Ftrl

Customizes the Ftrl optimizer.

Prototype

1
def create_hash_optimizer(learning_rate, use_locking=False, name="Ftrl", **kwargs)

Parameters

Parameter

Type

Mandatory/Optional

Description

learning_rate

float/tf.Tensor

Mandatory

Learning rate

Value range: [0.0, 10.0]

use_locking

bool

Optional

Prevents concurrent updates of variables.

Default value: False

Value range:

  • True
  • False

name

string

Optional

Name of the optimizer.

Default value: Ftrl

Name length range: [1, 200]

**kwargs Parameters

Parameter

Type

Mandatory/Optional

Description

learning_rate_power

float

Optional

Controls the decrease of learning rate during training.

Default value: –0.5

Value range: [–2147483647.0, 0.0]

initial_accumulator_value

float

Optional

Initial value of the accumulator

Default value: 0.1

Value range: (0.0, 1.0]

l1_regularization_strength

float

Optional

L1 regularization penalty

Default value: 0.0

Value range: [0.0, 10000.0]

l2_regularization_strength

float

Optional

L2 regularization penalty

Default value: 0.0

Value range: [0.0, 10000.0]

accum_name

string

Optional

Suffix of the variable of the gradient square accumulator

Default value: None

Value range: [1, 255]

linear_name

string

Optional

Suffix of the variable of the linear gradient accumulator

Default value: None

Value range: [1, 255]

l2_shrinkage_regularization_strength

float

Optional

L2 shrinkage penalty

Default value: 0.0

Value range: [0.0, 10000.0]

If kwargs is used to pass other parameters that are not described, Rec SDK TensorFlow does not use these parameters.

Return Value

An instance object of the CustomizedFtrl optimizer.

Example

1
2
from mx_rec.optimizers.ftrl import create_hash_optimizer
hashtable_optimizer = create_hash_optimizer(0.001)