def main():
import metropolis_hastings_sampler
sampling_options = {}
sampling_options["E"] = E
sampling_options["proposal_stddev"] = 0.1
sampling_options["thinning_factor"] = 10000
sampling_options["n_samples"] = 10000
sampling_options["mcmc_method"] = "metropolis_hastings_E"
results = metropolis_hastings_sampler.mcmc_generate_samples(sampling_options)
import cPickle
output_pkl_name = "ninja_star_metropolis_hasting_samples_n_%d_thin_%d_prop_stddev_%f.pkl" % (sampling_options["n_samples"], sampling_options["thinning_factor"], sampling_options["proposal_stddev"])
f = open(output_pkl_name, "w")
cPickle.dump(results['samples'], f)
f.close()
print "Acceptance ratio : %f" % results["acceptance_ratio"]
print "Wrote %s" % (output_pkl_name, )
def main():
import metropolis_hastings_sampler
sampling_options = {}
sampling_options["E"] = E
sampling_options["proposal_stddev"] = 0.1
sampling_options["thinning_factor"] = 10000
sampling_options["n_samples"] = 10000
sampling_options["mcmc_method"] = "metropolis_hastings_E"
results = metropolis_hastings_sampler.mcmc_generate_samples(
sampling_options)
import cPickle
output_pkl_name = "ninja_star_metropolis_hasting_samples_n_%d_thin_%d_prop_stddev_%f.pkl" % (
sampling_options["n_samples"], sampling_options["thinning_factor"],
sampling_options["proposal_stddev"])
f = open(output_pkl_name, "w")
cPickle.dump(results['samples'], f)
f.close()
print "Acceptance ratio : %f" % results["acceptance_ratio"]
print "Wrote %s" % (output_pkl_name, )
def main():
try:
opts, args = getopt.getopt(sys.argv[1:], "hv", [
"help", "n_samples=", "n_chains=", "thinning_factor=", "burn_in=",
"langevin_lambda=", "mcmc_method=", "proposal_stddev=",
"dataset_desc=", "output_dir_prefix="
])
except getopt.GetoptError as err:
# print help information and exit:
print str(err) # will print something like "option -a not recognized"
usage()
sys.exit(2)
sampling_options = {}
sampling_options["n_chains"] = None
output_options = {}
verbose = False
for o, a in opts:
if o == "-v":
# unused
verbose = True
elif o in ("-h", "--help"):
usage()
sys.exit()
elif o in ("--n_samples"):
sampling_options["n_samples"] = int(a)
elif o in ("--thinning_factor"):
sampling_options["thinning_factor"] = int(a)
elif o in ("--burn_in"):
sampling_options["burn_in"] = int(a)
elif o in ("--langevin_lambda"):
sampling_options["langevin_lambda"] = float(a)
elif o in ("--proposal_stddev"):
sampling_options["proposal_stddev"] = float(a)
elif o in ("--n_chains"):
sampling_options["n_chains"] = int(a)
elif o in ("--mcmc_method"):
sampling_options["mcmc_method"] = a
#if not a in ['langevin',
# 'metropolis_hastings_langevin_E',
# 'metropolis_hastings_langevin_grad_E',
# 'metropolis_hastings_E',
# 'metropolis_hastings_grad_E']:
# error("Bad name for mcmc_method.")
elif o in ("--dataset_desc"):
if a in ['ninja_star', 'Salah_DAE', 'guillaume_DAE']:
sampling_options["dataset_desc"] = a
else:
"Unrecognized dataset."
