Class Reference

FeatureSpec

Configuration description class for features to be queried, which applies to non-automatic graph modification mode.

Parameter	Type	Mandatory/Optional	Description
index_key	int/string	Optional	Index key. Default value: value of table_name. Value range: [0, 255] for an int value 1 to 255 characters for a string value
table_name	string	Optional.	Table name. The value can contain 1 to 255 characters.
access_threshold	int	Optional	Feature admission threshold. Value range: [-1, 2147483647]. 0: Enables the admission function. The occurrence count of keys in new batches is not accumulated. Historical feature counts are used. Greater than 0: Enables the admission function. The occurrence count of keys in new batches is accumulated, and feature count records are updated. -1: Disables the function.
eviction_threshold	int	Optional	Feature eviction threshold. Value range: [-1, 2147483647]. Greater than or equal to 0: Enables the eviction function. -1: Disables the eviction function. If you need to set a feature eviction threshold, you must also set the feature admission threshold.
is_timestamp	bool	Optional	Specifies whether it is a timestamp. Value range: True or False.
batch_size	int	Optional	Dataset batch size. Value range: [1, 2147483647].
faae_coefficient	int	Optional	Feature admission coefficient. Default value: 1. Value range: [1, 2147483647].
name	string	Mandatory	FeatureSpec name. The value can contain 1 to 255 characters.

Example

from mx_rec.core.asc.feature_spec import FeatureSpec
feature_spec_list = FeatureSpec("user_ids", table_name="user_table",
                                access_threshold=1,
                                eviction_threshold=1,
                                faae_coefficient=1)

GraphModifierHook

Automatic graph modification hook class, used only in Training with Estimator mode. The automatic graph modification feature is enabled after this hook is added.

Parameter	Type	Mandatory/Optional	Description
dump_graph	bool	Optional	Specifies whether to save the current TensorFlow computational graph. Default value: False.
modify_graph	bool	Optional	Specifies whether to enable automatic graph modification. Default value: True.

Example

from mx_rec.graph.modifier import GraphModifierHook

#Define the data processing function.
def input_fn():
     pass

est.train(input_fn=lambda: input_fn(), hooks=[GraphModifierHook()])   # est is the created NPUEstimator object.

EvictHook

Feature eviction hook class, used only in feature admission and eviction mode. It works with the feature eviction threshold eviction_threshold. The feature eviction function is enabled after this hook is added.

Parameter	Type	Mandatory/Optional	Description
evict_enable	bool	Optional	Specifies whether to enable feature eviction. Default value: False.
evict_time_interval	int	Optional	Interval for triggering the eviction function, in seconds. The default value is 24 * 60 * 60. Value range: [1, MAXINT32].
evict_step_interval	int	Optional	Interval for triggering the eviction function, in steps. Default value: None. Value range: [1, MAXINT32].

Example

from mx_rec.core.feature_process import EvictHook
hooks_list = []
hook_evict = EvictHook(evict_enable=True, evict_time_interval=30, evict_step_interval=20)
hooks_list.append(hook_evict)

#Define the data processing function.
def input_fn():
     pass

est.train(input_fn=lambda: input_fn(), hooks=hooks_list) # est is the created NPUEstimator object.

ConfigInitializer

Management class for saving global configuration information, which uses the singleton pattern.

This class is automatically initialized via the init() function and does not require manual construction. This section lists only the public interfaces of this class. Other interfaces not mentioned here are internal and should not be called directly.

Call Example

Interface	Purpose	Prototype
get_instance()	Gets the unique global instance of ConfigInitializer.	from mx_rec.util.initialize import ConfigInitializerConfigInitializer.get_instance()
use_dynamic_expansion()	See use_dynamic_expansion.
get_target_batch()	See get_target_batch.
if_load()	See if_load.
get_initializer(is_training)	See get_initializer.
ascend_global_hashtable_collection()	See ascend_global_hashtable_collection.

TrainParamsConfig

Data class for saving training task parameter configurations, such as the name of the hash table collection.

This class is automatically initialized via the init() function and does not require manual construction. This section lists only the public interfaces of this class. Other interfaces not mentioned here are internal and should not be called directly.

Call Example

Interface	Purpose
ascend_global_hashtable_collection()	See ascend_global_hashtable_collection.