def debug_laplace():
# Load data set
X, y, D, Xtest, ytest = gpml.load_mat(
'../data/kfold_data/r_concrete_500_fold_10_of_10.mat', y_dim=1)
# Load the suspicious kernel
sk = fk.repr_string_to_kernel(
'ScoredKernel(k_opt=ProductKernel([ MaskKernel(ndim=8, active_dimension=0, base_kernel=CubicKernel(offset=1.757755, output_variance=7.084045)), MaskKernel(ndim=8, active_dimension=7, base_kernel=SqExpPeriodicKernel(lengthscale=-2.701080, period=-0.380918, output_variance=-0.071214)) ]), nll=6348.096611, laplace_nle=-184450132.068237, bic_nle=12720.630212, noise=[-1.77276072])'
)
# Create some code to evaluate it
if X.ndim == 1: X = X[:, nax]
if y.ndim == 1: y = y[:, nax]
ndata = y.shape[0]
# Create data file
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X, 'y': y}) # Save regression data
# Move to fear
cblparallel.copy_to_remote(data_file)
scripts = [
gpml.OPTIMIZE_KERNEL_CODE % {
'datafile':
data_file.split('/')[-1],
'writefile':
'%(output_file)s', # N.B. cblparallel manages output files
'gpml_path':
cblparallel.gpml_path(local_computation=False),
'kernel_family':
sk.k_opt.gpml_kernel_expression(),
'kernel_params':
'[ %s ]' % ' '.join(str(p) for p in sk.k_opt.param_vector()),
'noise':
str(sk.noise),
'iters':
str(300)
}
]
#### Need to be careful with % signs
#### For the moment, cblparallel expects no single % signs - FIXME
scripts[0] = re.sub('% ', '%% ', scripts[0])
# Test
scripts[0] = re.sub('delta = 1e-6', 'delta = 1e-6', scripts[0])
#scripts[0] = re.sub('hyp.lik = [-1.77276072]', 'hyp.lik = [-0.77276072]', scripts[0])
output_file = cblparallel.run_batch_on_fear(scripts,
language='matlab',
max_jobs=600)[0]
# Read in results
output = gpml.read_outputs(output_file)
result = ScoredKernel.from_matlab_output(output, sk.k_opt.family(), ndata)
print result
print output.hessian
os.remove(output_file)
# Remove temporary data file (perhaps on the cluster server)
cblparallel.remove_temp_file(data_file, local_computation=False)
def make_predictions(X, y, Xtest, ytest, best_scored_kernel, local_computation=False, max_jobs=500, verbose=True, zero_mean=False, random_seed=0):
'''
Evaluates a kernel on held out data
Input:
- X - A matrix (data_points x dimensions) of input locations
- y - A matrix (data_points x 1) of output values
- Xtest - Held out X data
- ytest - Held out y data
- best_scored_kernel - A Scored Kernel object to be evaluated on the held out data
- ...
Return:
- A dictionary of results from the MATLAB script containing:
- loglik - an array of log likelihoods of test data
- predictions - an array of mean predictions for the held out data
- actuals - ytest
- model - I'm not sure FIXME
- timestamp - A time stamp of some sort
'''
# Make data into matrices in case they're unidimensional.
if X.ndim == 1: X = X[:, nax]
if y.ndim == 1: y = y[:, nax]
ndata = y.shape[0]
# Save temporary data file in standard temporary directory
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X, 'y': y, 'Xtest' : Xtest, 'ytest' : ytest})
# Copy onto cluster server if necessary
if not local_computation:
if verbose:
print 'Moving data file to fear'
cblparallel.copy_to_remote(data_file)
# Create prediction code
parameters ={'datafile': data_file.split('/')[-1],
'writefile': '%(output_file)s',
'gpml_path': cblparallel.gpml_path(local_computation),
'kernel_family': best_scored_kernel.k_opt.gpml_kernel_expression(),
'kernel_params': '[ %s ]' % ' '.join(str(p) for p in best_scored_kernel.k_opt.param_vector()),
'noise': str(best_scored_kernel.noise),
'iters': str(30),
'seed': str(random_seed)}
if zero_mean:
code = gpml.PREDICT_AND_SAVE_CODE_ZERO_MEAN % parameters
else:
code = gpml.PREDICT_AND_SAVE_CODE % parameters
code = re.sub('% ', '%% ', code) # HACK - cblparallel currently does not like % signs
# Evaluate code - potentially on cluster
if local_computation:
temp_results_file = cblparallel.run_batch_locally([code], language='matlab', max_cpu=1.1, max_mem=1.1, verbose=verbose)[0]
else:
temp_results_file = cblparallel.run_batch_on_fear([code], language='matlab', max_jobs=max_jobs, verbose=verbose)[0]
results = scipy.io.loadmat(temp_results_file)
# Remove temporary files (perhaps on the cluster server)
cblparallel.remove_temp_file(temp_results_file, local_computation)
cblparallel.remove_temp_file(data_file, local_computation)
# Return dictionary of MATLAB results
return results
def make_predictions(
X, y, Xtest, ytest, model, local_computation=False, max_jobs=500, verbose=True, random_seed=0, no_noise=False
):