elif o in ("--output_dir_prefix"):
output_options['output_dir_prefix'] = a
else:
assert False, "unhandled option"
if sampling_options["dataset_desc"] == "ninja_star":
import ninja_star_distribution
sampling_options["E"] = ninja_star_distribution.E
sampling_options["grad_E"] = ninja_star_distribution.grad_E
if not sampling_options["n_chains"] == None:
sampling_options["x0"] = np.random.normal(
size=(sampling_options["n_chains"], 2))
else:
sampling_options["x0"] = np.random.normal(size=(2, ))
# this field should be obsolete
#output_options["cross_entropy_function"] = ninja_star_distribution.cross_entropy
elif sampling_options["dataset_desc"] == "butterfly":
import butterfly_distribution
sampling_options["E"] = butterfly_distribution.E
sampling_options["grad_E"] = butterfly_distribution.grad_E
if not sampling_options["n_chains"] == None:
sampling_options["x0"] = np.random.normal(
size=(sampling_options["n_chains"], 2))
else:
sampling_options["x0"] = np.random.normal(size=(2, ))
elif sampling_options["dataset_desc"] == "Salah_DAE":
import cPickle
salah_dae_params = cPickle.load(
open('/data/lisatmp2/rifaisal/share/sm_dae/ae_std=0.4.pkl'))
import dae_untied_weights
n_inputs = 784
n_hiddens = 1024
Wc = salah_dae_params['e_weights']
Wb = salah_dae_params['e_weights']
c = salah_dae_params['e_bias']
b = salah_dae_params['d_bias']
s = 1.0 / 16
the_dae = dae_untied_weights.DAE_untied_weights(
n_inputs=n_inputs,
n_hiddens=n_hiddens,
Wc=Wc,
Wb=Wb,
c=c,
b=b,
s=s,
act_func=['sigmoid', 'id'])
reference_langevin_lambda = 0.4**2
# overwrite whichever langevin lambda was given as argument
#sampling_options["langevin_lambda"] = reference_langevin_lambda
r = lambda x: the_dae.encode_decode(x.reshape((-1, n_inputs))).reshape(
(n_inputs, ))
grad_E = lambda x: -(r(x) - x) / reference_langevin_lambda
sampling_options["grad_E"] = grad_E
# METHOD 1
mnist_dataset = cPickle.load(
open('/data/lisa/data/mnist/mnist.pkl', 'rb'))
mnist_train = mnist_dataset[0]
mnist_train_data, mnist_train_labels = mnist_train
ind = np.random.randint(0, mnist_train_data.shape[0] - 1)
sampling_options["x0"] = mnist_train_data[ind, :]
print "Starting the simulation from MNIST digit %d" % mnist_train_labels[
ind]
# METHOD 3
#sampling_options["x0"] = np.random.uniform(low=0.0, high=1.0,size=(n_inputs,))
# these are the only valid sampling methods because they rely only on grad_E
assert sampling_options["mcmc_method"] in [
"langevin", "metropolis_hastings_grad_E",
"metropolis_hastings_langevin_grad_E"
]
elif sampling_options["dataset_desc"] == "guillaume_DAE":
import cPickle
#guillaume_dae_params = cPickle.load(open("/u/alaingui/umontreal/denoising_autoencoder/mcmc_pof/trained_models/mydae_2013_02_07.pkl"))
guillaume_dae_params = cPickle.load(
open(
"/u/alaingui/umontreal/denoising_autoencoder/mcmc_pof/trained_models/mydae_2013_02_08.pkl"
))
import dae_untied_weights
n_inputs = 784
n_hiddens = 1024
Wc = guillaume_dae_params['Wc']
Wb = guillaume_dae_params['Wb']
c = guillaume_dae_params['c']
b = guillaume_dae_params['b']
s = guillaume_dae_params['s']
act_func = guillaume_dae_params['act_func']
the_dae = dae_untied_weights.DAE_untied_weights(n_inputs=n_inputs,
n_hiddens=n_hiddens,
Wc=Wc,
Wb=Wb,
c=c,
b=b,
s=s,
act_func=act_func)
