def build(self, hp: HyperParameters, *args: Any,
**kwargs: Any) -> keras.Model:
for i in range(self._max_fail_streak + 1):
# clean-up TF graph from previously stored (defunct) graph
keras.backend.clear_session()
gc.collect()
# Build a model, allowing max_fail_streak failed attempts.
try:
with maybe_distribute(self.distribution_strategy):
# /!\ Below line is the only one changed compared to the original version: /!\
# model = self.hypermodel.build(hp)
model = self.hypermodel.build(hp, *args, **kwargs)
except: # noqa: E722 do not use bare 'except'
if keras_tuner_config_module.DEBUG:
traceback.print_exc()
print("Invalid model %s/%s" % (i, self._max_fail_streak))
if i == self._max_fail_streak:
raise RuntimeError(
"Too many failed attempts to build model.")
continue
# Stop if `build()` does not return a valid model.
if not isinstance(model, keras.models.Model):
raise RuntimeError(
"Model-building function did not return "
"a valid Keras Model instance, found {}".format(model))
# Check model size.
size = maybe_compute_model_size(model)
if self.max_model_size and size > self.max_model_size:
print("Oversized model: %s parameters -- skipping" % (size))
if i == self._max_fail_streak:
raise RuntimeError(
"Too many consecutive oversized models.")
continue
break
return self._compile_model(model)
def get_best_model(self):
with hm_module.maybe_distribute(self.distribution_strategy):
model = tf.keras.models.load_model(self.best_model_path)
return model
def get_best_model(self):
model = self._build_best_model()
with hm_module.maybe_distribute(self.distribution_strategy):
model.load_weights(self.best_model_path)
return model