Averager¶
Averaged stochastic gradient descent.
This is a wrapper that can be applied to any stochastic gradient descent optimiser. Note that this implementation differs than what may be found elsewhere. Essentially, the average of the weights is usually only used at the end of the optimisation, once all the data has been seen. However, in this implementation the optimiser returns the current averaged weights.
Parameters¶
-
optimizer ('optim.base.Optimizer')
An optimizer for which the produced weights will be averaged.
-
start ('int') β defaults to
0
Indicates the number of iterations to wait before starting the average. Essentially, nothing happens differently before the number of iterations reaches this value.
Attributes¶
- learning_rate
Examples¶
>>> from river import datasets
>>> from river import evaluate
>>> from river import linear_model
>>> from river import metrics
>>> from river import optim
>>> from river import preprocessing
>>> dataset = datasets.Phishing()
>>> optimizer = optim.Averager(optim.SGD(0.01), 100)
>>> model = (
... preprocessing.StandardScaler() |
... linear_model.LogisticRegression(optimizer)
... )
>>> metric = metrics.F1()
>>> evaluate.progressive_val_score(dataset, model, metric)
F1: 87.89%
Methods¶
look_ahead
Updates a weight vector before a prediction is made.
Parameters: w (dict): A dictionary of weight parameters. The weights are modified in-place. Returns: The updated weights.
Parameters
- w ('dict')
step
Updates a weight vector given a gradient.
Parameters
- w ('dict | VectorLike')
- g ('dict | VectorLike')
Returns
dict | VectorLike: The updated weights.
References¶
-
Bottou, L., 2010. Large-scale machine learning with stochastic gradient descent. In Proceedings of COMPSTAT'2010 (pp. 177-186). Physica-Verlag HD. ↩
-
Stochastic Algorithms for One-Pass Learning slides by LΓ©on Bottou ↩
-
Xu, W., 2011. Towards optimal one pass large scale learning with averaged stochastic gradient descent. arXiv preprint arXiv:1107.2490. ↩