Training Result Check

The key steps are as follows:

1. View the training results.
   • To export the sparse table data in .npy format, call the export API.
   • To export the .pb model file, call the export_saved_model API of Estimator. The following is an example:

     import os
     import tensorflow as tf
     if tf.__version__.startswith("1"):
         from npu_bridge.npu_init import NPURunConfig, NPUEstimator
     else:
         from npu_device.compat.v1.npu_init import NPURunConfig, NPUEstimator
     # For details, see "Running Configuration" and "Creating an Estimator Object" in Porting with Estimator.
     run_config = NPURunConfig(...)
     est = NPUEstimator(...)
     # Generally, the export_saved_model API is called after the train or train_and_evaluate call.
     def _serving_input_fn():
         # Adjust the input based on the specific service model. The following uses the input of the little demo estimator model as an example.
         inputs = {
             "user_ids": tf.compat.v1.placeholder(shape=(None, 32), dtype=tf.int64, name="user_ids"),
             "item_ids": tf.compat.v1.placeholder(shape=(None, 8), dtype=tf.int64, name="item_ids"),
             "label_0": tf.compat.v1.placeholder(shape=(None,), dtype=tf.float32, name="label_0"),
             "label_1": tf.compat.v1.placeholder(shape=(None,), dtype=tf.float32, name="label_1"),
         }
         return tf.estimator.export.ServingInputReceiver(features=inputs, receiver_tensors=inputs)
     target_pb_path = os.path.abspath("pb_model_path")
     # Call the export_saved_model API of Estimator to save the model in .pb format.
     export_path = est.export_saved_model(target_pb_path, _serving_input_fn).decode("utf-8")
     print(f"The export saved model path is {export_path}.")

2. Call the terminate_config_initializer API to disable the data flow and release resources.