pylearn: pylearn/sampling/hmc.py annotate

annotate pylearn/sampling/hmc.py @ 1501:55534951dd91

Clean up import and remove deprecation warning.

author	Frederic Bastien <nouiz@nouiz.org>
date	Fri, 09 Sep 2011 10:53:46 -0400
parents	fbe470217937
children	4fa5ebe8a7ad

rev	line source
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	1
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	2 """Hybrid / Hamiltonian Monte Carlo Sampling
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	3
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	4 This algorithm is described in Radford Neal's PhD Thesis, pages 63--70.
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	5
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	6 :note:
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	7
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	8 The 'mass' of so-called particles is taken to be 1, so that kinetic energy (K) is the sum
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	9 of squared velocities (p).
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	10
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	11 :math:`K = \sum_i p_i^2 / 2`
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	12
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	13
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	14 The 'leap-frog' algorithm that advances by turns the velocity and the position is
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	15 currently implemented via several theano functions, rather than one complete theano
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	16 expression graph.
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	17
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	18 The initialize_dynamics() theano-function does several things:
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	19 1. samples a random velocity for each particle (saving it to self.velocities)
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	20 2. calculates the initial hamiltonian based on those velocities (saving it to
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	21 self.initial_hamiltonian)
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	22 3. saves self.positions to self.initial_positions.
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	23
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	24 The finalize_dynamics() theano-function re-calculates the Hamiltonian for each particle
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	25 based on the self.positions and self.velocities, and then implements the
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	26 Metropolis-Hastings accept/reject for each particle in the simulation by consulting the
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	27 self.initial_hamiltonian storing the result to self.
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	28
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	29
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	30 """
1270 d38cb039c662 debugging mcRBM James Bergstra <bergstrj@iro.umontreal.ca> parents: 1269 diff changeset	31 import numpy
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	32 from theano import function, shared
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	33 from theano import tensor as TT
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	34 import theano
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	35 from theano.printing import Print
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	36
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	37 def Print(msg):
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	38 return lambda x: x
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	39
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	40 def kinetic_energy(velocities, masses):
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	41 if masses != 1:
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	42 raise NotImplementedError()
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	43 return 0.5 * (velocities**2).sum(axis=1)
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	44
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	45 def metropolis_hastings_accept(energy_prev, energy_next, s_rng, shape=None):
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	46 """
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	47 Return acceptance of moves - energy_prev and energy_next are vectors of the energy of each
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	48 particle.
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	49 """
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	50 if shape is None:
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	51 shape = energy_prev.shape
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	52 ediff = Print('diff')(energy_prev - energy_next)
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	53 return (TT.exp(ediff) - s_rng.uniform(size=shape)) >= 0
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	54
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	55 def hamiltonian(pos, vel, energy_fn):
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	56 """Return a vector of energies - sum of kinetic and potential energy
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	57 """
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	58 # assuming mass is 1
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	59 return energy_fn(pos) + kinetic_energy(vel, 1)
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	60
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	61 def simulate_dynamics(initial_p, initial_v, stepsize, n_steps, energy_fn):
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	62 """
1341 fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	63 Return final (position, velocity) obtained after an `n_steps` leapfrog updates, using
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	64 Hamiltonian dynamics.
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	65
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	66 Parameters
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	67 ----------
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	68 initial_p: shared theano matrix
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	69 Initial position at which to start the simulation
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	70 initial_v: shared theano matrix
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	71 Initial velocity of particles
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	72 stepsize: shared theano scalar
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	73 Scalar value controlling amount by which to move
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	74 energy_fn: python function
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	75 Python function, operating on symbolic theano variables, used to compute the potential
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	76 energy at a given position.
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	77
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	78 Returns
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	79 -------
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	80 rval1: theano matrix
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	81 Final positions obtained after simulation
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	82 rval2: theano matrix
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	83 Final velocity obtained after simulation
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	84 """
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	85
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	86 def leapfrog(pos, vel, step):
1341 fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	87 """
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	88 Inside loop of Scan. Performs one step of leapfrog update, using Hamiltonian dynamics.
