def run_specify_abi(self, flags, configs, target_abi):
if target_abi not in self.target_abis:
six.print_('The device %s with soc %s do not support the abi %s' %
(self.device_name, self.target_socs, target_abi))
return
library_name = configs[YAMLKeyword.library_name]
model_output_dirs = []
for model_name in configs[YAMLKeyword.models]:
check_model_converted(library_name, model_name,
configs[YAMLKeyword.model_graph_format],
configs[YAMLKeyword.model_data_format],
target_abi)
if target_abi != ABIType.host:
self.clear_data_dir()
MaceLogger.header(
StringFormatter.block(
'Run model {} on {}'.format(model_name, self.device_name)))
model_config = configs[YAMLKeyword.models][model_name]
model_runtime = model_config[YAMLKeyword.runtime]
subgraphs = model_config[YAMLKeyword.subgraphs]
model_output_base_dir, model_output_dir, mace_model_dir = \
get_build_model_dirs(
library_name, model_name, target_abi, self,
model_config[YAMLKeyword.model_file_path])
# clear temp model output dir
if os.path.exists(model_output_dir):
sh.rm('-rf', model_output_dir)
os.makedirs(model_output_dir)
tuning = False
if not flags.address_sanitizer \
and target_abi != ABIType.host \
and (configs[YAMLKeyword.target_socs]
or flags.target_socs) \
and self.target_socs \
and model_runtime in [RuntimeType.gpu,
RuntimeType.cpu_gpu] \
and not flags.disable_tuning:
self.tuning(library_name, model_name, model_config,
configs[YAMLKeyword.model_graph_format],
configs[YAMLKeyword.model_data_format],
target_abi, flags.mace_lib_type)
model_output_dirs.append(model_output_dir)
self.clear_data_dir()
tuning = True
accuracy_validation_script = \
subgraphs[0][YAMLKeyword.accuracy_validation_script]
output_configs = []
if not accuracy_validation_script and flags.layers != "-1":
mace_check(configs[YAMLKeyword.model_graph_format] ==
ModelFormat.file and
configs[YAMLKeyword.model_data_format] ==
ModelFormat.file, "Device",
"'--layers' only supports model format 'file'.")
output_configs = self.get_layers(mace_model_dir,
model_name,
flags.layers)
# run for specified soc
if not subgraphs[0][YAMLKeyword.check_tensors]:
output_nodes = subgraphs[0][YAMLKeyword.output_tensors]
output_shapes = subgraphs[0][YAMLKeyword.output_shapes]
else:
output_nodes = subgraphs[0][YAMLKeyword.check_tensors]
output_shapes = subgraphs[0][YAMLKeyword.check_shapes]
model_path = "%s/%s.pb" % (mace_model_dir, model_name)
output_config = {YAMLKeyword.model_file_path: model_path,
YAMLKeyword.output_tensors: output_nodes,
YAMLKeyword.output_shapes: output_shapes}
output_configs.append(output_config)
runtime_list = []
if target_abi == ABIType.host:
runtime_list.append(RuntimeType.cpu)
elif model_runtime == RuntimeType.cpu_gpu:
runtime_list.extend([RuntimeType.cpu, RuntimeType.gpu])
else:
runtime_list.append(model_runtime)
if accuracy_validation_script:
flags.validate = False
flags.report = False
import imp
accuracy_val_module = imp.load_source(
'accuracy_val_module',
accuracy_validation_script)
for runtime in runtime_list:
accuracy_validator = \
accuracy_val_module.AccuracyValidator()
sample_size = accuracy_validator.sample_size()
val_batch_size = accuracy_validator.batch_size()
for i in range(0, sample_size, val_batch_size):
inputs = accuracy_validator.preprocess(