# Set to be like that because it's the value used to train
# the DAE originally. We've already got the pkl file hardcoded
# with the DAE parameters, so we might as well supply the lambda here.
reference_langevin_lambda = 0.01
#sampling_options["langevin_lambda"] = reference_langevin_lambda
r = lambda X: the_dae.encode_decode(X.reshape((-1, n_inputs))).reshape(
(n_inputs, ))
r_prime = lambda x: the_dae.jacobian_encode_decode(x)
f = lambda X: the_dae.encode(X.reshape((-1, n_inputs))).reshape(
(n_hiddens, ))
f_prime = lambda x: the_dae.jacobian_encode(x)
grad_E = lambda x: -(r(x) - x) / reference_langevin_lambda
sampling_options["grad_E"] = grad_E
sampling_options["r"] = r
sampling_options["r_prime"] = r_prime
sampling_options["f"] = f
sampling_options["f_prime"] = f_prime
# METHOD 1
mnist_dataset = cPickle.load(
open('/data/lisa/data/mnist/mnist.pkl', 'rb'))
mnist_train = mnist_dataset[0]
mnist_train_data, mnist_train_labels = mnist_train
ind = np.random.randint(0, mnist_train_data.shape[0] - 1)
sampling_options["x0"] = mnist_train_data[ind, :]
print "Starting the simulation from MNIST digit %d" % mnist_train_labels[
ind]
# METHOD 3
#sampling_options["x0"] = np.random.uniform(low=0.0, high=1.0,size=(n_inputs,))
# these are the only valid sampling methods because they rely only on grad_E
assert sampling_options["mcmc_method"] in [
"langevin", "metropolis_hastings_grad_E",
"metropolis_hastings_langevin_grad_E",
"metropolis_hastings_svd_grad_E"
]
else:
error("No dataset was supplied.")
results = metropolis_hastings_sampler.mcmc_generate_samples(
sampling_options)
#
# Returns a dictionary of this form.
#
# {'samples': numpy array (n_chains, n_samples, d),
# 'elapsed_time': seconds as float,
# 'proposals_per_second': float,
# 'acceptance_ratio': float in [0.0,1.0] }
print "Got the samples. Acceptance ratio was %f" % results[
'acceptance_ratio']
print "MCMC proposal speed was 10^%0.2f / s" % (
np.log(results['proposals_per_second']) / np.log(10), )
output_image_dir = "%s/%s/%d" % (output_options["output_dir_prefix"],
sampling_options['mcmc_method'],
int(time.time()))
os.makedirs(output_image_dir)
import cPickle
output_pkl_name = os.path.join(output_image_dir, "results_and_params.pkl")
for key in ["grad_E", "f", "f_prime", "r", "r_prime"]:
if sampling_options.has_key(key):
sampling_options[key] = "CANNOT BE PICKLED"
f = open(output_pkl_name, "w")
cPickle.dump(
{
'results': results,
'sampling_options': sampling_options,
'output_options': output_options
}, f)
f.close()
print "Wrote " + output_pkl_name
samples_only_pkl_name = os.path.join(output_image_dir, "samples.pkl")
f = open(samples_only_pkl_name, "w")
cPickle.dump(results['samples'], f)
f.close()
print "Wrote " + samples_only_pkl_name
if sampling_options["dataset_desc"] in ["Salah_DAE", "guillaume_DAE"]:
digits_image_file = os.path.join(output_image_dir, "digits.png")
plot_Salah_DAE_samples(results['samples'], digits_image_file)
def main():
try:
opts, args = getopt.getopt(
sys.argv[1:],
"hv",
[
"help",
"n_samples=",
"n_chains=",
"thinning_factor=",
"burn_in=",
"langevin_lambda=",
"mcmc_method=",
"proposal_stddev=",
"dataset_desc=",
"output_dir_prefix=",
],
)
except getopt.GetoptError as err:
# print help information and exit:
print str(err) # will print something like "option -a not recognized"
usage()
sys.exit(2)
sampling_options = {}
sampling_options["n_chains"] = None
output_options = {}
verbose = False
for o, a in opts:
if o == "-v":
# unused
verbose = True
elif o in ("-h", "--help"):
usage()
sys.exit()
elif o in ("--n_samples"):
sampling_options["n_samples"] = int(a)
elif o in ("--thinning_factor"):
sampling_options["thinning_factor"] = int(a)
elif o in ("--burn_in"):
sampling_options["burn_in"] = int(a)
elif o in ("--langevin_lambda"):
sampling_options["langevin_lambda"] = float(a)
elif o in ("--proposal_stddev"):
sampling_options["proposal_stddev"] = float(a)
elif o in ("--n_chains"):
sampling_options["n_chains"] = int(a)
elif o in ("--mcmc_method"):
