Skip to main content
This is a pre-production deployment of Warehouse. Changes made here affect the production instance of PyPI (pypi.python.org).
Help us improve Python packaging - Donate today!

Python package for wrapping gradient optimizers for models in Theano

Project Description
# Gradient Optimizers

Optimize you Theano Models with [Adagrad](http://www.magicbroom.info/Papers/DuchiHaSi10.pdf), 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](http://www.icml2010.org/papers/458.pdf)

Make sure the following parameters are not tampered with:

self._additional_params

self._num_updates
Release History

Release History

This version
History Node

0.0.4

History Node

0.0.3

History Node

0.0.2

History Node

0.0.1

Download Files

Download Files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

File Name & Checksum SHA256 Checksum Help Version File Type Upload Date
gradient-optimizers-0.0.4.tar.gz (9.2 kB) Copy SHA256 Checksum SHA256 Source Oct 30, 2014

Supported By

WebFaction WebFaction Technical Writing Elastic Elastic Search Pingdom Pingdom Monitoring Dyn Dyn DNS Sentry Sentry Error Logging CloudAMQP CloudAMQP RabbitMQ Heroku Heroku PaaS Kabu Creative Kabu Creative UX & Design Fastly Fastly CDN DigiCert DigiCert EV Certificate Rackspace Rackspace Cloud Servers DreamHost DreamHost Log Hosting