def test_model_converted_from_kaldi(self, mocker):
mock = mocker.patch(
'accuracy_checker.launcher.dlsdk_launcher.convert_model',
return_value=('converted_model', 'converted_weights'))
config = {
'framework': 'dlsdk',
'kaldi_model': '/path/to/source_models/custom_model',
'device': 'cpu',
'_models_prefix': '/path/to/source_models',
'adapter': 'classification',
'should_log_cmd': False
}
DLSDKLauncher(config)
mock.assert_called_once_with('custom_model',
'/path/to/source_models/custom_model',
'',
'',
FrameworkParameters('kaldi', False), [],
None,
None,
None,
None,
should_log_cmd=False)
def test_model_converted_with_mo_flags(self, mocker):
mock = mocker.patch(
'accuracy_checker.launcher.dlsdk_launcher.convert_model',
return_value=('converted_model', 'converted_weights'))
config = {
'framework': 'dlsdk',
'caffe_model': '/path/to/source_models/custom_model',
'caffe_weights': '/path/to/source_models/custom_weights',
'device': 'cpu',
'bitstream': Path('custom_bitstream'),
'_models_prefix': '/path/to/source_models',
'mo_flags': ['reverse_input_channels'],
'adapter': 'classification',
'should_log_cmd': False
}
DLSDKLauncher(config)
mock.assert_called_once_with('custom_model',
'/path/to/source_models/custom_model',
'/path/to/source_models/custom_weights',
'',
FrameworkParameters('caffe', False), [],
None, ['reverse_input_channels'],
None,
None,
should_log_cmd=False)
def test_model_converted_with_mo_params(self, mocker):
mock = mocker.patch(
'accuracy_checker.launcher.dlsdk_launcher.convert_model',
return_value=('converted_model', 'converted_weights')
)
config = {
'framework': "dlsdk",
'caffe_model': '/path/to/source_models/custom_model',
'caffe_weights': '/path/to/source_models/custom_weights',
'device': 'cpu',
'bitstream': Path('custom_bitstream'),
'mo_params': {'data_type': 'FP16'},
'adapter': 'classification',
'should_log_cmd': False
}
DLSDKLauncher(config)
mock.assert_called_once_with(
'custom_model', '/path/to/source_models/custom_model', '/path/to/source_models/custom_weights', '',
FrameworkParameters('caffe', False),
[], {'data_type': 'FP16'}, None, None, None, None, should_log_cmd=False
)
def process() -> CalibrationConfiguration:
args, unknown_args = CommandLineReader.parser().parse_known_args()
if unknown_args:
info("unknown command line arguments: {0}".format(unknown_args))
if not args.simplified_mode:
args.target_framework = "dlsdk"
args.aocl = None
merged_config, mode = ConfigReader.merge(args)
updated_config = ConfigurationFilter.filter(
merged_config, args.metric_name, args.metric_type,
default_logger)
if len(updated_config['models']) > 1:
raise ValueError("too much models")
if len(updated_config['models'][0]['launchers']) > 1:
raise ValueError("too much launchers")
launcher = updated_config['models'][0]['launchers'][0]
if 'caffe_model' in launcher or 'tf_model' in launcher or 'tf_meta' in launcher or 'mxnet_weights' in launcher or 'onnx_model' in launcher:
if args.converted_models:
tmp_directory = None
else:
tmp_directory = tempfile.mkdtemp(".converted_models")
launcher['mo_params']['output_dir'] = tmp_directory
if 'caffe_model' in launcher:
framework = FrameworkParameters('caffe', False)
output_model = Path.get_model(
str(launcher['caffe_model']), "_i8",
str(args.output_dir) if args.output_dir else None)
output_weights = Path.get_weights(
str(launcher['caffe_weights']), "_i8",
str(args.output_dir) if args.output_dir else None)
elif 'tf_model' in launcher:
framework = FrameworkParameters('tf', False)
output_model = Path.get_model(
str(launcher['tf_model']), "_i8",
str(args.output_dir) if args.output_dir else None)
output_weights = Path.get_weights(
str(launcher['tf_model']), "_i8",
str(args.output_dir) if args.output_dir else None)
elif 'tf_meta' in launcher:
framework = FrameworkParameters('tf', True)
output_model = Path.get_model(
str(launcher['tf_meta']), "_i8",
str(args.output_dir) if args.output_dir else None)
