def get_model(model_path):
hyperparams = utils.load_dict_from_json_file(os.path.join(model_path, "hyperparams"))
hyperparams['weights_initialization'] = "Zeros"
trainingparams = utils.load_dict_from_json_file(os.path.join(model_path, "trainingparams"))
dataset_name = trainingparams['dataset_name']
if dataset_name != "binarized_mnist":
raise ValueError("Invalid dataset. Only model trained on MNIST supported.")
#
# LOAD DATASET ####
dataset = Dataset.get(dataset_name)
if trainingparams['batch_size'] == -1:
trainingparams['batch_size'] = dataset['train']['length']
model = build_model(dataset, trainingparams, hyperparams, hyperparams['hidden_sizes'])
print("### Loading model | Hidden:{0} CondMask:{1} Activation:{2} ... ".format(
hyperparams['hidden_sizes'], hyperparams['use_cond_mask'], hyperparams['hidden_activation']), end=' ')
start_time = t.time()
load_model_params(model, model_path)
print(utils.get_done_text(start_time), "###")
return model, dataset_name, dataset
def get_model(model_path):
hyperparams = utils.load_dict_from_json_file(
os.path.join(model_path, "hyperparams"))
hyperparams['weights_initialization'] = "Zeros"
trainingparams = utils.load_dict_from_json_file(
os.path.join(model_path, "trainingparams"))
dataset_name = trainingparams['dataset_name']
if dataset_name != "binarized_mnist":
raise (Exception(
"Invalid dataset. Only model trained on MNIST supported."))
#
# LOAD DATASET ####
dataset = Dataset.get(dataset_name)
if trainingparams['batch_size'] == -1:
trainingparams['batch_size'] = dataset['train']['length']
model = build_model(dataset, trainingparams, hyperparams,
hyperparams['hidden_sizes'])
print "### Loading model | Hidden:{0} CondMask:{1} Activation:{2} ... ".format(
hyperparams['hidden_sizes'], hyperparams['use_cond_mask'],
hyperparams['hidden_activation']),
start_time = t.time()
load_model_params(model, model_path)
print utils.get_done_text(start_time), "###"
return model, dataset_name, dataset
def load(cls, path):
meta = load_dict_from_json_file(pjoin(path, "meta.json"))
assert meta['name'] == cls.__name__
hyperparams = load_dict_from_json_file(pjoin(path, "hyperparams.json"))
model = cls(**hyperparams)
parameters = np.load(pjoin(path, "params.npz"))
for param_name, param in model.parameters.items():
param.set_value(parameters[param_name])
return model
#
# Set the name of the experiment (remove the --force from the args to make sure it will generate the same uid)
if '--force' in sys.argv:
sys.argv.remove('--force')
experiment_name = args.name if args.name is not None else utils.generate_uid_from_string(
' '.join(sys.argv))
#
# Creating the experiments folder or resuming experiment
save_path_experiment = os.path.join('./experiments/', experiment_name)
if os.path.isdir(save_path_experiment):
if not args.force:
print "### Resuming experiment ({0}). ###\n".format(
experiment_name)
loaded_hyperparams = utils.load_dict_from_json_file(
os.path.join(save_path_experiment, "hyperparams"))
loaded_trainingparams = utils.load_dict_from_json_file(
os.path.join(save_path_experiment, "trainingparams"))
if loaded_trainingparams != trainingparams or loaded_hyperparams != hyperparams:
print "The arguments provided are different than the one saved. Use --force if you are certain.\nQuitting."
exit()
resume_mode = True
else:
os.makedirs(save_path_experiment)
utils.save_dict_to_json_file(
os.path.join(save_path_experiment, "hyperparams"), hyperparams)
utils.save_dict_to_json_file(
os.path.join(save_path_experiment, "trainingparams"),
hyperparams = vars(args.model)
trainingparams = vars(args.train)
#
# Set the name of the experiment (remove the --force from the args to make sure it will generate the same uid)
if '--force' in sys.argv:
sys.argv.remove('--force')
experiment_name = args.name if args.name is not None else utils.generate_uid_from_string(' '.join(sys.argv))
#
# Creating the experiments folder or resuming experiment
save_path_experiment = os.path.join('./experiments/', experiment_name)
if os.path.isdir(save_path_experiment):
if not args.force:
print "### Resuming experiment ({0}). ###\n".format(experiment_name)
loaded_hyperparams = utils.load_dict_from_json_file(os.path.join(save_path_experiment, "hyperparams"))
loaded_trainingparams = utils.load_dict_from_json_file(os.path.join(save_path_experiment, "trainingparams"))
if loaded_trainingparams != trainingparams or loaded_hyperparams != hyperparams:
print "The arguments provided are different than the one saved. Use --force if you are certain.\nQuitting."
