def generate_random_feeder(self,
config,
use_feed_fetch=True,
feeder_adapter=None):
if config is None or not isinstance(config, api_param.APIConfig):
raise ValueError(
"Argument \"config\" must be set to an instance of APIConfig.")
if feeder_adapter is not None and feeder_adapter.framework != "tensorflow":
assert use_feed_fetch, "Argument use_feed_fetch must be True when feeder_adapter is initialized by paddle."
if feeder_adapter is None or feeder_adapter.framework != "tensorflow":
self._need_feed = config.name == "feed"
self._need_fetch = use_feed_fetch or config.name == "fetch"
self._feed_spec = feeder.copy_feed_spec(config.feed_spec)
self._feed_dict = {}
self.__backward = False
self.build_graph(config=config)
if feeder_adapter is None:
feed_list = []
assert len(self._feed_dict) == len(self.feed_list)
for var in self.feed_list:
feed_list.append(self._feed_dict[var])
return feeder.FeederAdapter("tensorflow", config.feed_spec,
feed_list)
else:
return feeder_adapter
def generate_random_feeder(self,
config,
use_feed_fetch=True,
feeder_adapter=None):
if config is None or not isinstance(config, api_param.APIConfig):
raise ValueError(
"Argument \"config\" must be set to an instance of APIConfig.")
if feeder_adapter is None or feeder_adapter.framework != "paddle":
self._need_feed = config.name == "feed"
self._need_fetch = use_feed_fetch or config.name == "fetch"
self._feed_spec = feeder.copy_feed_spec(config.feed_spec)
self._feed_dict = {}
self.__backward = False
self.main_program = fluid.Program()
self.startup_program = fluid.Program()
with fluid.program_guard(self.main_program, self.startup_program):
self.build_program(config=config)
if feeder_adapter is None:
feed_list = []
for var in self.feed_vars:
feed_list.append(self._feed_dict[var])
return feeder.FeederAdapter("paddle", config.feed_spec, feed_list)
else:
return feeder_adapter
def run(self, config, args):
self.name = config.api_name
self._reset()
self._feed_spec = feeder.copy_feed_spec(config.feed_spec)
self._need_fetch = args.task == "accuracy"
if args.use_gpu and torch.cuda.is_available():
self._device = torch.device("cuda")
else:
self._device = torch.device("cpu")
outputs, stats = self.run_impl(use_gpu=args.use_gpu,
config=config,
repeat=args.repeat,
check_output=args.check_output,
profiler=args.profiler)
return outputs, stats
def run(self, config, args, feeder_adapter=None):
paddle.disable_static()
self.name = config.api_name
self._reset()
self._feed_spec = feeder.copy_feed_spec(config.feed_spec)
self._need_fetch = args.task == "accuracy"
if feeder_adapter:
self._feed_values = feeder_adapter.to_paddle()
outputs, stats = self.run_impl(use_gpu=args.use_gpu,
config=config,
repeat=args.repeat,
check_output=args.check_output,
profiler=args.profiler,
feeder_adapter=feeder_adapter)
return outputs, stats
def generate_random_feeder(self,
config,
use_feed_fetch=True,
feeder_adapter=None):
if config is None or not isinstance(config, api_param.APIConfig):
raise ValueError(
"Argument \"config\" must be set to an instance of APIConfig.")
if feeder_adapter is None or feeder_adapter.framework != "paddle":
self._need_feed = config.name == "feed"
self._need_fetch = use_feed_fetch or config.name == "fetch"
self._feed_spec = feeder.copy_feed_spec(config.feed_spec)
self._feed_dict = {}
self.__backward = False
self.main_program = fluid.Program()
self.startup_program = fluid.Program()
with fluid.program_guard(self.main_program, self.startup_program):
self.build_program(config=config)
# For backward benchmark, the program is composed of:
# xxx -> shape -> fill_constant -> xxx_grad
# The extra CUDA kernel of fill_constant will make the traced times
# larger than the actual, but tf can automatic optimize the execution
# of fill_constant. We call self._prune() to move the fill_constant op
# from main_program to startup_program for current benchmark and will
# optimize the execution strategy in the future.
self._prune(config)
if feeder_adapter is None:
feed_list = []
for var in self.feed_vars:
feed_list.append(self._feed_dict[var])
return feeder.FeederAdapter("paddle", config.feed_spec, feed_list)
else:
return feeder_adapter