output_weights = Path.get_weights(
str(launcher['tf_meta']), "_i8",
str(args.output_dir) if args.output_dir else None)
elif 'mxnet_weights' in launcher:
framework = FrameworkParameters('mxnet', False)
output_model = Path.get_model(
str(launcher['mxnet_weights']), "_i8",
str(args.output_dir) if args.output_dir else None)
output_weights = Path.get_weights(
str(launcher['mxnet_weights']), "_i8",
str(args.output_dir) if args.output_dir else None)
elif 'onnx_model' in launcher:
framework = FrameworkParameters('onnx', False)
output_model = Path.get_model(
str(launcher['onnx_model']), "_i8",
str(args.output_dir) if args.output_dir else None)
output_weights = Path.get_weights(
str(launcher['onnx_model']), "_i8",
str(args.output_dir) if args.output_dir else None)
else:
raise ValueError("unknown model framework")
model, weights = DLSDKLauncher.convert_model(
launcher, framework)
launcher['model'] = model
launcher['weights'] = weights
launcher.pop('caffe_model', None)
launcher.pop('caffe_weights', None)
launcher.pop('tf_model', None)
launcher.pop('tf_meta', None)
launcher.pop('mxnet_weights', None)
launcher.pop('onnx_model', None)
else:
model = launcher['model']
output_model = Path.get_model(
str(model), "_i8",
str(args.output_dir) if args.output_dir else None)
weights = launcher['weights']
output_weights = Path.get_weights(
str(weights), "_i8",
str(args.output_dir) if args.output_dir else None)
tmp_directory = None
batch_size = args.batch_size if args.batch_size else (
launcher['batch'] if 'batch' in launcher else None)
if not batch_size:
with Network(str(launcher['model']),
str(launcher['weights'])) as network:
batch_size = network.ie_network.batch_size
if 'cpu_extensions' in launcher:
cpu_extension = DLSDKLauncher.get_cpu_extension(
launcher['cpu_extensions'], args.cpu_extensions_mode)
launcher['cpu_extensions'] = cpu_extension
else:
cpu_extension = None
if not args.calibrate_fully_connected:
if args.ignore_layer_types is None:
args.ignore_layer_types = []
args.ignore_layer_types.append("FullyConnected")
return CalibrationConfiguration(
config=updated_config,
precision=args.precision,
model=str(model),
weights=str(weights),
tmp_directory=tmp_directory,
output_model=output_model,
output_weights=output_weights,
cpu_extension=str(cpu_extension) if cpu_extension else None,
gpu_extension=str(launcher['gpu_extensions'])
if 'gpu_extensions' in launcher else None,
device=launcher['device'],
batch_size=batch_size,
threshold=args.threshold,
ignore_layer_types=args.ignore_layer_types,
ignore_layer_types_path=args.ignore_layer_types_path,
ignore_layer_names=args.ignore_layer_names,
ignore_layer_names_path=args.ignore_layer_names_path,
benchmark_iterations_count=args.benchmark_iterations_count,
progress=(None if args.progress == 'None' else args.progress),
threshold_step=args.threshold_step,
threshold_boundary=args.threshold_boundary,
simplified_mode=args.simplified_mode)
else:
file_name = ntpath.basename(str(args.models))
model = os.path.splitext(file_name)
output_model = model[0] + "_i8"
if args.output_dir:
output_model = str(args.output_dir.joinpath(output_model))
batch_size = args.batch_size if args.batch_size else 0
precision = args.precision if args.precision.lower() in [
'fp16', 'fp32'
] else ''
return CalibrationConfiguration(
config=args,
precision=precision,
model=str(args.models),
weights=None,
tmp_directory=None,
output_model=output_model,
output_weights=None,
cpu_extension=str(args.extensions) if args.extensions else '',
gpu_extension=None,
device=args.target_devices,
batch_size=batch_size,
threshold=None,
ignore_layer_types=None,
ignore_layer_types_path=None,
ignore_layer_names=None,
ignore_layer_names_path=None,
benchmark_iterations_count=None,
progress=(None if args.progress == 'None' else args.progress),
threshold_step=None,
threshold_boundary=None,
simplified_mode=args.simplified_mode)