Overview

Applicability

Product

Supported

Atlas A3 training products/Atlas A3 inference products

Atlas A2 training products/Atlas A2 inference products

Atlas 200I/500 A2 inference products

Atlas inference products

Atlas training products

Function Description

When performing model training or online inference in Estimator mode on the Ascend AI Processor, you can use the constructor of the NPURunConfig class to specify the running configuration of the Estimator.

The NPURunConfig class inherits the RunConfig class of tf.estimator. For details about the support for native APIs of the RunConfig class, see Supported RunConfig Parameters.

Function Prototype

You can view the NPURunConfig prototype definition in the python/site-packages/npu_bridge/estimator/npu/npu_config.py file in the TensorFlow Adapter installation directory. The following is an example:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
class NPURunConfig(run_config_lib.RunConfig):
    def __init__(self,
                 iterations_per_loop=1,
                 profiling_config=None,
                 model_dir=None,
                 tf_random_seed=None,
                 save_summary_steps=0,
                 save_checkpoints_steps=None,
                 save_checkpoints_secs=None,
                 ...
                 )

For details about the parameters supported by NPURunConfig, see the parameter description in the following sections.

Constraints

In multi-device training scenarios, the save_checkpoints_secs parameter cannot be used to save files by time.

Returns

An object of the NPURunConfig class, as the initialization argument passed to the NPUEstimator call.

Example

The usage of NPURunConfig configurations is as follows:

1
2
3
4
5
6
from npu_bridge.npu_init import *
session_config=tf.ConfigProto()
config = NPURunConfig(
    session_config=session_config, 
    mix_compile_mode=False, 
    iterations_per_loop=1000)