interface UpdatesOptions
These options automates and hides complexity of learning rates parameter tuning.
Notice that by default adaptive, invariant and normalized are turned on and to remove them you need to pass sgd parameter.
Adaptive option tunes learning rates individually per feature. This option is highly recommended - if you do not specify it, then rare features will have too small weights.
Invariant option allows consider example's weight during online learning. Without this option example's weight will be ignored. Also this option usually improves model quality.
Normalize option turns on online feature scaling. In a lot of cases you want to remove this option, specially in case of very sparse boolean feature space.
Always play with these options first, it usually gives the biggest gain.
Modifier and Type | Method and Description |
---|---|
SGDVowpalWabbitBuilder |
adaptive()
use adaptive, individual learning rates.
|
SGDVowpalWabbitBuilder |
invariant()
use safe/importance aware updates.
|
SGDVowpalWabbitBuilder |
learningRate(double learningRate)
Set learning rate
|
SGDVowpalWabbitBuilder |
normalized()
use per feature normalized updates
|
SGDVowpalWabbitBuilder |
sgd()
use regular stochastic gradient descent update.
|
SGDVowpalWabbitBuilder sgd()
removes adaptive, invariant and normalized options
SGDVowpalWabbitBuilder adaptive()
SGDVowpalWabbitBuilder invariant()
SGDVowpalWabbitBuilder normalized()
SGDVowpalWabbitBuilder learningRate(double learningRate)
learningRate
- learning rate. Must be positiveCopyright © 2017. All rights reserved.