sampling_options["mcmc_method"] = a
# if not a in ['langevin',
# 'metropolis_hastings_langevin_E',
# 'metropolis_hastings_langevin_grad_E',
# 'metropolis_hastings_E',
# 'metropolis_hastings_grad_E']:
# error("Bad name for mcmc_method.")
elif o in ("--dataset_desc"):
if a in ["ninja_star", "Salah_DAE", "guillaume_DAE"]:
sampling_options["dataset_desc"] = a
else:
"Unrecognized dataset."
elif o in ("--output_dir_prefix"):
output_options["output_dir_prefix"] = a
else:
assert False, "unhandled option"
if sampling_options["dataset_desc"] == "ninja_star":
import ninja_star_distribution
sampling_options["E"] = ninja_star_distribution.E
sampling_options["grad_E"] = ninja_star_distribution.grad_E
if not sampling_options["n_chains"] == None:
sampling_options["x0"] = np.random.normal(size=(sampling_options["n_chains"], 2))
else:
sampling_options["x0"] = np.random.normal(size=(2,))
# this field should be obsolete
# output_options["cross_entropy_function"] = ninja_star_distribution.cross_entropy
elif sampling_options["dataset_desc"] == "butterfly":
import butterfly_distribution
sampling_options["E"] = butterfly_distribution.E
sampling_options["grad_E"] = butterfly_distribution.grad_E
if not sampling_options["n_chains"] == None:
sampling_options["x0"] = np.random.normal(size=(sampling_options["n_chains"], 2))
else:
sampling_options["x0"] = np.random.normal(size=(2,))
elif sampling_options["dataset_desc"] == "Salah_DAE":
import cPickle
salah_dae_params = cPickle.load(open("/data/lisatmp2/rifaisal/share/sm_dae/ae_std=0.4.pkl"))
import dae_untied_weights
n_inputs = 784
n_hiddens = 1024
Wc = salah_dae_params["e_weights"]
Wb = salah_dae_params["e_weights"]
c = salah_dae_params["e_bias"]
b = salah_dae_params["d_bias"]
s = 1.0 / 16
the_dae = dae_untied_weights.DAE_untied_weights(
n_inputs=n_inputs, n_hiddens=n_hiddens, Wc=Wc, Wb=Wb, c=c, b=b, s=s, act_func=["sigmoid", "id"]
)
reference_langevin_lambda = 0.4 ** 2
# overwrite whichever langevin lambda was given as argument
# sampling_options["langevin_lambda"] = reference_langevin_lambda
r = lambda x: the_dae.encode_decode(x.reshape((-1, n_inputs))).reshape((n_inputs,))
grad_E = lambda x: -(r(x) - x) / reference_langevin_lambda
sampling_options["grad_E"] = grad_E
# METHOD 1
mnist_dataset = cPickle.load(open("/data/lisa/data/mnist/mnist.pkl", "rb"))
mnist_train = mnist_dataset[0]
mnist_train_data, mnist_train_labels = mnist_train
ind = np.random.randint(0, mnist_train_data.shape[0] - 1)
sampling_options["x0"] = mnist_train_data[ind, :]
print "Starting the simulation from MNIST digit %d" % mnist_train_labels[ind]
