gradient-optimizers

Gradient Optimizers

Optimize you Theano Models with Adagrad, Hessian Free optimization, or linear updates.

pip3 install gradient-optimizers

See example notebook (TBD) for tutorial.

Two classes GradientModel, and GradientHFModel, for optimizing gradient based models (specifically built with indexed parameters in mind (e.g. for language models))

GradientModel

A gradient model for updating your model with hessian free, adagrad, or linear decay updates.

You will need to define the following attributes, and fill them as appropriate:

# a forward method for getting errors:
projection = self.projection_function(ivector <indices/>)

# a cost function (that takes the result of projection function and labels as input)
# and returns a symbolic differentiable theano variable
self.cost_function(projection, ivector <label/>).sum()

self.params = []
self.indexed_params = set()

self._l2_regularization = True / False

self.store_max_updates = True / False

# set this theano setting
self.theano_mode = "FAST_RUN"

# set this theano setting
self.disconnected_inputs = 'ignore' / None

# if L2 is true store this parameter:
self._l2_regularization_parameter = theano.shared(np.float64(l2_regularization).astype(REAL), name='l2_regularization_parameter')

Upon initialization you must run:

self._select_update_mechanism(update_method_name)

# then to compile this mechanism:
self.create_update_fun()

The update methods expect the input to be of the form:

ivector <indices/>, ivector <labels/>

If this is not the case you can modify them as appropriate.

GradientHFModel

Implements an symbolic one step of hessian-free [1] optimization that approximates the curvature, requires a _compute_cost method that takes an example as input or a _compute_cost_gradients that returns gradients for each example provided.

Model should have a params property containing symbolic theano variables.

[1] James Martens, ``Deep learning via Hessian-free optimization", ICML 2010

Make sure the following parameters are not tampered with:

self._additional_params

self._num_updates