def _broadcast(self, epoch=None):
record = Report().receive(self.trainer.step_name, self.trainer.worker_id)
if self.trainer.config.report_on_epoch:
record.epoch = self.trainer.epochs
# todo: remove in FinedGrainedSpace
if self.trainer.config.codec:
record.desc = self.trainer.config.codec
if not record.desc:
record.desc = self.trainer.model_desc
record.performance = self.trainer.performance
record.objectives = self.trainer.valid_metrics.objectives
if record.performance is not None:
for key in record.performance:
if key not in record.objectives:
if (key == 'flops' or key == 'params'):
record.objectives.update({key: 'MIN'})
else:
record.objectives.update({key: 'MAX'})
record.model_path = self.trainer.model_path
record.checkpoint_path = self.trainer.checkpoint_file
record.weights_file = self.trainer.weights_file
if self.trainer.runtime is not None:
record.runtime = self.trainer.runtime
Report().broadcast(record)
logging.debug("report_callback record: {}".format(record))
def _broadcast(self, pfms):
"""Boadcase pfrm to record."""
record = Report().receive(self.step_name, self.worker_id)
if record.performance:
record.performance.update(pfms)
else:
record.performance = pfms
Report().broadcast(record)
logging.info("valid record: {}".format(record))
def _save_best(self, desc):
record = Report().receive(self.step_name, self.sample_count + 1)
record.performance = {"accuracy": 100}
record.desc = desc
Report().broadcast(record)
def _broadcast(self, epoch, performance):
record = Report().receive(self.trainer.step_name, self.trainer.worker_id)
record.performance = performance
Report().broadcast(record)
logging.debug("report_callback record: {}".format(record))
