def add_onnx_loader(self, script, disable_outputs=None, suffix=None):
if self.model_args.model_type == "onnx":
script.add_import(imports=["OnnxFromPath"],
frm="polygraphy.backend.onnx")
loader_str = Script.invoke("OnnxFromPath",
self.model_args.model_file)
loader_name = script.add_loader(loader_str,
"load_onnx",
suffix=suffix)
else:
if self.tf2onnx_loader_args is None:
G_LOGGER.critical(
"Could not load: {:}. Is it an ONNX model?".format(
self.model_args.model_file))
loader_name = self.tf2onnx_loader_args.add_to_script(script)
modify_onnx_str = self._get_modify_onnx_str(
script, loader_name, disable_outputs=disable_outputs)
if modify_onnx_str is not None:
loader_name = script.add_loader(modify_onnx_str, "modify_onnx")
SAVE_ONNX = "SaveOnnx"
save_onnx_str = Script.invoke(SAVE_ONNX,
loader_name,
path=self.save_onnx)
if save_onnx_str != Script.invoke(SAVE_ONNX, loader_name):
script.add_import(imports=[SAVE_ONNX],
frm="polygraphy.backend.onnx")
loader_name = script.add_loader(save_onnx_str, "save_onnx")
return loader_name
def add_trt_runner(script, args, data_loader_name):
script.add_import(imports=["TrtRunner"], frm="polygraphy.backend.trt")
if args.model_type == "engine":
loader_name = tool_util.add_trt_serialized_engine_loader(script, args)
else:
script.add_import(imports=["EngineFromNetwork"],
frm="polygraphy.backend.trt")
loader_name = tool_util.add_trt_network_loader(script, args)
config_loader_name = tool_util.add_trt_config_loader(
script, args, data_loader_name=data_loader_name)
loader_str = Script.invoke("EngineFromNetwork",
loader_name,
config=config_loader_name)
loader_name = script.add_loader(loader_str, "build_engine")
SAVE_ENGINE = "SaveEngine"
save_engine = Script.invoke(SAVE_ENGINE,
loader_name,
path=args_util.get(args, "save_engine"))
if save_engine != Script.invoke(SAVE_ENGINE, loader_name):
script.add_import(imports=[SAVE_ENGINE], frm="polygraphy.backend.trt")
loader_name = script.add_loader(save_engine, "save_engine")
script.add_runner(Script.invoke("TrtRunner", loader_name))
def add_to_script(self, script, data_loader_name):
script.add_import(imports=["TrtRunner"], frm="polygraphy.backend.trt")
if self.model_args.model_type == "engine":
loader_name = self.trt_loader_args.add_trt_serialized_engine_loader(
script)
else:
script.add_import(imports=["EngineFromNetwork"],
frm="polygraphy.backend.trt")
loader_name = self.trt_loader_args.add_trt_network_loader(script)
config_loader_name = self.trt_loader_args.add_trt_config_loader(
script, data_loader_name)
loader_str = Script.invoke("EngineFromNetwork",
loader_name,
config=config_loader_name)
loader_name = script.add_loader(loader_str, "build_engine")
SAVE_ENGINE = "SaveEngine"
save_engine = Script.invoke(SAVE_ENGINE,
loader_name,
path=self.save_engine)
if save_engine != Script.invoke(SAVE_ENGINE, loader_name):
script.add_import(imports=[SAVE_ENGINE],
frm="polygraphy.backend.trt")
loader_name = script.add_loader(save_engine, "save_engine")
runner_name = script.add_loader(
Script.invoke("TrtRunner", loader_name), "trt_runner")
script.add_runner(runner_name)
return runner_name
def _make_data_loader(script):
data_loader_name = Inline("data_loader")
input_metadata_str = Inline(
repr(self.model_args.input_shapes
)) if self.model_args.input_shapes else None
if input_metadata_str:
script.add_import(imports=["TensorMetadata"],
frm="polygraphy.common")
data_loader = Script.invoke_if_nondefault(
"DataLoader",
seed=self.seed,
iterations=self.iterations,
input_metadata=input_metadata_str,
int_range=self.int_range,
float_range=self.float_range)