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	89
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	90 Parameters
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	91 ----------
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	92 pos: theano matrix
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	93 in leapfrog update equations, represents pos(t), position at time t
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	94 vel: theano matrix
1342 4ac393ec2eb7 Fixed comments in leapfrog method gdesjardins parents: 1341 diff changeset	95 in leapfrog update equations, represents vel(t - stepsize/2),
4ac393ec2eb7 Fixed comments in leapfrog method gdesjardins parents: 1341 diff changeset	96 velocity at time (t - stepsize/2)
1341 fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	97 step: theano scalar
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	98 scalar value controlling amount by which to move
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	99
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	100 Returns
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	101 -------
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	102 rval1: [theano matrix, theano matrix]
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	103 Symbolic theano matrices for new position pos(t + stepsize), and velocity
1342 4ac393ec2eb7 Fixed comments in leapfrog method gdesjardins parents: 1341 diff changeset	104 vel(t + stepsize/2)
1341 fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	105 rval2: dictionary
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	106 Dictionary of updates for the Scan Op
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	107 """
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	108 # from pos(t) and vel(t-sigma/2), compute vel(t+sigma/2)
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	109 dE_dpos = TT.grad(energy_fn(pos).sum(), pos)
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	110 new_vel = vel - step * dE_dpos
1341 fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	111 # from vel(t+sigma/2) compute pos(t+sigma)
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	112 new_pos = pos + step * new_vel
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	113 return [new_pos, new_vel],{}
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	114
1341 fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	115 # compute velocity at time-step: t + sigma/2
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	116 initial_energy = energy_fn(initial_p)
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	117 dE_dpos = TT.grad(initial_energy.sum(), initial_p)
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	118 v_half_step = initial_v - 0.5stepsizedE_dpos
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	119
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	120 p_full_step = initial_p + stepsize * v_half_step
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	121
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	122 # perform leapfrog updates: the scan op is used to repeatedly compute pos(t_1 + n*sigma) and
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	123 # vel(t_1 + n*sigma + 1/2) for n in [0,n_steps-2].
1501 55534951dd91 Clean up import and remove deprecation warning. Frederic Bastien <nouiz@nouiz.org> parents: 1447 diff changeset	124 (all_p, all_v), scan_updates = theano.scan(leapfrog,
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	125 outputs_info=[
1501 55534951dd91 Clean up import and remove deprecation warning. Frederic Bastien <nouiz@nouiz.org> parents: 1447 diff changeset	126 dict(initial=p_full_step),
55534951dd91 Clean up import and remove deprecation warning. Frederic Bastien <nouiz@nouiz.org> parents: 1447 diff changeset	127 dict(initial=v_half_step),
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	128 ],
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	129 non_sequences=[stepsize],
1341 fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	130 n_steps=n_steps-1)
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	131
1501 55534951dd91 Clean up import and remove deprecation warning. Frederic Bastien <nouiz@nouiz.org> parents: 1447 diff changeset	132 final_p = all_p[-1]
55534951dd91 Clean up import and remove deprecation warning. Frederic Bastien <nouiz@nouiz.org> parents: 1447 diff changeset	133 final_v = all_v[-1]
55534951dd91 Clean up import and remove deprecation warning. Frederic Bastien <nouiz@nouiz.org> parents: 1447 diff changeset	134
1390 746ebceeb46f added comments to hmc code (old outstanding changes) gdesjardins parents: 1342 diff changeset	135 # NOTE: Scan always returns an updates dictionary, in case the scanned function draws
746ebceeb46f added comments to hmc code (old outstanding changes) gdesjardins parents: 1342 diff changeset	136 # samples from a RandomStream. These updates must then be used when compiling the Theano
746ebceeb46f added comments to hmc code (old outstanding changes) gdesjardins parents: 1342 diff changeset	137 # function, to avoid drawing the same random numbers each time the function is called. In
746ebceeb46f added comments to hmc code (old outstanding changes) gdesjardins parents: 1342 diff changeset	138 # this case however, we consciously ignore "scan_updates" because we know it is empty.