# Make data into matrices in case they're unidimensional.
if X.ndim == 1:
X = X[:, nax]
if y.ndim == 1:
y = y[:, nax]
ndata = y.shape[0]
# Save temporary data file in standard temporary directory
data_file = cblparallel.create_temp_file(".mat")
scipy.io.savemat(data_file, {"X": X, "y": y, "Xtest": Xtest, "ytest": ytest})
# Copy onto cluster server if necessary
if not local_computation:
if verbose:
print "Moving data file to fear"
cblparallel.copy_to_remote(data_file)
# Create prediction code
parameters = {
"datafile": data_file.split("/")[-1],
"writefile": "%(output_file)s",
"gpml_path": cblparallel.gpml_path(local_computation),
"mean_syntax": model.mean.get_gpml_expression(dimensions=X.shape[1]),
"mean_params": "[ %s ]" % " ".join(str(p) for p in model.mean.param_vector),
"kernel_syntax": model.kernel.get_gpml_expression(dimensions=X.shape[1]),
"kernel_params": "[ %s ]" % " ".join(str(p) for p in model.kernel.param_vector),
"lik_syntax": model.likelihood.get_gpml_expression(dimensions=X.shape[1]),
"lik_params": "[ %s ]" % " ".join(str(p) for p in model.likelihood.param_vector),
"inference": model.likelihood.gpml_inference_method,
"iters": str(30),
"seed": str(random_seed),
}
code = gpml.PREDICT_AND_SAVE_CODE % parameters
code = re.sub("% ", "%% ", code) # HACK - cblparallel currently does not like % signs
# Evaluate code - potentially on cluster
if local_computation:
temp_results_file = cblparallel.run_batch_locally(
[code], language="matlab", max_cpu=1.1, max_mem=1.1, verbose=verbose
)[0]
else:
temp_results_file = cblparallel.run_batch_on_fear(
[code], language="matlab", max_jobs=max_jobs, verbose=verbose
)[0]
results = scipy.io.loadmat(temp_results_file)
# Remove temporary files (perhaps on the cluster server)
cblparallel.remove_temp_file(temp_results_file, local_computation)
cblparallel.remove_temp_file(data_file, local_computation)
# Return dictionary of MATLAB results
return results
def rf_fear_test_home(n=10,n_trees=10):
cblparallel.start_port_forwarding()
# Data
X, y = make_friedman1(n_samples=1200, random_state=0, noise=1.0)
X_train, X_test = X[:200], X[200:]
y_train, y_test = y[:200], y[200:]
# Params
#local_temp_path = os.path.abspath('../temp/')
#remote_temp_path = 'python/'
# Write data file locally
#data_file = mkstemp_safe(cblparallel.config.LOCAL_TEMP_PATH, '.p')
data_file = mkstemp_safe(cblparallel.config.HOME_TEMP_PATH, '.p')
with open(data_file, 'w') as f:
pickle.dump((X_train, y_train, X_test), f)
# Prepare code
scripts = [reduced_tree_code % {'data_file' : os.path.join(cblparallel.config.REMOTE_TEMP_PATH, os.path.split(data_file)[-1]),
'n_trees' : n_trees,
'random_state' : i * n_trees,
'output_file' : '%(output_file)s',
'flag_file' : '%(flag_file)s'} for i in range(n)]
# Submit to fear
with cblparallel.fear(via_gate=True) as fear:
fear.copy_to(data_file, os.path.join(cblparallel.config.REMOTE_TEMP_PATH, os.path.split(data_file)[-1]))
output_files = cblparallel.run_batch_on_fear(scripts, max_jobs=1000)
fear.rm(os.path.join(cblparallel.config.REMOTE_TEMP_PATH, os.path.split(data_file)[-1]))
# Kill local data file
os.remove(data_file)
# Now do something with the output
estimators = []
predictions = []
for output_file in output_files:
with open(output_file, 'r') as f:
