edflow.tf_util module¶

Summary¶

Functions:

`make_exponential_var`	Exponential from \((a, \alpha)\) to \((b, \beta)\) with decay rate decay.
`make_linear_var`	Linear from \((a, \alpha)\) to \((b, \beta)\), i.e.
`make_periodic_step`	Returns step within the unit period cycle specified
`make_periodic_wrapper`	A wrapper to wrap the step variable of a step function into a periodic step variable.
`make_staircase_var`	param step \(x\)
`make_var`	Example

Reference¶

edflow.tf_util.make_linear_var(step, start, end, start_value, end_value, clip_min=None, clip_max=None, **kwargs)[source]¶

Linear from \((a, \alpha)\) to \((b, \beta)\), i.e. \(y = (\beta - \alpha)/(b - a) * (x - a) + \alpha\)

Parameters

step (tf.Tensor) – \(x\)
start (int) – \(a\)
end (int) – \(b\)
start_value (float) – \(\alpha\)
end_value (float) – \(\beta\)
clip_min (int) – Minimal value returned.
clip_max (int) – Maximum value returned.

Returns

:math:`y`

Return type

tf.Tensor

edflow.tf_util.make_periodic_step(step, start_step: int, period_duration_in_steps: int, **kwargs)[source]¶

Returns step within the unit period cycle specified

Parameters

step (tf.Tensor) – step variable
start_step (int) – an offset parameter specifying when the first period begins
period_duration_in_steps (int) – period duration of step

Returns

step within unit cycle period

Return type

unit_step

edflow.tf_util.make_exponential_var(step, start, end, start_value, end_value, decay, **kwargs)[source]¶

Exponential from \((a, \alpha)\) to \((b, \beta)\) with decay rate decay.

Parameters

step (tf.Tensor) – \(x\)
start (int) – \(a\)
end (int) – \(b\)
start_value (float) – \(\alpha\)
end_value (float) – \(\beta\)
decay (int) – Decay rate

Returns

:math:`y`

Return type

tf.Tensor

edflow.tf_util.make_staircase_var(step, start, start_value, step_size, stair_factor, clip_min=0.0, clip_max=1.0, **kwargs)[source]¶

Parameters

step (tf.Tensor) – \(x\)
start (int) – \(a\)
start_value (float) – \(\alpha\)
step_size (int) – after how many steps the value should be changed
stair_factor (float) – factor that the value is multiplied with at every ‘step_size’ steps
clip_min (int) – Minimal value returned.
clip_max (int) – Maximum value returned.

Returns

:math:`y`

Return type

tf.Tensor

edflow.tf_util.make_periodic_wrapper(step_function)[source]¶

A wrapper to wrap the step variable of a step function into a periodic step variable. :param step_function: the step function where to exchange the step variable with a periodic step variable :type step_function: callable

Returns
Return type: a function with periodic steps

edflow.tf_util.make_var(step, var_type, options)[source]¶

Example

usage within trainer

grad_weight = make_var(step=self.global_step,
var_type=self.config["grad_weight"]["var_type"],
options=self.config["grad_weight"]["options"])

within yaml file

grad_weight:
  var_type: linear
  options:
    start:      50000
    end:        60000
    start_value:  0.0
    end_value: 1.0
    clip_min: 1.0e-6
    clip_max: 1.0

Parameters

step (tf.Tensor) – scalar tensor variable
var_type (str) – a string from [“linear”, “exponential”, “staircase”]
options (dict) – keyword arguments passed to specific ‘make_xxx_var’ function

Returns

:math:`y`

Return type

tf.Tensor