746ebceeb46f added comments to hmc code (old outstanding changes) gdesjardins parents: 1342 diff changeset	139 assert not scan_updates
746ebceeb46f added comments to hmc code (old outstanding changes) gdesjardins parents: 1342 diff changeset	140
1341 fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	141 # The last velocity returned by the scan op is at time-step: t + n_steps* stepsize - 1/2
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	142 # We therefore perform one more half-step to return vel(t + n_steps*stepsize)
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	143 energy = energy_fn(final_p)
fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	144 final_v = final_v - 0.5 * stepsize * TT.grad(energy.sum(), final_p)
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	145 return final_p, final_v
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	146
1341 fbe4a6383441 HMC: perform half-step on velocity first (instead of position). gdesjardins parents: 1280 diff changeset	147
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	148 def mcmc_move(s_rng, positions, energy_fn, stepsize, n_steps, positions_shape=None):
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	149 """Return new position
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	150 """
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	151 if positions_shape is None:
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	152 positions_shape = positions.shape
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	153
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	154 batchsize = positions_shape[0]
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	155
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	156 initial_v = s_rng.normal(size=positions_shape)
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	157
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	158 final_p, final_v = simulate_dynamics(
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	159 initial_p = positions,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	160 initial_v = initial_v,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	161 stepsize = stepsize,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	162 n_steps = n_steps,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	163 energy_fn = energy_fn)
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	164
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	165 accept = metropolis_hastings_accept(
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	166 energy_prev = Print('ep')(hamiltonian(positions, initial_v, energy_fn)),
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	167 energy_next = Print('en')(hamiltonian(final_p, final_v, energy_fn)),
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	168 s_rng=s_rng, shape=(batchsize,))
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	169
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	170 return Print('accept')(accept), final_p
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	171
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	172 def mcmc_updates(shrd_pos, shrd_stepsize, shrd_avg_acceptance_rate, final_p, accept,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	173 target_acceptance_rate,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	174 stepsize_inc,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	175 stepsize_dec,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	176 stepsize_min,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	177 stepsize_max,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	178 avg_acceptance_slowness
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	179 ):
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	180 return [
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	181 (shrd_pos,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	182 TT.switch(
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	183 accept.dimshuffle(0, (('x',)(final_p.ndim-1))),
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	184 final_p,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	185 shrd_pos)),
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	186 (shrd_stepsize,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	187 TT.clip(
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	188 TT.switch(
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	189 shrd_avg_acceptance_rate > target_acceptance_rate,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	190 shrd_stepsize * stepsize_inc,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	191 shrd_stepsize * stepsize_dec,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	192 ),
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	193 stepsize_min,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	194 stepsize_max)),
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	195 (shrd_avg_acceptance_rate,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	196 Print('arate')(TT.add(
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	197 avg_acceptance_slowness * shrd_avg_acceptance_rate,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	198 (1.0 - avg_acceptance_slowness) * accept.mean()))),
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	199 ]
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	200
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	201 class HMC_sampler(object):
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	202 """Convenience wrapper for HMC
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	203
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	204 The `draw` function advances the markov chain and returns the current sample by calling
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	205 `simulate` and `get_position` in sequence.
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	206
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	207 """
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	208
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	209 # Constants taken from Marc'Aurelio's 'train_mcRBM.py' file found in the code online for his
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	210 # paper.
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	211
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	212 def __init__(self, **kwargs):
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	213 # add things to __dict__
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	214 self.__dict__.update(kwargs)
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	215
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	216 @classmethod
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	217 def new_from_shared_positions(cls, shared_positions, energy_fn,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	218 initial_stepsize=0.01, target_acceptance_rate=.9, n_steps=20,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	219 stepsize_dec = 0.98,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	220 stepsize_min = 0.001,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	221 stepsize_max = 0.25,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	222 stepsize_inc = 1.02,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	223 avg_acceptance_slowness = 0.9, # used in geometric avg. 1.0 would be not moving at all
1280 1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	224 seed=12345, dtype=theano.config.floatX,
1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	225 shared_positions_shape=None,
1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	226 compile_simulate=True):
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	227 """
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	228 :param shared_positions: theano ndarray shared var with many particle [initial] positions
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	229 :param energy_fn:
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	230 callable such that energy_fn(positions)
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	231 returns theano vector of energies.
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	232 The len of this vector is the batchsize.
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	233
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	234 The sum of this energy vector must be differentiable (with theano.tensor.grad) with
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	235 respect to the positions for HMC sampling to work.
