Optimizers (recommendation.optimizers)

The classes presented in this section are optimizers to modify the SGD updates during the training of a model.

The update functions control the learning rate during the SGD optimization

SGD Stochastic Gradient Descent (SGD) updates
Momentum Stochastic Gradient Descent (SGD) updates with momentum
NesterovMomentum Stochastic Gradient Descent (SGD) updates with Nesterov momentum
AdaGrad AdaGrad updates
RMSProp Scale learning rates by dividing with the moving average of the root mean squared (RMS) gradients.
AdaDelta Scale learning rates by a the ratio of accumulated gradients to accumulated step sizes, see [4] and notes for further description.
Adam Adam updates implemented as in [5].
Adamax Adamax updates implemented as in [6].

Stochastic Gradient Descent

This is the optimizer by default in all models.

class orangecontrib.recommendation.optimizers.SGD(learning_rate=1.0)[source]

Stochastic Gradient Descent (SGD) updates

Generates update expressions of the form:

  • param := param - learning_rate * gradient
Args:
learning_rate: float, optional
The learning rate controlling the size of update steps
update(grads, params, indices=None)[source]

SGD updates

Args:
grads: array
List of gradient expressions
params: array
The variables to generate update expressions for
indices: array, optional
Indices in params to update

Momentum

class orangecontrib.recommendation.optimizers.Momentum(learning_rate=1.0, momentum=0.9)[source]

Stochastic Gradient Descent (SGD) updates with momentum

Generates update expressions of the form:

  • velocity := momentum * velocity - learning_rate * gradient
  • param := param + velocity
Args:
learning_rate: float
The learning rate controlling the size of update steps
momentum: float, optional
The amount of momentum to apply. Higher momentum results in smoothing over more update steps. Defaults to 0.9.
Notes:
Higher momentum also results in larger update steps. To counter that, you can optionally scale your learning rate by 1 - momentum.
See Also:
apply_momentum: Generic function applying momentum to updates nesterov_momentum: Nesterov’s variant of SGD with momentum
update(grads, params, indices=None)[source]

Momentum updates

Args:
grads: array
List of gradient expressions
params: array
The variables to generate update expressions for
indices: array
Indices in params to update

Nesterov’s Accelerated Gradient

class orangecontrib.recommendation.optimizers.NesterovMomentum(learning_rate=1.0, momentum=0.9)[source]

Stochastic Gradient Descent (SGD) updates with Nesterov momentum

Generates update expressions of the form:

  • param_ahead := param + momentum * velocity
  • velocity := momentum * velocity - learning_rate * gradient_ahead
  • param := param + velocity

In order to express the update to look as similar to vanilla SGD, this can be written as:

  • v_prev := velocity
  • velocity := momentum * velocity - learning_rate * gradient
  • param := -momentum * v_prev + (1 + momentum) * velocity
Args:
learning_rate : float
The learning rate controlling the size of update steps
momentum: float, optional
The amount of momentum to apply. Higher momentum results in smoothing over more update steps. Defaults to 0.9.
Notes:

Higher momentum also results in larger update steps. To counter that, you can optionally scale your learning rate by 1 - momentum.

The classic formulation of Nesterov momentum (or Nesterov accelerated gradient) requires the gradient to be evaluated at the predicted next position in parameter space. Here, we use the formulation described at https://github.com/lisa-lab/pylearn2/pull/136#issuecomment-10381617, which allows the gradient to be evaluated at the current parameters.

See Also:
apply_nesterov_momentum: Function applying momentum to updates
update(grads, params, indices=None)[source]

NAG updates

Args:
grads: array
List of gradient expressions
params: array
The variables to generate update expressions for
indices: array, optional
Indices in params to update
Returns
updates: list of float
Variables updated with the gradients

AdaGradient

class orangecontrib.recommendation.optimizers.AdaGrad(learning_rate=1.0, epsilon=1e-06)[source]

AdaGrad updates

Scale learning rates by dividing with the square root of accumulated squared gradients. See [1] for further description.

  • param := param - learning_rate * gradient
Args:
learning_rate : float or symbolic scalar
The learning rate controlling the size of update steps
epsilon: float or symbolic scalar
Small value added for numerical stability
Notes:

Using step size eta Adagrad calculates the learning rate for feature i at time step t as:

\[\eta_{t,i} = \frac{\eta} {\sqrt{\sum^t_{t^\prime} g^2_{t^\prime,i}+\epsilon}} g_{t,i}\]

as such the learning rate is monotonically decreasing.

Epsilon is not included in the typical formula, see [2].

