def map_structure(s):
if isinstance(s, int):
return flt[s]
elif isinstance(s, list):
return [map_structure(item) for item in s]
elif isinstance(s, tuple):
return tuple([map_structure(item) for item in s])
elif isinstance(s, dict):
return {key: map_structure(s[key]) for key in sorted(s.keys())}
else:
raise TypeError
def map_structure(s):
if isinstance(s, np.ndarray):
flt.append(s)
return len(flt) - 1
elif isinstance(s, list):
return [map_structure(item) for item in s]
elif isinstance(s, tuple):
return tuple([map_structure(item) for item in s])
elif isinstance(s, dict):
return {key: map_structure(s[key]) for key in sorted(s.keys())}
else:
raise TypeError
def export(self, exe, program, eval_model_spec, eval_result, state):
"""doc"""
log.debug('New evaluate result: %s \nold: %s' %
(repr(eval_result), repr(self._best)))
if self._best is None and state['best_model'] is not None:
self._best = state['best_model']
log.debug('restoring best state %s' % repr(self._best))
self._best_result = eval_result
if self._best is None or self.cmp_fn(old=self._best, new=eval_result):
if self.is_export:
log.debug('[Best Exporter]: export to %s' % self._export_dir)
eval_program = program.train_program
# FIXME: all eval datasets has same name/types/shapes now!!! so every eval program are the smae
saver = Saver(
self._export_dir, exe, program=program, max_ckpt_to_keep=1)
saver.save(state)
eval_result = map_structure(float, eval_result)
self._best = eval_result
state['best_model'] = eval_result
self._best_result = eval_result
else:
log.debug('[Best Exporter]: skip step %s' % state.gstep)
self._epoch_count += 1
if self._epoch_count % 10 == 0:
log.info("EPOCH_%s: %s" % (self._epoch_count, repr(self._best_result)))
def export(self, exe, program, eval_model_spec, eval_result, state):
"""doc"""
if self.model_class_or_model_fn is not None and self.hparams is not None \
and self.dataset is not None:
log.info('Building program by user defined model function')
if issubclass(self.model_class_or_model_fn, Model):
_model_fn = _build_model_fn(self.model_class_or_model_fn)
elif inspect.isfunction(self.model_class_or_model_fn):
_model_fn = self.model_class_or_model_fn
else:
raise ValueError('unknown model %s' %
self.model_class_or_model_fn)
# build net
infer_program = F.Program()
startup_prog = F.Program()
with F.program_guard(infer_program, startup_prog):
#share var with Train net
with F.unique_name.guard():
log.info('Building Infer Graph')
infer_fea = self.dataset.features()
# run_config is None
self.model_spec = _build_net(_model_fn, infer_fea,
RunMode.PREDICT, self.hparams,
None)
log.info('Done')
infer_program = infer_program.clone(for_test=True)
self.program = ProgramPair(train_program=infer_program,
startup_program=startup_prog)
else:
self.program = program
self.model_spec = eval_model_spec
if self._best is None and state['best_inf_model'] is not None:
self._best = state['best_inf_model']
log.debug('restoring best state %s' % repr(self._best))
log.debug('New evaluate result: %s \nold: %s' %
(repr(eval_result), repr(self._best)))
if self._best is None or self.cmp_fn(old=self._best, new=eval_result):
log.debug('[Best Exporter]: export to %s' % self._export_dir)
if self.model_spec.inference_spec is None:
raise ValueError('model_fn didnt return InferenceSpec')
inf_spec_dict = self.model_spec.inference_spec
if not isinstance(inf_spec_dict, dict):
inf_spec_dict = {'inference': inf_spec_dict}
for inf_spec_name, inf_spec in six.iteritems(inf_spec_dict):
if not isinstance(inf_spec, InferenceSpec):
raise ValueError('unknow inference spec type: %s' %
inf_spec)
save_dir = os.path.join(self._export_dir, inf_spec_name)
log.debug('[Best Exporter]: save inference model: "%s" to %s' %
(inf_spec_name, save_dir))
feed_var = [i.name for i in inf_spec.inputs]
fetch_var = inf_spec.outputs
infer_program = self.program.train_program
startup_prog = F.Program()
F.io.save_inference_model(save_dir,
feed_var,
fetch_var,
exe,
main_program=infer_program)
eval_result = map_structure(float, eval_result)
state['best_inf_model'] = eval_result
self._best = eval_result
else:
log.debug('[Best Exporter]: skip step %s' % state.gstep)