if data_loader is not None:
script.add_import(imports=["DataLoader"],
frm="polygraphy.comparator")
script.append_prefix(
Script.format_str("\n# Inference Inputs Loader\n{:} = {:}",
data_loader_name, Inline(data_loader)))
else:
data_loader_name = None
return data_loader_name
def add_onnx_loader(script, args, disable_outputs=None, suffix=None):
if args_util.get(args, "model_type") == "onnx":
script.add_import(imports=["OnnxFromPath"], frm="polygraphy.backend.onnx")
loader_str = Script.invoke("OnnxFromPath", args_util.get(args, "model_file"))
loader_name = script.add_loader(loader_str, "load_onnx", suffix=suffix)
else:
G_LOGGER.verbose("Attempting to load as a TensorFlow model, using TF2ONNX to convert to ONNX. "
"If this is not correct, please specify --model-type", mode=LogMode.ONCE)
script.add_import(imports=["OnnxFromTfGraph"], frm="polygraphy.backend.onnx")
loader_str = Script.invoke("OnnxFromTfGraph", add_tf_loader(script, args, disable_outputs=True, suffix=suffix),
opset=args_util.get(args, "opset"), fold_constant=False if args_util.get(args, "no_const_folding") else None)
loader_name = script.add_loader(loader_str, "export_onnx_from_tf", suffix=suffix)
modify_onnx_str = get_modify_onnx_str(script, args, loader_name, disable_outputs=disable_outputs)
if modify_onnx_str is not None:
loader_name = script.add_loader(modify_onnx_str, "modify_onnx")
save_onnx = args_util.get(args, "save_onnx")
SAVE_ONNX = "SaveOnnx"
save_onnx_str = Script.invoke(SAVE_ONNX, loader_name, path=save_onnx)
if save_onnx_str != Script.invoke(SAVE_ONNX, loader_name):
script.add_import(imports=[SAVE_ONNX], frm="polygraphy.backend.onnx")
loader_name = script.add_loader(save_onnx_str, "save_onnx")
return loader_name
def add_to_script(self, script, data_loader_name):
script.add_import(imports=["Comparator"], frm="polygraphy.comparator")
script.add_import(imports=["sys"])
RESULTS_VAR_NAME = Inline("results")
comparator_run = Script.invoke("Comparator.run",
script.get_runners(),
warm_up=self.warm_up,
data_loader=data_loader_name,
use_subprocess=self.use_subprocess,
save_inputs_path=self.save_inputs)
script.append_suffix(
Script.format_str("\n# Runner Execution\n{results} = {:}",
Inline(comparator_run),
results=RESULTS_VAR_NAME))
if self.save_results:
G_LOGGER.verbose("Will save runner results to: {:}".format(
self.save_results))
script.add_import(imports=["misc"], frm="polygraphy.util")
script.append_suffix(
Script.format_str(
"\n# Save results\nmisc.pickle_save({:}, {results})",
self.save_results,
results=RESULTS_VAR_NAME))
return RESULTS_VAR_NAME
def add_trt_config_loader(script, args, data_loader_name):
profiles = []
for (min_shape, opt_shape, max_shape) in args_util.get(args, "profiles"):
profile_str = "Profile()"
for name in min_shape.keys():
profile_str += Script.format_str(".add({:}, min={:}, opt={:}, max={:})", name, min_shape[name], opt_shape[name], max_shape[name])
profiles.append(Inline(profile_str))
if profiles:
script.add_import(imports=["Profile"], frm="polygraphy.backend.trt")
sep = Inline("\n{:}".format(constants.TAB))
profiles = Script.format_str("[{:}{:}\n]", sep, Inline((",{:}".format(sep)).join(profiles)))
profile_name = script.add_loader(profiles, "profiles")
else:
profile_name = None
calibrator = None
if args_util.get(args, "int8"):
script.add_import(imports=["DataLoader"], frm="polygraphy.comparator")
script.add_import(imports=["Calibrator"], frm="polygraphy.backend.trt")
calibrator = Script.invoke("Calibrator", data_loader=Inline(data_loader_name) if data_loader_name else Inline("DataLoader()"),
