def generate_initialization(distribution):
""" Run mjhmc for BURN_IN_STEPS on distribution, generating a fair set of initial states
:param distribution: Distribution object. Must have nbatch == MAX_N_PARTICLES
:returns: a set of fair initial states and an estimate of the variance for emc and true both
:rtype: tuple: (array of shape (distribution.ndims, MAX_N_PARTICLES), float, float)
"""
print('Generating fair initialization for {} by burning in {} steps'.format(
type(distribution).__name__, BURN_IN_STEPS))
assert BURN_IN_STEPS > VAR_STEPS
assert distribution.nbatch == MAX_N_PARTICLES
mjhmc = MarkovJumpHMC(distribution=distribution, resample=False)
for _ in xrange(BURN_IN_STEPS - VAR_STEPS):
mjhmc.sampling_iteration()
assert mjhmc.resample == False
emc_var_estimate, mjhmc = online_variance(mjhmc, distribution)
# we discard v since p(x,v) = p(x)p(v)
fair_x = mjhmc.state.copy().X
# otherwise will go into recursive loop
distribution.mjhmc = False
control = ControlHMC(distribution=distribution.reset())
for _ in xrange(BURN_IN_STEPS - VAR_STEPS):
control.sampling_iteration()
true_var_estimate, control = online_variance(control, distribution)
return (fair_x, emc_var_estimate, true_var_estimate)
def sampler_speedometer():
""" Tests average sample speed of different samplers in different configurations
:returns: Just prints info
:rtype: None
"""
gaussian = Gaussian()
np.random.seed(2015)
pot = ProductOfT(ndims=36,nbasis=36)
mjhmc_gauss = MarkovJumpHMC(distribution=gaussian)
mjhmc_gauss_nr = MarkovJumpHMC(distribution=gaussian, resample=False)
control_gauss = ControlHMC(distribution=gaussian)
mjhmc_pot = MarkovJumpHMC(distribution=pot)
mjhmc_pot_nr = MarkovJumpHMC(distribution=pot, resample=False)
control_pot = ControlHMC(distribution=pot)
m_g_r_avg = time_per_sample(mjhmc_gauss)
m_g_nr_avg = time_per_sample(mjhmc_gauss_nr)
c_g_avg = time_per_sample(control_gauss)
m_p_r_avg = time_per_sample(mjhmc_pot)
m_p_nr_avg = time_per_sample(mjhmc_pot_nr)
c_p_avg = time_per_sample(control_pot)
print "Average times per samples..."
print "resampled MJHMC numpy gradient: {}".format(m_g_r_avg)
print "not resampled MJHMC numpy gradient: {}".format(m_g_nr_avg)
print "control HMC numpy gradient: {}".format(c_g_avg)
print "resampled MJHMC theano gradient: {}".format(m_p_r_avg)
print "not resampled MJHMC theano gradient: {}".format(m_p_nr_avg)
print "control HMC theano gradient: {}".format(c_p_avg)
def generate_figure_samples(samples_per_frame, n_frames, burnin = int(1e4)):
""" Generates the figure
:param samples_per_frame: number of sample steps between each frame
:param n_frames: number of frames to draw
:returns: None
:rtype: None
"""
n_samples = samples_per_frame * n_frames
ndims = 36
nbasis = 72
rand_val = rand(ndims,nbasis/2,density=0.25)
W = np.concatenate([rand_val.toarray(), -rand_val.toarray()],axis=1)
logalpha = np.random.randn(nbasis, 1)
poe = ProductOfT(nbatch=1, W=W, logalpha=logalpha)
