TensorFlow Serving 记录
2021-04-20
1 min read
模型保存
我们主要使用 Estimator API 进行模型训练，在模型保存接口中
def export_saved_model(self,
export_dir_base,
serving_input_receiver_fn,
assets_extra=None,
as_text=False,
checkpoint_path=None,
experimental_mode=ModeKeys.PREDICT):
if not serving_input_receiver_fn:
raise ValueError('An input_receiver_fn must be defined.')
input_receiver_fn_map = {experimental_mode: serving_input_receiver_fn}
return self._export_all_saved_models(
export_dir_base,
input_receiver_fn_map,
assets_extra=assets_extra,
as_text=as_text,
checkpoint_path=checkpoint_path,
strip_default_attrs=True)
有一个关键参数 serving_input_receiver_fn，用来指定将来使用 TensorFlow Serving 时 输入的Tensor 类型