cache=args_util.get(args, "calibration_cache"))
config_loader_str = Script.invoke_if_nondefault("CreateTrtConfig", max_workspace_size=args_util.get(args, "workspace"), tf32=args_util.get(args, "tf32"),
fp16=args_util.get(args, "fp16"), int8=args_util.get(args, "int8"), strict_types=args_util.get(args, "strict_types"),
profiles=profile_name, calibrator=Inline(calibrator) if calibrator else None)
if config_loader_str is not None:
script.add_import(imports=["CreateConfig as CreateTrtConfig"], frm="polygraphy.backend.trt")
config_loader_name = script.add_loader(config_loader_str, "create_trt_config")
else:
config_loader_name = None
return config_loader_name
def add_onnxrt_runner(script, args):
script.add_import(imports=["OnnxrtRunner"],
frm="polygraphy.backend.onnxrt")
onnx_name = tool_util.add_serialized_onnx_loader(script, args)
script.add_import(imports=["SessionFromOnnxBytes"],
frm="polygraphy.backend.onnxrt")
loader_name = script.add_loader(
Script.invoke("SessionFromOnnxBytes", onnx_name),
"build_onnxrt_session")
script.add_runner(Script.invoke("OnnxrtRunner", loader_name))
def add_trt_serialized_engine_loader(self, script):
script.add_import(imports=["EngineFromBytes"],
frm="polygraphy.backend.trt")
script.add_import(imports=["BytesFromPath"],
frm="polygraphy.backend.common")
load_engine = script.add_loader(
Script.invoke("BytesFromPath", self.model_args.model_file),
"load_engine")
return script.add_loader(
Script.invoke("EngineFromBytes",
self._wrap_if_plugins(script, load_engine)),
"deserialize_engine")
def add_trt_config_loader(self, script, data_loader_name):
profiles = []
profile_args = tools_util.parse_profile_shapes(
self.model_args.input_shapes, self.trt_min_shapes,
self.trt_opt_shapes, self.trt_max_shapes)
for (min_shape, opt_shape, max_shape) in profile_args:
profile_str = "Profile()"
for name in min_shape.keys():
profile_str += Script.format_str(
".add({:}, min={:}, opt={:}, max={:})", name,
min_shape[name], opt_shape[name], max_shape[name])
profiles.append(Inline(profile_str))
if profiles:
script.add_import(imports=["Profile"],
frm="polygraphy.backend.trt")
sep = Inline("\n{:}".format(constants.TAB))
profiles = Script.format_str(
"[{:}{:}\n]", sep, Inline((",{:}".format(sep)).join(profiles)))
profile_name = script.add_loader(profiles, "profiles")
else:
profile_name = None
calibrator = None
if self.int8:
script.add_import(imports=["Calibrator"],
frm="polygraphy.backend.trt")
script.add_import(imports=["DataLoader"],
frm="polygraphy.comparator")
calibrator = Script.invoke(
"Calibrator",
data_loader=Inline(data_loader_name)
if data_loader_name else Inline("DataLoader()"),
cache=self.calibration_cache)
config_loader_str = Script.invoke_if_nondefault(
"CreateTrtConfig",
max_workspace_size=self.workspace,
tf32=self.tf32,
fp16=self.fp16,
int8=self.int8,
strict_types=self.strict_types,
profiles=profile_name,
calibrator=Inline(calibrator) if calibrator else None)
if config_loader_str is not None:
script.add_import(imports=["CreateConfig as CreateTrtConfig"],
frm="polygraphy.backend.trt")
config_loader_name = script.add_loader(config_loader_str,
"create_trt_config")
else:
config_loader_name = None
return config_loader_name
def add_tf_loader(script, args, disable_outputs=None, suffix=None):
if disable_outputs:
outputs = None
else:
outputs = _get_outputs_arg(script, args, "tf_outputs")
model_file = args_util.get(args, "model_file")
model_type = args_util.get(args, "model_type")
save_pb = args_util.get(args, "save_pb")
save_tensorboard = args_util.get(args, "save_tensorboard")
