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

Several optimization algorithms are included:

  • First-order stochastic gradient descent: SGD and NAG.
  • First-order stochastic techniques with adaptive learning rates: RProp, RMSProp, Equilibrated SGD, Adam, and ADADELTA.
  • Wrappers for several algorithms from scipy.optimize.minimize.

The source code for downhill lives at, the documentation lives at, and announcements and discussion happen on the mailing list.

Example Code

Let’s say you want to compute a sparse, low-rank approximation for some 1000-dimensional data that you have lying around. You can represent a batch of \(m\) of data points \(X \in \mathbb{R}^{m \times 1000}\) as the product of a sparse coefficient matrix \(U \in \mathbb{R}^{m \times k}\) and a low-rank basis matrix \(V \in \mathbb{R}^{k \times 1000}\). You might represent the loss as

\[\mathcal{L} = \| X - UV \|_2^2 + \alpha \| U \|_1 + \beta \| V \|_2\]

where the first term represents the approximation error, the second represents the sparsity of the representation, and the third prevents the basis vectors from growing too large.

This is pretty straightforward to model using Theano. Once you set up the appropriate variables and an expression for the loss, you can optimize the loss with respect to the variables using downhill:

import climate
import theano
import theano.tensor as TT
import downhill
import my_data_set


A, B, K = 1000, 2000, 10

x = TT.matrix('x')

u = theano.shared(np.random.randn(A, K).astype('f'), name='u')
v = theano.shared(np.random.randn(K, B).astype('f'), name='v')

err = TT.sqr(x -, v))

    loss=err.mean() + abs(u).mean() + (v * v).mean(),
    params=[u, v],
        ('u<0.1', 100 * (abs(u) < 0.1).mean()),
        ('v<0.1', 100 * (abs(v) < 0.1).mean()),

After optimization, you can get the \(u\) and \(v\) matrix values out of the shared variables using u.get_value() and v.get_value().