# METHOD 3
# sampling_options["x0"] = np.random.uniform(low=0.0, high=1.0,size=(n_inputs,))
# these are the only valid sampling methods because they rely only on grad_E
assert sampling_options["mcmc_method"] in [
"langevin",
"metropolis_hastings_grad_E",
"metropolis_hastings_langevin_grad_E",
]
elif sampling_options["dataset_desc"] == "guillaume_DAE":
import cPickle
# guillaume_dae_params = cPickle.load(open("/u/alaingui/umontreal/denoising_autoencoder/mcmc_pof/trained_models/mydae_2013_02_07.pkl"))
guillaume_dae_params = cPickle.load(
open("/u/alaingui/umontreal/denoising_autoencoder/mcmc_pof/trained_models/mydae_2013_02_08.pkl")
)
import dae_untied_weights
n_inputs = 784
n_hiddens = 1024
Wc = guillaume_dae_params["Wc"]
Wb = guillaume_dae_params["Wb"]
c = guillaume_dae_params["c"]
b = guillaume_dae_params["b"]
s = guillaume_dae_params["s"]
act_func = guillaume_dae_params["act_func"]
the_dae = dae_untied_weights.DAE_untied_weights(
n_inputs=n_inputs, n_hiddens=n_hiddens, Wc=Wc, Wb=Wb, c=c, b=b, s=s, act_func=act_func
)
# Set to be like that because it's the value used to train
# the DAE originally. We've already got the pkl file hardcoded
# with the DAE parameters, so we might as well supply the lambda here.
reference_langevin_lambda = 0.01
# sampling_options["langevin_lambda"] = reference_langevin_lambda
r = lambda X: the_dae.encode_decode(X.reshape((-1, n_inputs))).reshape((n_inputs,))
r_prime = lambda x: the_dae.jacobian_encode_decode(x)
f = lambda X: the_dae.encode(X.reshape((-1, n_inputs))).reshape((n_hiddens,))
f_prime = lambda x: the_dae.jacobian_encode(x)
grad_E = lambda x: -(r(x) - x) / reference_langevin_lambda
sampling_options["grad_E"] = grad_E
sampling_options["r"] = r
sampling_options["r_prime"] = r_prime
sampling_options["f"] = f
sampling_options["f_prime"] = f_prime
# METHOD 1
mnist_dataset = cPickle.load(open("/data/lisa/data/mnist/mnist.pkl", "rb"))
mnist_train = mnist_dataset[0]
mnist_train_data, mnist_train_labels = mnist_train
ind = np.random.randint(0, mnist_train_data.shape[0] - 1)
sampling_options["x0"] = mnist_train_data[ind, :]
print "Starting the simulation from MNIST digit %d" % mnist_train_labels[ind]
# METHOD 3
# sampling_options["x0"] = np.random.uniform(low=0.0, high=1.0,size=(n_inputs,))
# these are the only valid sampling methods because they rely only on grad_E
assert sampling_options["mcmc_method"] in [
"langevin",
"metropolis_hastings_grad_E",
"metropolis_hastings_langevin_grad_E",
"metropolis_hastings_svd_grad_E",
]
else:
error("No dataset was supplied.")
results = metropolis_hastings_sampler.mcmc_generate_samples(sampling_options)
