WindowedAdagrad

class viabel.WindowedAdagrad(learning_rate, *, weight_decay=0, window_size=10, jitter=1e-08, diagnostics=False)

Windowed Adagrad optimization method (Default optimizer in Pymc3)

Uses a running window (w) to get the mean squared gradient to rescale the current stochastic gradient:

\[\frac{\hat{g}^{(k+1)}}{\sqrt{\sum^k_{k^\prime = k-w} \hat{g}^{(k^\prime)} \cdot \hat{g}^{(k^\prime)}}}\]

Parameters:

window sizeint optional

Window size used to store the square of the gradients. The default is 10

jitter: `float` optional

Small value used for numerical stability. The default is 1e-8

Returns:

descent_dirnumpy.ndarray, shape(var_param_dim,)

Descent direction of the optimization algorithm

Methods

descent_direction(grad)	Compute descent direction for optimization.
optimize(n_iters, objective, init_param[, ...])	Parameters:
reset_state()	resetting the running squared gradients

__init__(learning_rate, *, weight_decay=0, window_size=10, jitter=1e-08, diagnostics=False)

Parameters:

learning_ratefloat

Tuning parameter that determines the step size

weight_decay: `float`

L2 regularization weight

iterate_avg_propfloat

Proportion of iterates to use for computing iterate average. None means no iterate averaging. The default is 0.2.

diagnosticsbool, optional

Record diagnostic information if True. The default is False.

descent_direction(grad)

Compute descent direction for optimization.

Default implementation returns grad.

Parameters:

gradnumpy.ndarray, shape(var_param_dim,)

(stochastic) gradient of the objective function

Returns:

descent_dirnumpy.ndarray, shape(var_param_dim,)

Descent direction of the optimization algorithm

reset_state()

resetting the running squared gradients