Bases: BayesianNeuralFieldEstimator
Fits models using stochastic ensembles of maximum-a-posteriori estimates.
Implementation of
BayesianNeuralFieldEstimator.
Source code in bayesnf/spatiotemporal.py
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541 | class BayesianNeuralFieldMAP(BayesianNeuralFieldEstimator):
"""Fits models using stochastic ensembles of maximum-a-posteriori estimates.
Implementation of
[BayesianNeuralFieldEstimator](BayesianNeuralFieldEstimator.md).
"""
_ensemble_dims = 2
def fit(
self,
table,
seed,
ensemble_size=16,
learning_rate=0.005,
num_epochs=5_000,
batch_size=None,
num_splits=1,
) -> BayesianNeuralFieldEstimator:
"""Run inference using stochastic MAP ensembles.
Args:
table (pandas.DataFrame):
See documentation of
[`table`][bayesnf.spatiotemporal.BayesianNeuralFieldEstimator.fit]
in the base class.
seed (jax.random.PRNGKey): The jax random key.
ensemble_size (int): Number of particles in the ensemble. It currently
an error if `ensemble_size < jax.device_count`, but will be fixed
in https://github.com/google/bayesnf/issues/28.
learning_rate (float): Learning rate for SGD.
num_epochs (int): Number of full epochs through the training data.
batch_size (None | int): Batch size for SGD. Default is `None`,
meaning full-batch. Each epoch will perform `len(table) // batch_size`
SGD updates.
num_splits (int): Number of splits over the data to run training.
Defaults to 1, meaning there are no splits.
Returns:
Instance of `self`.
"""
if ensemble_size < jax.device_count():
raise ValueError('ensemble_size cannot be smaller than device_count. '
'https://github.com/google/bayesnf/issues/28.')
train_data = self.data_handler.get_train(table)
train_target = self.data_handler.get_target(table)
if batch_size is None:
batch_size = train_data.shape[0]
if self._scale_epochs_by_batch_size:
num_epochs = num_epochs * (train_data.shape[0] // batch_size)
model_args = self._model_args((batch_size, train_data.shape[-1]))
self.params_, self.losses_ = inference.fit_map(
train_data,
train_target,
seed=seed,
observation_model=self.observation_model,
model_args=model_args,
num_particles=ensemble_size,
learning_rate=learning_rate,
num_epochs=num_epochs,
prior_weight=self._prior_weight,
batch_size=batch_size,
num_splits=num_splits)
return self
|
fit
fit(table, seed, ensemble_size=16, learning_rate=0.005, num_epochs=5000, batch_size=None, num_splits=1)
Run inference using stochastic MAP ensembles.
PARAMETER |
DESCRIPTION |
table |
See documentation of
table
in the base class.
TYPE:
DataFrame
|
seed |
TYPE:
PRNGKey
|
ensemble_size |
Number of particles in the ensemble. It currently
an error if ensemble_size < jax.device_count , but will be fixed
in https://github.com/google/bayesnf/issues/28.
TYPE:
int
DEFAULT:
16
|
learning_rate |
TYPE:
float
DEFAULT:
0.005
|
num_epochs |
Number of full epochs through the training data.
TYPE:
int
DEFAULT:
5000
|
batch_size |
Batch size for SGD. Default is None ,
meaning full-batch. Each epoch will perform len(table) // batch_size
SGD updates.
TYPE:
None | int
DEFAULT:
None
|
num_splits |
Number of splits over the data to run training.
Defaults to 1, meaning there are no splits.
TYPE:
int
DEFAULT:
1
|
Source code in bayesnf/spatiotemporal.py
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541 | def fit(
self,
table,
seed,
ensemble_size=16,
learning_rate=0.005,
num_epochs=5_000,
batch_size=None,
num_splits=1,
) -> BayesianNeuralFieldEstimator:
"""Run inference using stochastic MAP ensembles.
Args:
table (pandas.DataFrame):
See documentation of
[`table`][bayesnf.spatiotemporal.BayesianNeuralFieldEstimator.fit]
in the base class.
seed (jax.random.PRNGKey): The jax random key.
ensemble_size (int): Number of particles in the ensemble. It currently
an error if `ensemble_size < jax.device_count`, but will be fixed
in https://github.com/google/bayesnf/issues/28.
learning_rate (float): Learning rate for SGD.
num_epochs (int): Number of full epochs through the training data.
batch_size (None | int): Batch size for SGD. Default is `None`,
meaning full-batch. Each epoch will perform `len(table) // batch_size`
SGD updates.
num_splits (int): Number of splits over the data to run training.
Defaults to 1, meaning there are no splits.
Returns:
Instance of `self`.
"""
if ensemble_size < jax.device_count():
raise ValueError('ensemble_size cannot be smaller than device_count. '
'https://github.com/google/bayesnf/issues/28.')
train_data = self.data_handler.get_train(table)
train_target = self.data_handler.get_target(table)
if batch_size is None:
batch_size = train_data.shape[0]
if self._scale_epochs_by_batch_size:
num_epochs = num_epochs * (train_data.shape[0] // batch_size)
model_args = self._model_args((batch_size, train_data.shape[-1]))
self.params_, self.losses_ = inference.fit_map(
train_data,
train_target,
seed=seed,
observation_model=self.observation_model,
model_args=model_args,
num_particles=ensemble_size,
learning_rate=learning_rate,
num_epochs=num_epochs,
prior_weight=self._prior_weight,
batch_size=batch_size,
num_splits=num_splits)
return self
|