#
# Returns a dictionary of this form.
#
# {'samples': numpy array (n_chains, n_samples, d),
# 'elapsed_time': seconds as float,
# 'proposals_per_second': float,
# 'acceptance_ratio': float in [0.0,1.0] }
print "Got the samples. Acceptance ratio was %f" % results["acceptance_ratio"]
print "MCMC proposal speed was 10^%0.2f / s" % (np.log(results["proposals_per_second"]) / np.log(10),)
output_image_dir = "%s/%s/%d" % (
output_options["output_dir_prefix"],
sampling_options["mcmc_method"],
int(time.time()),
)
os.makedirs(output_image_dir)
import cPickle
output_pkl_name = os.path.join(output_image_dir, "results_and_params.pkl")
for key in ["grad_E", "f", "f_prime", "r", "r_prime"]:
if sampling_options.has_key(key):
sampling_options[key] = "CANNOT BE PICKLED"
f = open(output_pkl_name, "w")
cPickle.dump({"results": results, "sampling_options": sampling_options, "output_options": output_options}, f)
f.close()
print "Wrote " + output_pkl_name
samples_only_pkl_name = os.path.join(output_image_dir, "samples.pkl")
f = open(samples_only_pkl_name, "w")
cPickle.dump(results["samples"], f)
f.close()
print "Wrote " + samples_only_pkl_name
if sampling_options["dataset_desc"] in ["Salah_DAE", "guillaume_DAE"]:
digits_image_file = os.path.join(output_image_dir, "digits.png")
plot_Salah_DAE_samples(results["samples"], digits_image_file)
def main():
try:
opts, args = getopt.getopt(sys.argv[1:], "hv", ["help", "omit_ninja_star_from_plots", "n_samples=", "n_chains=", "thinning_factor=", "burn_in=", "langevin_lambda=", "mcmc_method=", "proposal_stddev=", "dataset=", "output_dir_prefix=", "reference_pickled_samples_for_KL=", "reference_stddev_for_KL=", "no_plots"])
except getopt.GetoptError as err:
# print help information and exit:
print str(err) # will print something like "option -a not recognized"
usage()
sys.exit(2)
sampling_options = {}
sampling_options["n_chains"] = None
sampling_options["no_plots"] = False
output_options = {}
verbose = False
for o, a in opts:
if o == "-v":
# unused
verbose = True
elif o in ("-h", "--help"):
usage()
sys.exit()
if o == "--no_plots":
sampling_options["no_plots"] = True
elif o in ("--n_samples"):
sampling_options["n_samples"] = int(a)
elif o in ("--thinning_factor"):
sampling_options["thinning_factor"] = int(a)
elif o in ("--burn_in"):
sampling_options["burn_in"] = int(a)
elif o in ("--langevin_lambda"):
sampling_options["langevin_lambda"] = float(a)
elif o in ("--proposal_stddev"):
sampling_options["proposal_stddev"] = float(a)
elif o in ("--n_chains"):
sampling_options["n_chains"] = int(a)
elif o in ("--mcmc_method"):
sampling_options["mcmc_method"] = a
#if not a in ['langevin',
# 'metropolis_hastings_langevin_E',
# 'metropolis_hastings_langevin_grad_E',
# 'metropolis_hastings_E',
# 'metropolis_hastings_grad_E']:
# error("Bad name for mcmc_method.")
elif o in ("--dataset"):
if a == 'ninja_star':
import ninja_star_distribution
sampling_options["dataset_description"] = "ninja_star"
sampling_options["E"] = ninja_star_distribution.E
sampling_options["grad_E"] = ninja_star_distribution.grad_E
if a == 'butterfly':
import butterfly_distribution
sampling_options["dataset_description"] = "butterfly"
sampling_options["E"] = butterfly_distribution.E
sampling_options["grad_E"] = butterfly_distribution.grad_E
else:
"Unrecognized dataset."
elif o in ("--output_dir_prefix"):
output_options['output_dir_prefix'] = a
elif o in ("--reference_pickled_samples_for_KL"):
output_options['reference_pickled_samples_for_KL'] = a
elif o in ("--reference_stddev_for_KL"):
output_options['reference_stddev_for_KL'] = float(a)
elif o in ("--omit_ninja_star_from_plots"):
output_options['omit_ninja_star_from_plots'] = True
else:
assert False, "unhandled option"
if sampling_options["dataset_description"] == "ninja_star":
if not sampling_options["n_chains"] == None:
sampling_options["x0"] = np.random.normal(size=(sampling_options["n_chains"],2))
else:
sampling_options["x0"] = np.random.normal(size=(2,))
# osbsolete
#output_options["cross_entropy_function"] = ninja_star_distribution.cross_entropy
elif sampling_options["dataset_description"] == "butterfly":
if not sampling_options["n_chains"] == None:
sampling_options["x0"] = np.random.normal(size=(sampling_options["n_chains"],2))
else:
sampling_options["x0"] = np.random.normal(size=(2,))
else:
error("No dataset was supplied.")