if model_type == "ckpt":
G_LOGGER.verbose("Loading a TensorFlow checkpoint. Please ensure you are not using the --use-subprocess flag".format(model_file), mode=LogMode.ONCE)
script.add_import(imports=["GraphFromCkpt"], frm="polygraphy.backend.tf")
loader_id = "load_ckpt"
loader_str = Script.invoke("GraphFromCkpt", model_file, args_util.get(args, "ckpt"))
elif model_type == "keras":
script.add_import(imports=["GraphFromKeras"], frm="polygraphy.backend.tf")
loader_id = "load_keras"
loader_str = Script.invoke("GraphFromKeras", model_file)
else:
script.add_import(imports=["GraphFromFrozen"], frm="polygraphy.backend.tf")
G_LOGGER.verbose("Attempting to load as a frozen graph. If this is not correct, please specify --model-type", mode=LogMode.ONCE)
loader_id = "load_frozen"
loader_str = Script.invoke("GraphFromFrozen", model_file)
loader_name = script.add_loader(loader_str, loader_id, suffix=suffix)
if args_util.get(args, "freeze_graph"):
script.add_import(imports=["OptimizeGraph"], frm="polygraphy.backend.tf")
loader_name = script.add_loader(Script.invoke("OptimizeGraph", loader_name), "optimize_graph", suffix=suffix)
if args_util.get(args, "tftrt"):
script.add_import(imports=["UseTfTrt"], frm="polygraphy.backend.tf")
loader_str = Script.invoke("UseTfTrt", loader_name, max_workspace_size=args_util.get(args, "workspace"), fp16=args_util.get(args, "fp16"), int8=args_util.get(args, "int8"),
max_batch_size=args_util.get(args, "batch_size"), is_dynamic_op=args_util.get(args, "dynamic_op"), minimum_segment_size=args_util.get(args, "minimum_segment_size"))
loader_name = script.add_loader(loader_str, "use_tftrt", suffix=suffix)
MODIFY_TF = "ModifyGraph"
modify_tf_str = Script.invoke(MODIFY_TF, loader_name, outputs=outputs)
if modify_tf_str != Script.invoke(MODIFY_TF, loader_name):
script.add_import(imports=[MODIFY_TF], frm="polygraphy.backend.tf")
loader_name = script.add_loader(modify_tf_str, "modify_tf")
engine_dir = None
if args_util.get(args, "tftrt"):
engine_dir = args_util.get(args, "save_engine")
WRITE_TF = "SaveGraph"
write_tf_str = Script.invoke(WRITE_TF, loader_name, path=save_pb, tensorboard_dir=save_tensorboard, engine_dir=engine_dir)
if write_tf_str != Script.invoke(WRITE_TF, loader_name):
script.add_import(imports=[WRITE_TF], frm="polygraphy.backend.tf")
loader_name = script.add_loader(write_tf_str, "save_tf")
return loader_name
def add_serialized_onnx_loader(script, args, disable_outputs=None):
model_file = args_util.get(args, "model_file")
needs_modify = get_modify_onnx_str(script, args, "check_needs_modify", disable_outputs) is not None
should_import_raw = args_util.get(args, "model_type") == "onnx" and not needs_modify
if should_import_raw:
script.add_import(imports=["BytesFromPath"], frm="polygraphy.backend.common")
onnx_loader = script.add_loader(Script.invoke("BytesFromPath", model_file), "load_serialized_onnx")
else:
script.add_import(imports=["BytesFromOnnx"], frm="polygraphy.backend.onnx")
onnx_loader = add_onnx_loader(script, args, disable_outputs=disable_outputs)
onnx_loader = script.add_loader(Script.invoke("BytesFromOnnx", onnx_loader), "serialize_onnx")
return onnx_loader
def add_to_script(self, script):
script.add_import(imports=["OnnxrtRunner"],
frm="polygraphy.backend.onnxrt")
onnx_name = self.onnx_loader_args.add_serialized_onnx_loader(script)
script.add_import(imports=["SessionFromOnnxBytes"],
frm="polygraphy.backend.onnxrt")
loader_name = script.add_loader(
Script.invoke("SessionFromOnnxBytes", onnx_name),
"build_onnxrt_session")
runner_name = script.add_loader(