#(estimator, prediction) = pickle.load(f)
prediction = np.genfromtxt(output_file, delimiter=',')
os.remove(output_file)
#estimators.append(estimator)
predictions.append(prediction)
#ens = EnsembleRegressor(estimators)
#return RMSE(X_test, y_test, ens)
ens_pred = np.mean(predictions, axis=0)
return RMSE_y(y_test, ens_pred)
def remote_matlab_test(n=10):
cblparallel.start_port_forwarding()
# Prepare code
scripts = [matlab_code] * n
# Run bacth in parallel
output_files = cblparallel.run_batch_on_fear(scripts, language='matlab', max_jobs=1000)
# Now do something with the output
estimators = []
for output_file in output_files:
with open(output_file, 'r') as f:
estimator = np.genfromtxt(output_file, delimiter=',')
os.remove(output_file)
estimators.append(estimator)
ens_pred = np.mean(estimators)
return ens_pred
def debug_laplace():
# Load data set
X, y, D, Xtest, ytest = gpml.load_mat('../data/kfold_data/r_concrete_500_fold_10_of_10.mat', y_dim=1)
# Load the suspicious kernel
sk = fk.repr_string_to_kernel('ScoredKernel(k_opt=ProductKernel([ MaskKernel(ndim=8, active_dimension=0, base_kernel=CubicKernel(offset=1.757755, output_variance=7.084045)), MaskKernel(ndim=8, active_dimension=7, base_kernel=SqExpPeriodicKernel(lengthscale=-2.701080, period=-0.380918, output_variance=-0.071214)) ]), nll=6348.096611, laplace_nle=-184450132.068237, bic_nle=12720.630212, noise=[-1.77276072])')
# Create some code to evaluate it
if X.ndim == 1: X = X[:, nax]
if y.ndim == 1: y = y[:, nax]
ndata = y.shape[0]
# Create data file
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X, 'y': y}) # Save regression data
# Move to fear
cblparallel.copy_to_remote(data_file)
scripts = [gpml.OPTIMIZE_KERNEL_CODE % {'datafile': data_file.split('/')[-1],
'writefile': '%(output_file)s', # N.B. cblparallel manages output files
'gpml_path': cblparallel.gpml_path(local_computation=False),
'kernel_family': sk.k_opt.gpml_kernel_expression(),
'kernel_params': '[ %s ]' % ' '.join(str(p) for p in sk.k_opt.param_vector()),
'noise': str(sk.noise),
'iters': str(300)}]
#### Need to be careful with % signs
#### For the moment, cblparallel expects no single % signs - FIXME
scripts[0] = re.sub('% ', '%% ', scripts[0])
# Test
scripts[0] = re.sub('delta = 1e-6', 'delta = 1e-6', scripts[0])
#scripts[0] = re.sub('hyp.lik = [-1.77276072]', 'hyp.lik = [-0.77276072]', scripts[0])
output_file = cblparallel.run_batch_on_fear(scripts, language='matlab', max_jobs=600)[0]
# Read in results
output = gpml.read_outputs(output_file)
result = ScoredKernel.from_matlab_output(output, sk.k_opt.family(), ndata)
print result
print output.hessian
os.remove(output_file)
# Remove temporary data file (perhaps on the cluster server)
cblparallel.remove_temp_file(data_file, local_computation=False)
def covariance_distance(kernels, X, local_computation=True, verbose=True):
'''
Evaluate a distance matrix of kernels, in terms of their covariance matrix evaluated on training inputs
Input:
- kernels - A list of fk.ScoredKernel
- X - A matrix (data_points x dimensions) of input locations
- local_computation - Boolean indicating if computation should be performed on cluster or on local machine
Return:
- A matrix of similarities between the input kernels
'''
