding_info = parser.GetNetworkInputBindingInfo(
graph_id, input_names[0])
input_tensor_id = input_binding_info[0]
input_tensor_info = input_binding_info[1]
# Create a runtime object that will perform inference.
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
# Backend choices earlier in the list have higher preference.
preferredBackends = [ann.BackendId('CpuAcc'), ann.BackendId('CpuRef')]
opt_network, messages = ann.Optimize(network, preferredBackends,
runtime.GetDeviceSpec(),
ann.OptimizerOptions())
# Load the optimized network into the runtime.
net_id, _ = runtime.LoadNetwork(opt_network)
# Create an inputTensor for inference.
input_tensors = ann.make_input_tensors([input_binding_info], [image])
# Get output binding information for an output layer by using the layer
# name.
output_names = parser.GetSubgraphOutputTensorNames(graph_id)
output_binding_info = parser.GetNetworkOutputBindingInfo(
0, output_names[0])
output_tensors = ann.make_output_tensors([output_binding_info])
start = timer()
runtime.EnqueueWorkload(0, input_tensors, output_tensors)
end = timer()
print('Elapsed time is ', (end - start) * 1000, 'ms')
output, output_tensor_info = ann.from_output_tensor(
output_tensors[0][1])
print(f"Output tensor info: {output_tensor_info}")
print(output)
j = np.argmax(output)
if j == 0:
print("Non-Fire")
else:
print("Fire")
def test_deserializer_end_to_end(shared_data_folder):
parser = ann.IDeserializer()
network = parser.CreateNetworkFromBinary(
os.path.join(shared_data_folder, "mock_model.armnn"))
# use 0 as a dummy value for layer_id, which is unused in the actual implementation
layer_id = 0
input_name = 'input_1'
output_name = 'dense/Softmax'
input_binding_info = parser.GetNetworkInputBindingInfo(
layer_id, input_name)
preferred_backends = [ann.BackendId('CpuAcc'), ann.BackendId('CpuRef')]
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
opt_network, messages = ann.Optimize(network, preferred_backends,
runtime.GetDeviceSpec(),
ann.OptimizerOptions())
assert 0 == len(messages)
net_id, messages = runtime.LoadNetwork(opt_network)
assert "" == messages
# Load test image data stored in input_lite.npy
input_tensor_data = np.load(
os.path.join(shared_data_folder, 'deserializer/input_lite.npy'))
input_tensors = ann.make_input_tensors([input_binding_info],
[input_tensor_data])
output_tensors = []
out_bind_info = parser.GetNetworkOutputBindingInfo(layer_id, output_name)
out_tensor_info = out_bind_info[1]
out_tensor_id = out_bind_info[0]
output_tensors.append((out_tensor_id, ann.Tensor(out_tensor_info)))
runtime.EnqueueWorkload(net_id, input_tensors, output_tensors)
output_vectors = []
for index, out_tensor in enumerate(output_tensors):
output_vectors.append(out_tensor[1].get_memory_area())
# Load golden output file for result comparison.
expected_outputs = np.load(
os.path.join(shared_data_folder,
'deserializer/golden_output_lite.npy'))
# Check that output matches golden output
assert (expected_outputs == output_vectors[0]).all()
def test_profiler_workloads(mock_profiler, exec_times, unit, backend,
workload):
preferred_backends = [
ann.BackendId('CpuRef'),
ann.BackendId('CpuAcc'),
ann.BackendId('GpuAcc'),
ann.BackendId('EthosNAcc')
]
profiling_data_obj = ann.get_profiling_data(mock_profiler,
preferred_backends)
work_load_exec = profiling_data_obj.per_workload_execution_data[workload]
assert work_load_exec["execution_time"] == exec_times
assert work_load_exec["time_unit"] == unit
assert work_load_exec["backend"] == backend
def mock_model_runtime(shared_data_folder):
parser = ann.ITfLiteParser()
network = parser.CreateNetworkFromBinaryFile(os.path.join(shared_data_folder, 'mock_model.tflite'))
graph_id = 0
input_binding_info = parser.GetNetworkInputBindingInfo(graph_id, "input_1")
input_tensor_data = np.load(os.path.join(shared_data_folder, 'tflite_parser/input_lite.npy'))
preferred_backends = [ann.BackendId('CpuRef')]
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
opt_network, messages = ann.Optimize(network, preferred_backends, runtime.GetDeviceSpec(), ann.OptimizerOptions())
print(messages)
net_id, messages = runtime.LoadNetwork(opt_network)
print(messages)
input_tensors = ann.make_input_tensors([input_binding_info], [input_tensor_data])
output_names = parser.GetSubgraphOutputTensorNames(graph_id)
outputs_binding_info = []
for output_name in output_names:
outputs_binding_info.append(parser.GetNetworkOutputBindingInfo(graph_id, output_name))
output_tensors = ann.make_output_tensors(outputs_binding_info)
yield runtime, net_id, input_tensors, output_tensors
def test_optimize_executes_successfully_for_neon_backend_only(network_file, get_runtime):
preferred_backends = [ann.BackendId('CpuAcc')]