Script.invoke("OnnxrtRunner", loader_name), "onnxrt_runner")
script.add_runner(runner_name)
return runner_name
def add_onnxtf_runner(script, args):
script.add_import(imports=["OnnxTfRunner", "OnnxFromPath"],
frm="polygraphy.backend.onnx")
script.add_runner(
Script.invoke(
"OnnxTfRunner",
tool_util.add_onnx_loader(script, args, suffix="_onnxtf")))
def _wrap_if_plugins(script, args, obj_name):
plugins = args_util.get(args, "plugins")
if plugins:
script.add_import(imports=["LoadPlugins"], frm="polygraphy.backend.trt")
loader_str = Script.invoke("LoadPlugins", obj_name, plugins=plugins)
obj_name = script.add_loader(loader_str, "load_plugins")
return obj_name
def add_tf_runner(script, args):
script.add_import(imports=["TfRunner"], frm="polygraphy.backend.tf")
graph_name = tool_util.add_tf_loader(script, args)
config_name = tool_util.add_tf_config_loader(script, args)
script.add_import(imports=["SessionFromGraph"],
frm="polygraphy.backend.tf")
loader_name = script.add_loader(
Script.invoke("SessionFromGraph", graph_name, config=config_name),
"build_tf_session")
runner_str = Script.invoke("TfRunner",
loader_name,
timeline_path=args.save_timeline)
script.add_runner(runner_str)
def build_script(args, cmd_run=None):
script = Script(summary=generate_summary(args.model_file, args.runners,
args.load_results))
tool_util.add_logger_settings(script, args)
data_loader_name = tool_util.add_data_loader(script, args)
for runner_arg in args.runners:
add_runner_func = {
"tf":
add_tf_runner,
"onnxrt":
add_onnxrt_runner,
"onnxtf":
add_onnxtf_runner,
"cntk":
add_cntk_runner,
"trt":
lambda script, args: add_trt_runner(script, args, data_loader_name
),
"trt_legacy":
add_trt_legacy_runner,
}[runner_arg]
add_runner_func(script, args)
add_comparator(script,
args,
data_loader_name=data_loader_name,
cmd_run=cmd_run)
return str(script)
def get_data_loader(self):
script = Script()
data_loader_name = self.add_to_script(script)
if data_loader_name is None: # All arguments are default
from polygraphy.comparator import DataLoader
return DataLoader()
exec(str(script), globals(), locals())
return locals()[data_loader_name]
def add_to_script(self, script):
script.add_import(imports=["TfRunner"], frm="polygraphy.backend.tf")
graph_name = self.tf_loader_args.add_to_script(script)
config_name = self.tf_config_args.add_to_script(script)
script.add_import(imports=["SessionFromGraph"],
frm="polygraphy.backend.tf")
loader_name = script.add_loader(
Script.invoke("SessionFromGraph", graph_name, config=config_name),
"build_tf_session")
runner_name = script.add_loader(
Script.invoke("TfRunner",
loader_name,
timeline_path=self.timeline_path), "tf_runner")
script.add_runner(runner_name)
return runner_name
def _wrap_if_plugins(self, script, obj_name):
if self.plugins:
script.add_import(imports=["LoadPlugins"],
frm="polygraphy.backend.trt")
loader_str = Script.invoke("LoadPlugins",
obj_name,
plugins=self.plugins)
obj_name = script.add_loader(loader_str, "load_plugins")
return obj_name
def add_tf_config_loader(script, args):
config_loader_str = Script.invoke_if_nondefault("CreateConfig", gpu_memory_fraction=args_util.get(args, "gpu_memory_fraction"),
allow_growth=args_util.get(args, "allow_growth"), use_xla=args_util.get(args, "xla"))
if config_loader_str is not None:
script.add_import(imports=["CreateConfig"], frm="polygraphy.backend.tf")
config_loader_name = script.add_loader(config_loader_str, "create_tf_config")
else:
config_loader_name = None
return config_loader_name
def add_to_script(self, script):