## NUTS uses a different number of grad evals for each update step!!
## makes it very hard to compare against others w/ same number of update steps
# # NUTS
# print "NUTS"
# nuts_init = poe.Xinit[:, 0]
# nuts_samples = nuts6(poe.reset(), n_samples, nuts_burnin, nuts_init)[0]
# nuts_frames = [nuts_samples[f_idx * samples_per_frame, :] for f_idx in xrange(0, n_frames)]
# x_init = nuts_samples[0, :].reshape(ndims, 1)
## burnin
print "MJHMC burnin"
x_init = poe.Xinit #[:, [0]]
mjhmc = MarkovJumpHMC(distribution=poe.reset(), **mjhmc_params)
mjhmc.state = HMCState(x_init.copy(), mjhmc)
mjhmc_samples = mjhmc.sample(burnin)
print mjhmc_samples.shape
x_init = mjhmc_samples[:, [0]]
# control HMC
print "Control"
hmc = ControlHMC(distribution=poe.reset(), **control_params)
hmc.state = HMCState(x_init.copy(), hmc)
hmc_samples = hmc.sample(n_samples)
hmc_frames = [hmc_samples[:, f_idx * samples_per_frame].copy() for f_idx in xrange(0, n_frames)]
# MJHMC
print "MJHMC"
mjhmc = MarkovJumpHMC(distribution=poe.reset(), resample=False, **mjhmc_params)
mjhmc.state = HMCState(x_init.copy(), mjhmc)
mjhmc_samples = mjhmc.sample(n_samples)
mjhmc_frames = [mjhmc_samples[:, f_idx * samples_per_frame].copy() for f_idx in xrange(0, n_frames)]
print mjhmc.r_count, hmc.r_count
print mjhmc.l_count, hmc.l_count
print mjhmc.f_count, hmc.f_count
print mjhmc.fl_count, hmc.fl_count
frames = [mjhmc_frames, hmc_frames]
names = ['MJHMC', 'ControlHMC']
frame_grads = [f_idx * samples_per_frame for f_idx in xrange(0, n_frames)]
return frames, names, frame_grads
def generate_figure_samples(samples_per_frame, n_frames, burnin=int(1e4)):
""" Generates the figure
:param samples_per_frame: number of sample steps between each frame
:param n_frames: number of frames to draw
:returns: None
:rtype: None
"""
n_samples = samples_per_frame * n_frames
ndims = 36
nbasis = 72
rand_val = rand(ndims, nbasis / 2, density=0.25)
W = np.concatenate([rand_val.toarray(), -rand_val.toarray()], axis=1)
logalpha = np.random.randn(nbasis, 1)
poe = ProductOfT(nbatch=1, W=W, logalpha=logalpha)
## NUTS uses a different number of grad evals for each update step!!
## makes it very hard to compare against others w/ same number of update steps
# # NUTS
# print "NUTS"
# nuts_init = poe.Xinit[:, 0]
# nuts_samples = nuts6(poe.reset(), n_samples, nuts_burnin, nuts_init)[0]
# nuts_frames = [nuts_samples[f_idx * samples_per_frame, :] for f_idx in xrange(0, n_frames)]
# x_init = nuts_samples[0, :].reshape(ndims, 1)
## burnin
print "MJHMC burnin"
x_init = poe.Xinit # [:, [0]]
mjhmc = MarkovJumpHMC(distribution=poe.reset(), **mjhmc_params)
mjhmc.state = HMCState(x_init.copy(), mjhmc)
mjhmc_samples = mjhmc.sample(burnin)
print mjhmc_samples.shape
x_init = mjhmc_samples[:, [0]]
# control HMC
print "Control"
hmc = ControlHMC(distribution=poe.reset(), **control_params)
hmc.state = HMCState(x_init.copy(), hmc)
hmc_samples = hmc.sample(n_samples)
hmc_frames = [hmc_samples[:, f_idx * samples_per_frame].copy() for f_idx in xrange(0, n_frames)]
# MJHMC
print "MJHMC"
mjhmc = MarkovJumpHMC(distribution=poe.reset(), resample=False, **mjhmc_params)
mjhmc.state = HMCState(x_init.copy(), mjhmc)
mjhmc_samples = mjhmc.sample(n_samples)
mjhmc_frames = [mjhmc_samples[:, f_idx * samples_per_frame].copy() for f_idx in xrange(0, n_frames)]
print mjhmc.r_count, hmc.r_count
print mjhmc.l_count, hmc.l_count
print mjhmc.f_count, hmc.f_count
print mjhmc.fl_count, hmc.fl_count
frames = [mjhmc_frames, hmc_frames]
names = ["MJHMC", "ControlHMC"]
frame_grads = [f_idx * samples_per_frame for f_idx in xrange(0, n_frames)]
return frames, names, frame_grads
def generate_initialization(distribution):
""" Run mjhmc for BURN_IN_STEPS on distribution, generating a fair set of initial states
:param distribution: Distribution object. Must have nbatch == MAX_N_PARTICLES
:returns: a set of fair initial states and an estimate of the variance for emc and true both
:rtype: tuple: (array of shape (distribution.ndims, MAX_N_PARTICLES), float, float)
"""
print(
'Generating fair initialization for {} by burning in {} steps'.format(
type(distribution).__name__, BURN_IN_STEPS))
assert BURN_IN_STEPS > VAR_STEPS
# must rebuild graph to nbatch=MAX_N_PARTICLES
if distribution.backend == 'tensorflow':
distribution.build_graph()
mjhmc = MarkovJumpHMC(distribution=distribution, resample=False)
for _ in xrange(BURN_IN_STEPS - VAR_STEPS):
mjhmc.sampling_iteration()