network = get_runtime[1]
runtime = get_runtime[2]
opt_network, messages = ann.Optimize(network, preferred_backends, runtime.GetDeviceSpec(), ann.OptimizerOptions())
assert 0 == len(messages)
assert opt_network
def test_inference_exec(mock_profiler):
preferred_backends = [
ann.BackendId('CpuRef'),
ann.BackendId('CpuAcc'),
ann.BackendId('GpuAcc'),
ann.BackendId('EthosNAcc')
]
profiling_data_obj = ann.get_profiling_data(mock_profiler,
preferred_backends)
assert (len(profiling_data_obj.inference_data) > 0)
assert (len(profiling_data_obj.per_workload_execution_data) > 0)
# Check each total execution time
assert (profiling_data_obj.inference_data["execution_time"] == [
1.1, 2.2, 3.3, 4.4, 5.5, 6.6
])
assert (profiling_data_obj.inference_data["time_unit"] == "us")
def test_optimize_executes_successfully(network_file, get_runtime):
preferred_backends = [ann.BackendId('CpuRef')]
network = get_runtime[1]
runtime = get_runtime[2]
opt_network, messages = ann.Optimize(network, preferred_backends, runtime.GetDeviceSpec(), ann.OptimizerOptions())
assert len(messages) == 0, 'With only CpuRef, there should be no warnings irrelevant of architecture.'
assert opt_network
def test_optimize_fails_for_invalid_backends(network_file, get_runtime):
invalid_backends = [ann.BackendId('Unknown')]
network = get_runtime[1]
runtime = get_runtime[2]
with pytest.raises(RuntimeError) as err:
ann.Optimize(network, invalid_backends, runtime.GetDeviceSpec(), ann.OptimizerOptions())
expected_error_message = "None of the preferred backends [Unknown ] are supported."
assert expected_error_message in str(err.value)
def create_network(model_file: str, backends: list):
"""
Creates a network based on the model file and a list of backends.
Args:
model_file: User-specified model file.
backends: List of backends to optimize network.
Returns:
net_id: Unique ID of the network to run.
runtime: Runtime context for executing inference.
input_binding_info: Contains essential information about the model input.
output_binding_info: Used to map output tensor and its memory.
"""
if not os.path.exists(model_file):
raise FileNotFoundError(f'Model file not found for: {model_file}')
# Determine which parser to create based on model file extension
parser = None
_, ext = os.path.splitext(model_file)
if ext == '.tflite':
parser = ann.ITfLiteParser()
elif ext == '.pb':
parser = ann.ITfParser()
elif ext == '.onnx':
parser = ann.IOnnxParser()
assert (parser is not None)
network = parser.CreateNetworkFromBinaryFile(model_file)
# Specify backends to optimize network
preferred_backends = []
for b in backends:
preferred_backends.append(ann.BackendId(b))
# Select appropriate device context and optimize the network for that device
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
opt_network, messages = ann.Optimize(network, preferred_backends, runtime.GetDeviceSpec(),
ann.OptimizerOptions())
print(f'Preferred backends: {backends}\n{runtime.GetDeviceSpec()}\n'
f'Optimization warnings: {messages}')
# Load the optimized network onto the Runtime device
net_id, _ = runtime.LoadNetwork(opt_network)
# Get input and output binding information
graph_id = parser.GetSubgraphCount() - 1
input_names = parser.GetSubgraphInputTensorNames(graph_id)
input_binding_info = parser.GetNetworkInputBindingInfo(graph_id, input_names[0])
output_names = parser.GetSubgraphOutputTensorNames(graph_id)
output_binding_info = []
for output_name in output_names:
outBindInfo = parser.GetNetworkOutputBindingInfo(graph_id, output_name)
output_binding_info.append(outBindInfo)
return net_id, runtime, input_binding_info, output_binding_info
def test_onnx_parser_end_to_end(shared_data_folder):
parser = ann.IOnnxParser = ann.IOnnxParser()
network = parser.CreateNetworkFromBinaryFile(
os.path.join(shared_data_folder, 'mock_model.onnx'))
# load test image data stored in input_onnx.npy
input_binding_info = parser.GetNetworkInputBindingInfo("input")
input_tensor_data = np.load(
os.path.join(shared_data_folder,
'onnx_parser/input_onnx.npy')).astype(np.float32)
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
preferred_backends = [ann.BackendId('CpuAcc'), ann.BackendId('CpuRef')]
opt_network, messages = ann.Optimize(network, preferred_backends,
runtime.GetDeviceSpec(),
ann.OptimizerOptions())
assert 0 == len(messages)
net_id, messages = runtime.LoadNetwork(opt_network)
assert "" == messages
input_tensors = ann.make_input_tensors([input_binding_info],
[input_tensor_data])
output_tensors = ann.make_output_tensors(
[parser.GetNetworkOutputBindingInfo("output")])
runtime.EnqueueWorkload(net_id, input_tensors, output_tensors)
output = ann.workload_tensors_to_ndarray(output_tensors)