config_loader_str = Script.invoke_if_nondefault("CreateConfig", gpu_memory_fraction=self.gpu_memory_fraction,
allow_growth=self.allow_growth, use_xla=self.xla)
if config_loader_str is not None:
script.add_import(imports=["CreateConfig"], frm="polygraphy.backend.tf")
config_loader_name = script.add_loader(config_loader_str, "create_tf_config")
else:
config_loader_name = None
return config_loader_name
def get_modify_onnx_str(script, args, loader_name, disable_outputs=None):
if disable_outputs:
outputs = None
exclude_outputs = None
else:
outputs = _get_outputs_arg(script, args, "onnx_outputs")
exclude_outputs = args_util.get(args, "onnx_exclude_outputs")
if hasattr(args, "shape_inference"):
do_shape_inference = args_util.get(args, "shape_inference")
else:
do_shape_inference = None if args_util.get(args, "no_shape_inference") else True
MODIFY_ONNX = "ModifyOnnx"
modify_onnx_str = Script.invoke(MODIFY_ONNX, loader_name, do_shape_inference=do_shape_inference,
outputs=outputs, exclude_outputs=exclude_outputs)
if modify_onnx_str != Script.invoke(MODIFY_ONNX, loader_name):
script.add_import(imports=[MODIFY_ONNX], frm="polygraphy.backend.onnx")
return modify_onnx_str
return None
def _get_modify_onnx_str(self, script, loader_name, disable_outputs=None):
if disable_outputs:
outputs = None
exclude_outputs = None
else:
outputs = tools_util.get_outputs_for_script(script, self.outputs)
exclude_outputs = self.exclude_outputs
MODIFY_ONNX = "ModifyOnnx"
modify_onnx_str = Script.invoke(
MODIFY_ONNX,
loader_name,
do_shape_inference=self.do_shape_inference,
outputs=outputs,
exclude_outputs=exclude_outputs)
if modify_onnx_str != Script.invoke(MODIFY_ONNX, loader_name):
script.add_import(imports=[MODIFY_ONNX],
frm="polygraphy.backend.onnx")
return modify_onnx_str
return None
def add_trt_network_loader(script, args):
model_file = args_util.get(args, "model_file")
outputs = _get_outputs_arg(script, args, "trt_outputs")
if args_util.get(args, "network_api"):
CREATE_NETWORK_FUNC = Inline("create_network")
script.add_import(imports=["CreateNetwork"], frm="polygraphy.backend.trt")
script.add_import(imports=["extend"], frm="polygraphy.common.func")
script.append_prefix("# Manual TensorRT network creation")
script.append_prefix("@extend(CreateNetwork())")
script.append_prefix("def {:}(builder, network):".format(CREATE_NETWORK_FUNC))
script.append_prefix("{tab}import tensorrt as trt\n".format(tab=constants.TAB))
script.append_prefix("{tab}# Define your network here. Make sure to mark outputs!".format(tab=constants.TAB))
net_inputs = args_util.get(args, "inputs")
if net_inputs:
for name, (dtype, shape) in net_inputs.items():
script.append_prefix("{tab}{name} = network.add_input(name='{name}', shape={shape}, dtype=trt.float32) # TODO: Set dtype".format(
name=name, shape=shape, tab=constants.TAB))
script.append_prefix("{tab}# TODO: network.mark_output(...)\n".format(tab=constants.TAB))
return CREATE_NETWORK_FUNC
if args_util.get(args, "ext"):
script.add_import(imports=["NetworkFromOnnxPath"], frm="polygraphy.backend.trt")
loader_str = Script.invoke("NetworkFromOnnxPath", _wrap_if_plugins(script, args, model_file), explicit_precision=args_util.get(args, "explicit_precision"))
loader_name = script.add_loader(loader_str, "parse_network_from_onnx")
else:
script.add_import(imports=["NetworkFromOnnxBytes"], frm="polygraphy.backend.trt")
onnx_loader = add_serialized_onnx_loader(script, args, disable_outputs=True)
loader_str = Script.invoke("NetworkFromOnnxBytes", _wrap_if_plugins(script, args, onnx_loader), explicit_precision=args_util.get(args, "explicit_precision"))