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	236 """
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	237 # allocate shared vars
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	238
1280 1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	239 if shared_positions_shape==None:
1447 fbe470217937 Use .get_value() and .set_value() of shared instead of the .value property Pascal Lamblin <lamblinp@iro.umontreal.ca> parents: 1390 diff changeset	240 shared_positions_shape = shared_positions.get_value(borrow=True).shape
1280 1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	241 batchsize = shared_positions_shape[0]
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	242
1270 d38cb039c662 debugging mcRBM James Bergstra <bergstrj@iro.umontreal.ca> parents: 1269 diff changeset	243 stepsize = shared(numpy.asarray(initial_stepsize).astype(theano.config.floatX), 'hmc_stepsize')
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	244 avg_acceptance_rate = shared(target_acceptance_rate, 'avg_acceptance_rate')
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	245 s_rng = TT.shared_randomstreams.RandomStreams(seed)
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	246
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	247 accept, final_p = mcmc_move(
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	248 s_rng,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	249 shared_positions,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	250 energy_fn,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	251 stepsize,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	252 n_steps,
1280 1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	253 shared_positions_shape)
1266 bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	254 simulate_updates = mcmc_updates(
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	255 shared_positions,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	256 stepsize,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	257 avg_acceptance_rate,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	258 final_p=final_p,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	259 accept=accept,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	260 stepsize_min=stepsize_min,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	261 stepsize_max=stepsize_max,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	262 stepsize_inc=stepsize_inc,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	263 stepsize_dec=stepsize_dec,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	264 target_acceptance_rate=target_acceptance_rate,
bc4d98995bad hmc - separated simulate into simulate_updates James Bergstra <bergstrj@iro.umontreal.ca> parents: 1265 diff changeset	265 avg_acceptance_slowness=avg_acceptance_slowness)
1280 1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	266 if compile_simulate:
1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	267 simulate = function([], [], updates=simulate_updates)
1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	268 else:
1eaa015e88c1 extended hmc to allow for use as a cd1 sampler James Bergstra <bergstrj@iro.umontreal.ca> parents: 1274 diff changeset	269 simulate = None
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	270 return cls(
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	271 positions=shared_positions,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	272 stepsize=stepsize,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	273 stepsize_min=stepsize_min,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	274 stepsize_max=stepsize_max,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	275 avg_acceptance_rate=avg_acceptance_rate,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	276 target_acceptance_rate=target_acceptance_rate,
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	277 s_rng=s_rng,
1274 9d5905d6d879 hmc - changed updates to member fn from lambda for pickling James Bergstra <bergstrj@iro.umontreal.ca> parents: 1271 diff changeset	278 _updates=simulate_updates,
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	279 simulate=simulate)
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	280
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	281 def draw(self, **kwargs):
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	282 """Return the current sample in the Markov chain as a list of numpy arrays
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	283
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	284 The size of the arrays will match the size of the `initial_position` argument to
492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	285 __init__.
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	286
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	287 The `kwargs` dictionary is passed to the shared variable (self.positions) `get_value()`
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	288 function. So for example, to avoid copying the shared variable value, consider passing
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	289 `borrow=True`.
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	290 """
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	291 self.simulate()
1447 fbe470217937 Use .get_value() and .set_value() of shared instead of the .value property Pascal Lamblin <lamblinp@iro.umontreal.ca> parents: 1390 diff changeset	292 return self.positions.get_value(borrow=False)
955 492473059b37 Adding sampling module James Bergstra <bergstrj@iro.umontreal.ca> parents: diff changeset	293
1274 9d5905d6d879 hmc - changed updates to member fn from lambda for pickling James Bergstra <bergstrj@iro.umontreal.ca> parents: 1271 diff changeset	294 def updates(self):
9d5905d6d879 hmc - changed updates to member fn from lambda for pickling James Bergstra <bergstrj@iro.umontreal.ca> parents: 1271 diff changeset	295 """Returns the update expressions required to simulate the Markov Chain
9d5905d6d879 hmc - changed updates to member fn from lambda for pickling James Bergstra <bergstrj@iro.umontreal.ca> parents: 1271 diff changeset	296
9d5905d6d879 hmc - changed updates to member fn from lambda for pickling James Bergstra <bergstrj@iro.umontreal.ca> parents: 1271 diff changeset	297 :TODO: :WRITEME: prescriptive definition of what this method does (API)
9d5905d6d879 hmc - changed updates to member fn from lambda for pickling James Bergstra <bergstrj@iro.umontreal.ca> parents: 1271 diff changeset	298 """
9d5905d6d879 hmc - changed updates to member fn from lambda for pickling James Bergstra <bergstrj@iro.umontreal.ca> parents: 1271 diff changeset	299 return list(self._updates)
9d5905d6d879 hmc - changed updates to member fn from lambda for pickling James Bergstra <bergstrj@iro.umontreal.ca> parents: 1271 diff changeset	300
1265 d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	301 #TODO:
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	302 # Consider a heuristic for updating the MASS of the particles. We might want the mass to be
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	303 # such that the momentum is in the same range as the gradient on the energy. Look at Radford's
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	304 # recent book chapter, maybe there are hints. (2010).
d6665a5af743 hmc - replaced with new refactored code James Bergstra <bergstrj@iro.umontreal.ca> parents: 1264 diff changeset	305

Mercurial > pylearn

annotate pylearn/sampling/hmc.py @ 1501:55534951dd91