afcarl / downhill

Stochastic gradient routines for Theano

DOWNHILL

The downhill package provides algorithms for minimizing scalar loss functions that are defined using Theano.

Several optimization algorithms are included:

• ADADELTA
• ADAGRAD
• Adam
• Equilibrated SGD
• Nesterov's Accelerated Gradient
• RMSProp
• Resilient Backpropagation
• Stochastic Gradient Descent

All algorithms permit the use of regular or Nesterov-style momentum as well.

Quick Start: Matrix Factorization

Let's say you have 100 samples of 1000-dimensional data, and you want to represent your data as 100 coefficients in a 10-dimensional basis. This is pretty straightforward to model using Theano: you can use a matrix multiplication as the data model, a squared-error term for optimization, and a sparse regularizer to encourage small coefficient values.

Once you have constructed an expression for the loss, you can optimize it with a single call to downhill.minimize:

import downhill
import numpy as np
import theano
import theano.tensor as TT

FLOAT = 'df'[theano.config.floatX == 'float32']

def rand(a, b):
    return np.random.randn(a, b).astype(FLOAT)

A, B, K = 20, 5, 3

# Set up a matrix factorization problem to optimize.
u = theano.shared(rand(A, K), name='u')
v = theano.shared(rand(K, B), name='v')
z = TT.matrix()
err = TT.sqr(z - TT.dot(u, v))
loss = err.mean() + abs(u).mean() + (v * v).mean()

# Minimize the regularized loss with respect to a data matrix.
y = np.dot(rand(A, K), rand(K, B)) + rand(A, B)

# Monitor during optimization.
monitors = (('err', err.mean()),
            ('|u|<0.1', (abs(u) < 0.1).mean()),
            ('|v|<0.1', (abs(v) < 0.1).mean()))

downhill.minimize(
    loss=loss,
    train=[y],
    patience=0,
    batch_size=A,                 # Process y as a single batch.
    max_gradient_norm=1,          # Prevent gradient explosion!
    learning_rate=0.1,
    monitors=monitors,
    monitor_gradients=True)

# Print out the optimized coefficients u and basis v.
print('u =', u.get_value())
print('v =', v.get_value())

If you prefer to maintain more control over your model during optimization, downhill provides an iterative optimization interface:

opt = downhill.build(algo='rmsprop',
                     loss=loss,
                     monitors=monitors,
                     monitor_gradients=True)

for metrics, _ in opt.iterate(train=[[y]],
                              patience=0,
                              batch_size=A,
                              max_gradient_norm=1,
                              learning_rate=0.1):
    print(metrics)

If that's still not enough, you can just plain ask downhill for the updates to your model variables and do everything else yourself:

updates = downhill.build('rmsprop', loss).get_updates(
    batch_size=A, max_gradient_norm=1, learning_rate=0.1)
func = theano.function([z], loss, updates=list(updates))
for _ in range(100):
    print(func(y))  # Evaluate func and apply variable updates.

More Information

Source: http://github.com/lmjohns3/downhill

Documentation: http://downhill.readthedocs.org

Mailing list: https://groups.google.com/forum/#!forum/downhill-users