# Load golden output file for result comparison.
golden_output = np.load(
os.path.join(shared_data_folder, 'onnx_parser/golden_output_onnx.npy'))
# Check that output matches golden output to 4 decimal places (there are slight rounding differences after this)
np.testing.assert_almost_equal(output[0], golden_output, decimal=4)
def __create_network(model_file: str, backends: list, parser=None):
"""Creates a network based on a file and parser type.
Args:
model_file (str): Path of the model file.
backends (list): List of backends to use when running inference.
parser_type: Parser instance. (pyarmnn.ITFliteParser/pyarmnn.IOnnxParser...)
Returns:
int: Network ID.
int: Graph ID.
IParser: TF Lite parser instance.
IRuntime: Runtime object instance.
"""
args = parse_command_line()
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
if parser is None:
# try to determine what parser to create based on model extension
_, ext = os.path.splitext(model_file)
if ext == ".onnx":
parser = ann.IOnnxParser()
elif ext == ".tflite":
parser = ann.ITfLiteParser()
assert (parser is not None)
network = parser.CreateNetworkFromBinaryFile(model_file)
preferred_backends = []
for b in backends:
preferred_backends.append(ann.BackendId(b))
opt_network, messages = ann.Optimize(network, preferred_backends,
runtime.GetDeviceSpec(),
ann.OptimizerOptions())
if args.verbose:
for m in messages:
warnings.warn(m)
net_id, w = runtime.LoadNetwork(opt_network)
if args.verbose and w:
warnings.warn(w)
return net_id, parser, runtime
def random_runtime(shared_data_folder):
parser = ann.ITfLiteParser()
network = parser.CreateNetworkFromBinaryFile(
os.path.join(shared_data_folder, 'mock_model.tflite'))
preferred_backends = [ann.BackendId('CpuRef')]
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
graphs_count = parser.GetSubgraphCount()
graph_id = graphs_count - 1
input_names = parser.GetSubgraphInputTensorNames(graph_id)
input_binding_info = parser.GetNetworkInputBindingInfo(
graph_id, input_names[0])
input_tensor_id = input_binding_info[0]
input_tensor_info = input_binding_info[1]
output_names = parser.GetSubgraphOutputTensorNames(graph_id)
input_data = np.random.randint(255,
size=input_tensor_info.GetNumElements(),
dtype=np.uint8)
const_tensor_pair = (input_tensor_id,
ann.ConstTensor(input_tensor_info, input_data))
input_tensors = [const_tensor_pair]
output_tensors = []
for index, output_name in enumerate(output_names):
out_bind_info = parser.GetNetworkOutputBindingInfo(
graph_id, output_name)
out_tensor_info = out_bind_info[1]
out_tensor_id = out_bind_info[0]
output_tensors.append((out_tensor_id, ann.Tensor(out_tensor_info)))
yield preferred_backends, network, runtime, input_tensors, output_tensors
input_names = parser.GetSubgraphInputTensorNames(graph_id)
input_binding_info = parser.GetNetworkInputBindingInfo(graph_id,
input_names[0])
input_tensor_id = input_binding_info[0]
input_tensor_info = input_binding_info[1]
print(f"""
tensor id: {input_tensor_id},
tensor info: {input_tensor_info}
""")