loader_name = script.add_loader(loader_str, "parse_network_from_onnx")
MODIFY_NETWORK = "ModifyNetwork"
modify_network_str = Script.invoke(MODIFY_NETWORK, loader_name, outputs=outputs, exclude_outputs=args_util.get(args, "trt_exclude_outputs"))
if modify_network_str != Script.invoke(MODIFY_NETWORK, loader_name):
script.add_import(imports=[MODIFY_NETWORK], frm="polygraphy.backend.trt")
loader_name = script.add_loader(modify_network_str, "modify_network")
return loader_name
def add_to_script(self, script):
script.add_import(imports=["TrtLegacyRunner"], frm="polygraphy.backend.trt_legacy")
G_LOGGER.warning("Legacy TensorRT runner only supports implicit batch TensorFlow/UFF, ONNX, and Caffe models")
if self.model_args.model_type == "onnx":
script.add_import(imports=["ParseNetworkFromOnnxLegacy"], frm="polygraphy.backend.trt_legacy")
onnx_loader = self.onnx_loader_args.add_onnx_loader(script, disable_outputs=True)
loader_name = script.add_loader(Script.format_str("ParseNetworkFromOnnxLegacy({:})", onnx_loader), "parse_network_from_onnx_legacy")
elif self.model_args.model_type == "caffe":
script.add_import(imports=["LoadNetworkFromCaffe"], frm="polygraphy.backend.trt_legacy")
loader_name = script.add_loader(Script.format_str("LoadNetworkFromCaffe({:}, {:}, {:}, {:})", self.model_args.model_file, self.caffe_model,
self.trt_outputs, self.batch_size), "parse_network_from_caffe")
else:
script.add_import(imports=["LoadNetworkFromUff"], frm="polygraphy.backend.trt_legacy")
if self.model_args.model_type == "uff":
script.add_import(imports=["LoadUffFile"], frm="polygraphy.backend.trt_legacy")
shapes = {name: shape for name, (_, shape) in self.trt_loader_args.input_shapes.items()}
loader_name = script.add_loader(Script.format_str("LoadUffFile({:}, {:}, {:})", self.model_args.model_file, misc.default_value(shapes, {}), self.trt_outputs), "load_uff_file")
else:
script.add_import(imports=["ConvertToUff"], frm="polygraphy.backend.trt_legacy")
loader_name = script.add_loader(Script.format_str("ConvertToUff({:}, save_uff={:}, preprocessor={:})", self.tf_loader_args.add_to_script(script), self.save_uff, self.preprocessor), "convert_to_uff")
loader_name = script.add_loader(Script.format_str("LoadNetworkFromUff({:}, uff_order={:})", loader_name, self.uff_order), "uff_network_loader")
runner_str = Script.format_str("TrtLegacyRunner({:}, {:}, {:}, fp16={:}, tf32={:}, load_engine={:}, save_engine={:}, layerwise={:}, plugins={:})",
loader_name, self.trt_loader_args.workspace, self.batch_size, self.trt_loader_args.fp16, self.trt_loader_args.tf32,
self.model_args.model_file if self.model_args.model_type == "engine" else None,
self.trt_runner_args.save_engine, self.trt_outputs==constants.MARK_ALL, self.trt_loader_args.plugins)
runner_name = script.add_loader(runner_str, "trt_legacy_runner")
script.add_runner(runner_name)
return runner_name
def add_data_loader(script, args):
def omit_none_tuple(tup):
if all([elem is None for elem in tup]):
return None
return tup
int_range = omit_none_tuple(tup=(args_util.get(args, "int_min"), args_util.get(args, "int_max")))
float_range = omit_none_tuple(tup=(args_util.get(args, "float_min"), args_util.get(args, "float_max")))
input_metadata_str = Inline(repr(args_util.get(args, "inputs"))) if args_util.get(args, "inputs") else None
if input_metadata_str:
script.add_import(imports=["TensorMetadata"], frm="polygraphy.common")
data_loader = Script.invoke_if_nondefault("DataLoader", seed=args_util.get(args, "seed"), iterations=args_util.get(args, "iterations"),
