apgpy

Python package that implements an accelerated proximal gradient method for minimizing convex functions (Nesterov 2007, Beck and Teboulle 2009).

solves:

minimize f(x) + h(x)
over x \in R^dim_x

where f is smooth, convex - user supplies function to evaluate gradient of f
h is convex - user supplies function to evaluate the proximal operator of h

call as:

x = apgpy.solve( grad_f, prox_h, dim_x )

solve has call signature:

def solve(grad_f, prox_h, dim_x,
          max_iters=2000,
          eps=1e-6,
          alpha=1.01,
          beta=0.5,
          gen_plots=True,
          use_restart=True,
          x_init=False,
          quiet=False,
          use_gra=False,
          step_size=False,
          fixed_step_size=False,
          debug=False)

this takes in two functions:

grad_f(v) = df(v)/dv 
(gradient of f at v)

prox_h(v, t) = argmin_x ( t * h(x) + 1/2 * norm(x-v)^2 )
where t is the (scalar, positive) step size at that iteration

each iteration of apg requires one gradient evaluation of f and one prox step with h

quits when:

norm( y(k) - x(k+1) ) / t < eps * max( 1,norm( x(k+1) )

apg implements something similar to TFOCS step-size adaptation (Becker, Candes and Grant 2010)
and gradient-scheme adaptive restarting (O'Donoghue and Candes 2013)

see notebooks/ for usage instances

to install type python setup.py install

krsna6/apgpy

apgpy