# Create a runtime object that will perform inference.
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
# Backend choices earlier in the list have higher preference.
preferredBackends = [ann.BackendId('CpuAcc'), ann.BackendId('CpuRef')]
opt_network, messages = ann.Optimize(network, preferredBackends,
runtime.GetDeviceSpec(),
ann.OptimizerOptions())
# Load the optimized network into the runtime.
net_id, _ = runtime.LoadNetwork(opt_network)
print(f"Loaded network, id={net_id}")
# Create an inputTensor for inference.
input_tensors = ann.make_input_tensors([input_binding_info], [image])
# Get output binding information for an output layer by using the layer name.
output_names = parser.GetSubgraphOutputTensorNames(graph_id)
output_binding_info = parser.GetNetworkOutputBindingInfo(0, output_names[0])
output_tensors = ann.make_output_tensors([output_binding_info])
def test_add_constant_layer_to_fully_connected():
inputWidth = 1
inputHeight = 1
inputChannels = 5
inputNum = 2
outputChannels = 3
outputNum = 2
inputShape = (inputNum, inputChannels, inputHeight, inputWidth)
outputShape = (outputNum, outputChannels)
weightsShape = (inputChannels, outputChannels)
biasShape = (outputChannels, )
input = np.array([[1.0, 2.0, 3.0, 4.0, 5.0], [5.0, 4.0, 3.0, 2.0, 1.0]],
dtype=np.float32)
weights = np.array(
[[.5, 2., .5], [.5, 2., 1.], [.5, 2., 2.], [.5, 2., 3.], [.5, 2., 4.]],
dtype=np.float32)
biasValues = np.array([10, 20, 30], dtype=np.float32)
expectedOutput = np.array([[
0.5 + 1.0 + 1.5 + 2.0 + 2.5 + biasValues[0], 2.0 + 4.0 + 6.0 + 8.0 +
10. + biasValues[1], 0.5 + 2.0 + 6.0 + 12. + 20. + biasValues[2]
],
[
2.5 + 2.0 + 1.5 + 1.0 + 0.5 + biasValues[0],
10.0 + 8.0 + 6.0 + 4.0 + 2. + biasValues[1],
2.5 + 4.0 + 6.0 + 6. + 4. + biasValues[2]
]],
dtype=np.float32)
network = ann.INetwork()
input_info = ann.TensorInfo(ann.TensorShape(inputShape),
ann.DataType_Float32, 0, 0, True)
input_tensor = ann.ConstTensor(input_info, input)
input_layer = network.AddInputLayer(0, "input")
w_info = ann.TensorInfo(ann.TensorShape(weightsShape),
ann.DataType_Float32, 0, 0, True)
w_tensor = ann.ConstTensor(w_info, weights)
w_layer = network.AddConstantLayer(w_tensor, "weights")
b_info = ann.TensorInfo(ann.TensorShape(biasShape), ann.DataType_Float32,
0, 0, True)
b_tensor = ann.ConstTensor(b_info, biasValues)
b_layer = network.AddConstantLayer(b_tensor, "bias")
fc_descriptor = ann.FullyConnectedDescriptor()
fc_descriptor.m_BiasEnabled = True
fc_descriptor.m_ConstantWeights = True
fully_connected = network.AddFullyConnectedLayer(fc_descriptor, "fc")
output_info = ann.TensorInfo(ann.TensorShape(outputShape),
ann.DataType_Float32)
output_tensor = ann.Tensor(output_info, np.zeros([1, 1], dtype=np.float32))
output = network.AddOutputLayer(0, "output")
input_layer.GetOutputSlot(0).Connect(fully_connected.GetInputSlot(0))
w_layer.GetOutputSlot(0).Connect(fully_connected.GetInputSlot(1))
b_layer.GetOutputSlot(0).Connect(fully_connected.GetInputSlot(2))
fully_connected.GetOutputSlot(0).Connect(output.GetInputSlot(0))
input_layer.GetOutputSlot(0).SetTensorInfo(input_info)
w_layer.GetOutputSlot(0).SetTensorInfo(w_info)
b_layer.GetOutputSlot(0).SetTensorInfo(b_info)
fully_connected.GetOutputSlot(0).SetTensorInfo(output_info)
preferred_backends = [ann.BackendId('CpuRef')]
options = ann.CreationOptions()
runtime = ann.IRuntime(options)
opt_network, messages = ann.Optimize(network, preferred_backends,
runtime.GetDeviceSpec(),
ann.OptimizerOptions())
net_id, messages = runtime.LoadNetwork(opt_network)
input_tensors = [(0, input_tensor)]
output_tensors = [(0, output_tensor)]
runtime.EnqueueWorkload(net_id, input_tensors, output_tensors)
output_vectors = ann.workload_tensors_to_ndarray(output_tensors)
assert (output_vectors == expectedOutput).all()