References:
[1]Duchi, J., Hazan, E., & Singer, Y. (2011): Adaptive subgradient methods for online learning and stochastic optimization. JMLR, 12:2121-2159.
[2]Chris Dyer: Notes on AdaGrad. http://www.ark.cs.cmu.edu/cdyer/adagrad.pdf
update(grads, params, indices=None)[source]

AdaGrad updates

Args:
grads: array
List of gradient expressions
params: array
The variables to generate update expressions for
indices: array, optional
Indices in params to update

RMSProp

class orangecontrib.recommendation.optimizers.RMSProp(learning_rate=1.0, rho=0.9, epsilon=1e-06)[source]

Scale learning rates by dividing with the moving average of the root mean squared (RMS) gradients. See [3] for further description.

Args:
learning_rate: float
The learning rate controlling the size of update steps
rho: float
Gradient moving average decay factor
epsilon: float
Small value added for numerical stability
Notes:

rho should be between 0 and 1. A value of rho close to 1 will decay the moving average slowly and a value close to 0 will decay the moving average fast.

Using the step size \(\eta\) and a decay factor \(\rho\) the learning rate \(\eta_t\) is calculated as:

\[\begin{split}r_t &= \rho r_{t-1} + (1-\rho)*g^2\\ \eta_t &= \frac{\eta}{\sqrt{r_t + \epsilon}}\end{split}\]
References:
[3]Tieleman, T. and Hinton, G. (2012): Neural Networks for Machine Learning, Lecture 6.5 - rmsprop. Coursera. http://www.youtube.com/watch?v=O3sxAc4hxZU (formula @5:20)
update(grads, params, indices=None)[source]

RMSProp updates

Args:
grads: array
List of gradient expressions
params: array
The variables to generate update expressions for
indices: array, optional
Indices in params to update

AdaDelta

class orangecontrib.recommendation.optimizers.AdaDelta(learning_rate=1.0, rho=0.95, epsilon=1e-06)[source]

Scale learning rates by a the ratio of accumulated gradients to accumulated step sizes, see [4] and notes for further description.

Args:
learning_rate: float
The learning rate controlling the size of update steps
rho: float
Squared gradient moving average decay factor
epsilon: float
Small value added for numerical stability
Notes:

rho should be between 0 and 1. A value of rho close to 1 will decay the moving average slowly and a value close to 0 will decay the moving average fast.

rho = 0.95 and epsilon=1e-6 are suggested in the paper and reported to work for multiple datasets (MNIST, speech).

In the paper, no learning rate is considered (so learning_rate=1.0). Probably best to keep it at this value. epsilon is important for the very first update (so the numerator does not become 0).

Using the step size eta and a decay factor rho the learning rate is calculated as:

\[\begin{split}r_t &= \rho r_{t-1} + (1-\rho)*g^2\\ \eta_t &= \eta \frac{\sqrt{s_{t-1} + \epsilon}} {\sqrt{r_t + \epsilon}}\\ s_t &= \rho s_{t-1} + (1-\rho)*g^2\end{split}\]
References:
[4](1, 2) Zeiler, M. D. (2012): ADADELTA: An Adaptive Learning Rate Method. arXiv Preprint arXiv:1212.5701.
update(grads, params, indices=None)[source]

AdaDelta updates

Args:
grads: array
List of gradient expressions
params: array
The variables to generate update expressions for
indices: array, optional
Indices in params to update

Adam

class orangecontrib.recommendation.optimizers.Adam(learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-08)[source]

Adam updates implemented as in [5].

Args:
learning_rate : float
The learning rate controlling the size of update steps
beta_1 : float
Exponential decay rate for the first moment estimates.
beta_2 : float
Exponential decay rate for the second moment estimates.
epsilon : float
Constant for numerical stability.
Notes:
The paper [5] includes an additional hyperparameter lambda. This is only needed to prove convergence of the algorithm and has no practical use, it is therefore omitted here.
References:
[5](1, 2, 3) Kingma, Diederik, and Jimmy Ba (2014): Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980.
update(grads, params, indices=None)[source]

Adam updates

Args:
grads: array
List of gradient expressions
params: array
The variables to generate update expressions for
indices: array, optional
Indices of parameters (‘params’) to update. If None (default), all parameters will be updated.
Returns
updates: list of float
Variables updated with the gradients

Adamax

class orangecontrib.recommendation.optimizers.Adamax(learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-08)[source]

Adamax updates implemented as in [6]. This is a variant of of the Adam algorithm based on the infinity norm.

Args:
learning_rate : float
The learning rate controlling the size of update steps
beta_1 : float
Exponential decay rate for the first moment estimates.
beta_2 : float
Exponential decay rate for the second moment estimates.
epsilon : float
Constant for numerical stability.
References:
[6](1, 2) Kingma, Diederik, and Jimmy Ba (2014): Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980.
update(grads, params, indices=None)[source]

Adamax updates

Args:
grads: array
List of gradient expressions
params: array
The variables to generate update expressions for
indices: array, optional
Indices of parameters (‘params’) to update. If None (default), all parameters will be updated.
Returns
updates: list of float
Variables updated with the gradients