input_metadata=input_metadata_str, int_range=int_range, float_range=float_range)
if data_loader is not None:
data_loader_name = Inline("data_loader")
script.add_import(imports=["DataLoader"], frm="polygraphy.comparator")
script.append_prefix(Script.format_str("\n# Inference Inputs Loader\n{:} = {:}\n", data_loader_name, Inline(data_loader)))
else:
data_loader_name = None
return data_loader_name
def add_to_script(self, script, suffix=None):
G_LOGGER.verbose(
"Attempting to load as a TensorFlow model, using TF2ONNX to convert to ONNX. "
"If this is not correct, please specify --model-type",
mode=LogMode.ONCE)
script.add_import(imports=["OnnxFromTfGraph"],
frm="polygraphy.backend.onnx")
loader_str = Script.invoke("OnnxFromTfGraph",
self.tf_loader_args.add_to_script(
script,
disable_outputs=True,
suffix=suffix),
opset=self.opset,
fold_constant=self.fold_constant)
loader_name = script.add_loader(loader_str,
"export_onnx_from_tf",
suffix=suffix)
return loader_name
def add_to_script(self, script):
def _make_data_loader(script):
data_loader_name = Inline("data_loader")
input_metadata_str = Inline(
repr(self.model_args.input_shapes
)) if self.model_args.input_shapes else None
if input_metadata_str:
script.add_import(imports=["TensorMetadata"],
frm="polygraphy.common")
data_loader = Script.invoke_if_nondefault(
"DataLoader",
seed=self.seed,
iterations=self.iterations,
input_metadata=input_metadata_str,
int_range=self.int_range,
float_range=self.float_range)
if data_loader is not None:
script.add_import(imports=["DataLoader"],
frm="polygraphy.comparator")
script.append_prefix(
Script.format_str("\n# Inference Inputs Loader\n{:} = {:}",
data_loader_name, Inline(data_loader)))
else:
data_loader_name = None
return data_loader_name
if self.load_inputs:
script.add_import(imports=["misc"], frm="polygraphy.util")
data_loader_name = Inline("data_loader")
script.append_prefix(
Script.format_str(
"# Load inputs\n{data_loader} = []\nfor input_data_path in {load_inputs}:"
"\n{tab}{data_loader}.extend(misc.pickle_load(input_data_path))",
data_loader=data_loader_name,
load_inputs=self.load_inputs,
tab=Inline(constants.TAB)))
else:
data_loader_name = _make_data_loader(script)
script.append_prefix("") # Newline
return data_loader_name
def build_script(self, args):
script = Script(
summary=generate_summary(self.makers[ModelArgs].model_file,
args.runners, args.load_results))
self.makers[LoggerArgs].add_to_script(script)
data_loader_name = self.makers[DataLoaderArgs].add_to_script(script)
for runner_arg in args.runners:
add_runner_func = {
"tf":
self.makers[TfRunnerArgs].add_to_script,
"onnxrt":
self.makers[OnnxrtRunnerArgs].add_to_script,
"onnxtf":
self.makers[OnnxtfRunnerArgs].add_to_script,
"trt":
lambda script: self.makers[TrtRunnerArgs].add_to_script(
script, data_loader_name),
"trt_legacy":
self.makers[TrtLegacyArgs].add_to_script,
}[runner_arg]
add_runner_func(script)
RESULTS_VAR_NAME = self.makers[ComparatorRunArgs].add_to_script(
script, data_loader_name=data_loader_name)
SUCCESS_VAR_NAME = self.makers[ComparatorCompareArgs].add_to_script(
script, results_name=RESULTS_VAR_NAME)
cmd_run = Inline("' '.join(sys.argv)")
script.append_suffix(
Script.format_str(
'# Report Results\ncmd_run={cmd}\nif {success}:\n{tab}G_LOGGER.finish("PASSED | Command: {{}}".format(cmd_run))\nelse:\n{tab}G_LOGGER.error("FAILED | Command: {{}}".format(cmd_run))',
cmd=cmd_run,
success=SUCCESS_VAR_NAME,
tab=Inline(constants.TAB)))
script.append_suffix(
"sys.exit(0 if {success} else 1)".format(success=SUCCESS_VAR_NAME))
return str(script)