assert mjhmc.resample == False
emc_var_estimate, mjhmc = online_variance(mjhmc, distribution)
# we discard v since p(x,v) = p(x)p(v)
mjhmc_endpt = mjhmc.state.copy().X
# otherwise will go into recursive loop
distribution.mjhmc = False
try:
distribution.gen_init_X()
except NotImplementedError:
print("No explicit init method found, using mjhmc endpoint")
distribution.E_count = 0
distribution.dEdX_count = 0
control = ControlHMC(distribution=distribution)
for _ in xrange(BURN_IN_STEPS - VAR_STEPS):
control.sampling_iteration()
true_var_estimate, control = online_variance(control, distribution)
control_endpt = control.state.copy().X
return mjhmc_endpt, emc_var_estimate, true_var_estimate, control_endpt
def generate_initialization(distribution):
""" Run mjhmc for BURN_IN_STEPS on distribution, generating a fair set of initial states
:param distribution: Distribution object. Must have nbatch == MAX_N_PARTICLES
:returns: a set of fair initial states and an estimate of the variance for emc and true both
:rtype: tuple: (array of shape (distribution.ndims, MAX_N_PARTICLES), float, float)
"""
print('Generating fair initialization for {} by burning in {} steps'.format(
type(distribution).__name__, BURN_IN_STEPS))
assert BURN_IN_STEPS > VAR_STEPS
# must rebuild graph to nbatch=MAX_N_PARTICLES
if distribution.backend == 'tensorflow':
distribution.build_graph()
mjhmc = MarkovJumpHMC(distribution=distribution, resample=False)
for _ in xrange(BURN_IN_STEPS - VAR_STEPS):
mjhmc.sampling_iteration()
assert mjhmc.resample == False
emc_var_estimate, mjhmc = online_variance(mjhmc, distribution)
# we discard v since p(x,v) = p(x)p(v)
mjhmc_endpt = mjhmc.state.copy().X
# otherwise will go into recursive loop
distribution.mjhmc = False
try:
distribution.gen_init_X()
except NotImplementedError:
print("No explicit init method found, using mjhmc endpoint")
distribution.E_count = 0
distribution.dEdX_count = 0
control = ControlHMC(distribution=distribution)
for _ in xrange(BURN_IN_STEPS - VAR_STEPS):
control.sampling_iteration()
true_var_estimate, control = online_variance(control, distribution)
control_endpt = control.state.copy().X
return mjhmc_endpt, emc_var_estimate, true_var_estimate, control_endpt
def ladder_numerical_err_hist(distr=None, n_steps=int(1e5)):
""" Compute a histogram of the numerical integration error
on the state ladder. Implicitly assumes that such a distribution exists
and is shared by all ladders
Args:
distr: distribution object to run on, make sure n_batch is big
Returns:
energies = {E(L^j \zeta) : j \in {0, ..., k}}^{n_batch}
run_lengths: list of observed ladder sizes
"""
distr = distr or Gaussian(nbatch=1)
sampler = ControlHMC(distribution=distr)
# [[ladder_energies]]
energies = []
run_lengths = []
r_counts = [0]
ladder_energies = [np.squeeze(sampler.state.H())]
run_length = 0
for _ in range(n_steps):
if sampler.r_count == r_counts[-1]:
run_length += 1
ladder_energies.append(np.squeeze(sampler.state.H()))
else:
run_lengths.append(run_length)
run_length = 0
energies.append(np.array(ladder_energies))
ladder_energies = [np.squeeze(sampler.state.H())]
r_counts.append(sampler.r_count)
sampler.sampling_iteration()
centered_energies = []
for ladder_energies in energies:
centered_energies += list(ladder_energies - ladder_energies[0])
return centered_energies, run_lengths
def ladder_numerical_err_hist(distr=None, n_steps=int(1e5)):
""" Compute a histogram of the numerical integration error
on the state ladder. Implicitly assumes that such a distribution exists
and is shared by all ladders
Args:
distr: distribution object to run on, make sure n_batch is big
Returns:
energies = {E(L^j \zeta) : j \in {0, ..., k}}^{n_batch}
run_lengths: list of observed ladder sizes
"""
distr = distr or Gaussian(nbatch=1)
sampler = ControlHMC(distribution=distr)
# [[ladder_energies]]
energies = []
run_lengths = []
r_counts = [0]
ladder_energies = [np.squeeze(sampler.state.H())]
run_length = 0
for _ in range(n_steps):
if sampler.r_count == r_counts[-1]:
run_length += 1
ladder_energies.append(np.squeeze(sampler.state.H()))
else:
run_lengths.append(run_length)
run_length = 0
energies.append(np.array(ladder_energies))
ladder_energies = [np.squeeze(sampler.state.H())]
r_counts.append(sampler.r_count)
sampler.sampling_iteration()
centered_energies = []
for ladder_energies in energies:
centered_energies += list(ladder_energies - ladder_energies[0])
return centered_energies, run_lengths