results = metropolis_hastings_sampler.mcmc_generate_samples(sampling_options)
#
# Returns a dictionary of this form.
#
# {'samples': numpy array (n_chains, n_samples, d),
# 'elapsed_time': seconds as float,
# 'proposals_per_second': float,
# 'acceptance_ratio': float in [0.0,1.0] }
print "Got the samples. Acceptance ratio was %f" % results['acceptance_ratio']
print "MCMC proposal speed was 10^%0.2f / s" % (np.log(results['proposals_per_second']) / np.log(10), )
if len(results['samples'].shape) == 2:
# We will evaluate an approximate KL divergence if we are given a
# reference set of samples from the true distribution.
if output_options.has_key("reference_pickled_samples_for_KL"):
import KL_approximation
import cPickle
assert os.path.exists(output_options["reference_pickled_samples_for_KL"])
f = open(output_options["reference_pickled_samples_for_KL"])
reference_sample = cPickle.load(f)
# Let's leave out that argument for now.
#if output_options.has_key("reference_stddev_for_KL"):
# reference_stddev_for_KL = output_options["reference_stddev_for_KL"]
#else:
# reference_stddev_for_KL = None
KL_value = KL_approximation.KL(reference_sample, results['samples'], 0.5, 0.5)
print "We got a KL divergence value of %f" % KL_value
output_image_dir = "%s/%s/%d" % (output_options["output_dir_prefix"], sampling_options['mcmc_method'], int(time.time()) )
os.makedirs(output_image_dir)
import cPickle
output_pkl_name = os.path.join(output_image_dir, "results_and_params.pkl")
f = open(output_pkl_name, "w")
if sampling_options.has_key("grad_E"):
sampling_options["grad_E"] = "CANNOT BE PICKLED"
cPickle.dump({'results':results, 'sampling_options':sampling_options, 'output_options':output_options}, f)
f.close()
print "Wrote " + output_pkl_name
samples_only_pkl_name = os.path.join(output_image_dir, "samples.pkl")
f = open(samples_only_pkl_name, "w")
cPickle.dump(results['samples'], f)
f.close()
print "Wrote " + samples_only_pkl_name
if sampling_options["dataset_description"] == "ninja_star" and not (sampling_options["no_plots"]):
if len(results['samples'].shape) == 2:
output_image_path = os.path.join(output_image_dir, "whole_chain.png")
plot_one_slice_of_ninja_star_samples(results['samples'],
output_image_path,
dpi=200,
omit_ninja_star_from_plots = output_options.has_key('omit_ninja_star_from_plots'))
elif len(results['samples'].shape) == 3:
#plot_one_slice_of_ninja_star_samples_2(results['samples'],
# lambda n : os.path.join(output_image_dir, "frame_%0.6d.png" % n),
# dpi=100)
for n in np.arange(sampling_options['n_samples']):
output_image_path = os.path.join(output_image_dir, "frame_%0.6d.png" % n)
plot_one_slice_of_ninja_star_samples(results['samples'][:,n,:],
output_image_path,
dpi=100,
omit_ninja_star_from_plots = output_options.has_key('omit_ninja_star_from_plots'))
import subprocess
#export MOVIEDIR=${HOME}/Documents/tmp/metropolis_hastings_E/1359965409
#ffmpeg -i ${MOVIEDIR}/frame_%06d.png -vcodec libx264 -vpre hq -crf 22 -r 10 ${MOVIEDIR}/seq.mp4
subprocess.check_output(["ffmpeg", "-i", output_image_dir + r"/frame_%06d.png", "-y", "-vcodec", "libx264", "-vpre", "hq", "-crf", "22", "-r", "10", output_image_dir + r"/assembled.mp4"])
print "Generated movie at %s." % (output_image_dir + r"/assembled.mp4", )
else:
raise("Wrong shape for samples returned !")