i, i + val_batch_size)
sh_commands.gen_input(
model_output_dir,
subgraphs[0][YAMLKeyword.input_tensors],
subgraphs[0][YAMLKeyword.input_shapes],
input_data_types=subgraphs[0][YAMLKeyword.input_data_types], # noqa
input_data_map=inputs)
self.run_model(flags, configs, target_abi, model_name,
output_configs[-1], runtime, tuning)
accuracy_validator.postprocess(
i, i + val_batch_size,
self.get_output_map(
target_abi,
output_nodes,
subgraphs[0][YAMLKeyword.output_shapes],
model_output_dir))
accuracy_validator.result()
else:
sh_commands.gen_input(
model_output_dir,
subgraphs[0][YAMLKeyword.input_tensors],
subgraphs[0][YAMLKeyword.input_shapes],
subgraphs[0][YAMLKeyword.validation_inputs_data],
input_ranges=subgraphs[0][YAMLKeyword.input_ranges],
input_data_types=subgraphs[0][YAMLKeyword.input_data_types]
)
for runtime in runtime_list:
device_type = parse_device_type(runtime)
log_dir = mace_model_dir + "/" + runtime
if os.path.exists(log_dir):
sh.rm('-rf', log_dir)
os.makedirs(log_dir)
for output_config in output_configs:
self.run_model(flags, configs, target_abi, model_name,
output_config, runtime, tuning)
if flags.validate:
log_file = ""
if flags.layers != "-1":
log_file = log_dir + "/log.csv"
model_file_path, weight_file_path = \
get_model_files(
model_config[YAMLKeyword.model_file_path],
model_config[
YAMLKeyword.model_sha256_checksum],
BUILD_DOWNLOADS_DIR,
model_config[YAMLKeyword.weight_file_path],
model_config[
YAMLKeyword.weight_sha256_checksum])
validate_type = device_type
if device_type in [DeviceType.CPU,
DeviceType.GPU] and \
(model_config[YAMLKeyword.quantize] == 1 or
model_config[YAMLKeyword.quantize_large_weights] == 1): # noqa
validate_type = DeviceType.QUANTIZE
dockerfile_path, docker_image_tag = \
get_dockerfile_info(
model_config.get(
YAMLKeyword.dockerfile_path),
model_config.get(
YAMLKeyword.dockerfile_sha256_checksum), # noqa
model_config.get(
YAMLKeyword.docker_image_tag)
) if YAMLKeyword.dockerfile_path \
in model_config \
else ("third_party/caffe", "lastest")
sh_commands.validate_model(
abi=target_abi,
device=self,
model_file_path=model_file_path,
weight_file_path=weight_file_path,
docker_image_tag=docker_image_tag,
dockerfile_path=dockerfile_path,
platform=model_config[YAMLKeyword.platform],
device_type=device_type,
input_nodes=subgraphs[0][
YAMLKeyword.input_tensors],
output_nodes=output_config[
YAMLKeyword.output_tensors],
input_shapes=subgraphs[0][
YAMLKeyword.input_shapes],
output_shapes=output_config[
YAMLKeyword.output_shapes],
input_data_formats=subgraphs[0][
YAMLKeyword.input_data_formats],
output_data_formats=subgraphs[0][
YAMLKeyword.output_data_formats],
model_output_dir=model_output_dir,
input_data_types=subgraphs[0][
YAMLKeyword.input_data_types],
caffe_env=flags.caffe_env,
validation_threshold=subgraphs[0][
YAMLKeyword.validation_threshold][
validate_type],
backend=subgraphs[0][YAMLKeyword.backend],
validation_outputs_data=subgraphs[0][
YAMLKeyword.validation_outputs_data],
log_file=log_file,
)
if flags.round > 0:
tuned = tuning and device_type == DeviceType.GPU
self.report_run_statistics(
target_abi=target_abi,
model_name=model_name,
device_type=device_type,
flags=flags,
tuned=tuned)
if model_output_dirs:
opencl_output_bin_path = get_opencl_binary_output_path(
library_name, target_abi, self
)
opencl_parameter_bin_path = get_opencl_parameter_output_path(
library_name, target_abi, self
)
# clear opencl output dir
if os.path.exists(opencl_output_bin_path):
sh.rm('-rf', opencl_output_bin_path)
if os.path.exists(opencl_parameter_bin_path):
sh.rm('-rf', opencl_parameter_bin_path)
# merge all model's opencl binaries together
sh_commands.merge_opencl_binaries(
model_output_dirs, CL_COMPILED_BINARY_FILE_NAME,
opencl_output_bin_path
)
# merge all model's opencl parameter together
sh_commands.merge_opencl_parameters(
model_output_dirs, CL_TUNED_PARAMETER_FILE_NAME,
opencl_parameter_bin_path
)
sh_commands.gen_opencl_binary_cpps(
opencl_output_bin_path,
opencl_parameter_bin_path,
opencl_output_bin_path + '.cc',
opencl_parameter_bin_path + '.cc')
def tuning(self, library_name, model_name, model_config,
model_graph_format, model_data_format,
target_abi, mace_lib_type):
six.print_('* Tuning, it may take some time')
build_tmp_binary_dir = get_build_binary_dir(library_name, target_abi)
mace_run_name = MACE_RUN_STATIC_NAME
link_dynamic = False
if mace_lib_type == MACELibType.dynamic:
mace_run_name = MACE_RUN_DYNAMIC_NAME
link_dynamic = True
embed_model_data = model_data_format == ModelFormat.code
# build for specified soc
# device_wrapper = DeviceWrapper(device)
model_output_base_dir, model_output_dir, mace_model_dir = \
get_build_model_dirs(
library_name, model_name, target_abi, self,
model_config[YAMLKeyword.model_file_path])
self.clear_data_dir()
subgraphs = model_config[YAMLKeyword.subgraphs]
# generate input data
sh_commands.gen_input(
model_output_dir,
subgraphs[0][YAMLKeyword.input_tensors],
subgraphs[0][YAMLKeyword.input_shapes],
subgraphs[0][YAMLKeyword.validation_inputs_data],
input_ranges=subgraphs[0][YAMLKeyword.input_ranges],
input_data_types=subgraphs[0][YAMLKeyword.input_data_types]
)
self.tuning_run(
abi=target_abi,
target_dir=build_tmp_binary_dir,
target_name=mace_run_name,
vlog_level=0,
embed_model_data=embed_model_data,
model_output_dir=model_output_dir,
input_nodes=subgraphs[0][YAMLKeyword.input_tensors],
output_nodes=subgraphs[0][YAMLKeyword.output_tensors],
input_shapes=subgraphs[0][YAMLKeyword.input_shapes],
output_shapes=subgraphs[0][YAMLKeyword.output_shapes],
input_data_formats=subgraphs[0][YAMLKeyword.input_data_formats],
output_data_formats=subgraphs[0][YAMLKeyword.output_data_formats],
mace_model_dir=mace_model_dir,
model_tag=model_name,
device_type=DeviceType.GPU,
running_round=0,
restart_round=1,
limit_opencl_kernel_time=model_config[
YAMLKeyword.limit_opencl_kernel_time],
opencl_queue_window_size=model_config[
YAMLKeyword.opencl_queue_window_size],
tuning=True,
out_of_range_check=False,
model_graph_format=model_graph_format,
opencl_binary_file='',
opencl_parameter_file='',
libmace_dynamic_library_path=LIBMACE_DYNAMIC_PATH,
link_dynamic=link_dynamic,
)
# pull opencl library
self.pull(self.interior_dir, CL_COMPILED_BINARY_FILE_NAME,
'{}/{}'.format(model_output_dir,
BUILD_TMP_OPENCL_BIN_DIR))
# pull opencl parameter
self.pull_from_data_dir(CL_TUNED_PARAMETER_FILE_NAME,
'{}/{}'.format(model_output_dir,
BUILD_TMP_OPENCL_BIN_DIR))
six.print_('Tuning done! \n')