exit()
resume_mode = True
else:
os.makedirs(save_path_experiment)
utils.save_dict_to_json_file(os.path.join(save_path_experiment, "hyperparams"), hyperparams)
utils.save_dict_to_json_file(os.path.join(save_path_experiment, "trainingparams"), trainingparams)
#
# LOAD DATASET ####
def _compute_AIS(model, M=100, betas=BETAS, batch_size=None, seed=1234, ais_working_dir=".", force=False):
ais_results_json = pjoin(ais_working_dir, "ais_results.part.json")
if batch_size is None:
batch_size = M
# Will be executing M AIS's runs.
last_sample_chain = np.zeros((M, model.input_size), dtype=config.floatX)
M_log_w_ais = np.zeros(M, dtype=np.float64)
model.set_rng_seed(seed)
ais_results = {}
if os.path.isfile(ais_results_json) and not force:
print "Resuming AIS using info from {}".format(ais_results_json)
ais_results = utils.load_dict_from_json_file(ais_results_json)
M_log_w_ais = ais_results['M_log_w_ais']
last_sample_chain = ais_results['last_sample_chain']
lnZ_trivial = ais_results['lnZ_trivial']
# Iterate through all AIS runs.
for i in range(0, M, batch_size):
if i <= ais_results.get('batch_id', -1):
continue
model.set_rng_seed(seed+i)
actual_size = min(M - i, batch_size)
print "AIS run: {}/{} (using batch size of {})".format(i, M, batch_size)
ais_partial_results = _compute_AIS_samples(model, M=actual_size, betas=betas)
M_log_w_ais[i:i+batch_size] = ais_partial_results['M_log_w_ais']
last_sample_chain[i:i+batch_size] = ais_partial_results['last_sample_chain']
lnZ_trivial = ais_partial_results['lnZ_trivial']
# Save partial results
if os.path.isfile(ais_results_json):
shutil.copy(ais_results_json, ais_results_json[:-4] + "old.json")
ais_results = {'batch_id': i,
'M': M,
'batch_size': batch_size,
'last_sample_chain': last_sample_chain,
'M_log_w_ais': M_log_w_ais,
'lnZ_trivial': lnZ_trivial}
utils.save_dict_to_json_file(ais_results_json, ais_results)
# We compute the mean of the estimated `r_AIS`
Ms = np.arange(1, M+1)
log_sum_w_ais = np.logaddexp.accumulate(M_log_w_ais)
logcummean_Z = log_sum_w_ais - np.log(Ms)
# We compute the standard deviation of the estimated `r_AIS`
logstd_AIS = np.zeros_like(M_log_w_ais)
for k in Ms[1:]:
m = np.max(M_log_w_ais[:k])
logstd_AIS[k-1] = np.log(np.std(np.exp(M_log_w_ais[:k]-m), ddof=1)) - np.log(np.sqrt(k))
logstd_AIS[k-1] += m
logstd_AIS[0] = np.nan # Standard deviation of only one sample does not exist.
# The authors report AIS error using ln(Ẑ ± 3\sigma)
m = max(np.nanmax(logstd_AIS), np.nanmax(logcummean_Z))
logcumstd_Z_up = np.log(np.exp(logcummean_Z-m) + 3*np.exp(logstd_AIS-m)) + m - logcummean_Z
logcumstd_Z_down = -(np.log(np.exp(logcummean_Z-m) - 3*np.exp(logstd_AIS-m)) + m) + logcummean_Z
# Compute the standard deviation of ln(Z)
std_lnZ = np.array([np.std(M_log_w_ais[:k], ddof=1) for k in Ms[1:]])
std_lnZ = np.r_[np.nan, std_lnZ] # Standard deviation of only one sample does not exist.
return {"logcummean_Z": logcummean_Z.astype(config.floatX),
"logcumstd_Z_down": logcumstd_Z_down.astype(config.floatX),
"logcumstd_Z_up": logcumstd_Z_up.astype(config.floatX),
"logcumstd_Z": logstd_AIS.astype(config.floatX),
"M_log_w_ais": M_log_w_ais,
"lnZ_trivial": lnZ_trivial,
"std_lnZ": std_lnZ,
"last_sample_chain": last_sample_chain,
"batch_size": batch_size,
"seed": seed,
"nb_temperatures": len(betas),
"nb_samples": M
}