assert(len(kernels) > 0) #### FIXME - This sort of check should happen earlier
# Make data into matrices in case they're unidimensional.
if X.ndim == 1: X = X[:, nax]
# Save temporary data file in standard temporary directory
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X})
# Copy onto cluster server if necessary
if not local_computation:
if verbose:
print 'Moving data file to fear'
cblparallel.copy_to_remote(data_file)
# Construct testing code
code = gpml.DISTANCE_CODE_HEADER % {'datafile': data_file.split('/')[-1],
'gpml_path': cblparallel.gpml_path(local_computation)}
for (i, kernel) in enumerate([k.k_opt for k in kernels]):
code = code + gpml.DISTANCE_CODE_COV % {'iter' : i + 1,
'kernel_family': kernel.gpml_kernel_expression(),
'kernel_params': '[ %s ]' % ' '.join(str(p) for p in kernel.param_vector())}
code = code + gpml.DISTANCE_CODE_FOOTER_HIGH_MEM % {'writefile': '%(output_file)s'} # N.B. cblparallel manages output files
code = re.sub('% ', '%% ', code) # HACK - cblparallel not fond of % signs at the moment
# Run code - either locally or on cluster - returning location of output file
if local_computation:
output_file = cblparallel.run_batch_locally([code], language='matlab', max_cpu=1.1, max_mem=1.1, job_check_sleep=30, verbose=verbose, single_thread=False)[0]
else:
output_file = cblparallel.run_batch_on_fear([code], language='matlab', max_jobs=500, verbose=verbose)[0]
# Read in results from experiment
gpml_result = scipy.io.loadmat(output_file)
distance = gpml_result['sim_matrix']
# Remove temporary files (perhaps on the cluster server)
cblparallel.remove_temp_file(output_file, local_computation)
cblparallel.remove_temp_file(data_file, local_computation)
# Return distance matrix
return distance
def make_predictions(X, y, Xtest, ytest, model, local_computation=False, max_jobs=500, verbose=True, random_seed=0, no_noise=False):
# Make data into matrices in case they're unidimensional.
if X.ndim == 1: X = X[:, nax]
if y.ndim == 1: y = y[:, nax]
ndata = y.shape[0]
# Save temporary data file in standard temporary directory
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X, 'y': y, 'Xtest' : Xtest, 'ytest' : ytest})
# Copy onto cluster server if necessary
if not local_computation:
if verbose:
print 'Moving data file to fear'
cblparallel.copy_to_remote(data_file)
# Create prediction code
parameters ={'datafile': data_file.split('/')[-1],
'writefile': '%(output_file)s',
'gpml_path': cblparallel.gpml_path(local_computation),
'mean_syntax': model.mean.get_gpml_expression(dimensions=X.shape[1]),
'mean_params': '[ %s ]' % ' '.join(str(p) for p in model.mean.param_vector),
'kernel_syntax': model.kernel.get_gpml_expression(dimensions=X.shape[1]),
'kernel_params': '[ %s ]' % ' '.join(str(p) for p in model.kernel.param_vector),
'lik_syntax': model.likelihood.get_gpml_expression(dimensions=X.shape[1]),
'lik_params': '[ %s ]' % ' '.join(str(p) for p in model.likelihood.param_vector),
'inference': model.likelihood.gpml_inference_method,
'iters': str(30),
'seed': str(random_seed)}
code = gpml.PREDICT_AND_SAVE_CODE % parameters
code = re.sub('% ', '%% ', code) # HACK - cblparallel currently does not like % signs
# Evaluate code - potentially on cluster
if local_computation:
temp_results_file = cblparallel.run_batch_locally([code], language='matlab', max_cpu=1.1, max_mem=1.1, verbose=verbose)[0]
else:
temp_results_file = cblparallel.run_batch_on_fear([code], language='matlab', max_jobs=max_jobs, verbose=verbose)[0]
results = scipy.io.loadmat(temp_results_file)
# Remove temporary files (perhaps on the cluster server)
cblparallel.remove_temp_file(temp_results_file, local_computation)
cblparallel.remove_temp_file(data_file, local_computation)
# Return dictionary of MATLAB results
return results
def evaluate_models(models,
X,
y,
verbose=True,
iters=300,
local_computation=False,
zip_files=False,
max_jobs=500,
random_seed=0,
subset=False,
subset_size=250,
full_iters=0,
bundle_size=1):
# Make data into matrices in case they're unidimensional.
if X.ndim == 1: X = X[:, nax]
if y.ndim == 1: y = y[:, nax]
ndata = y.shape[0]
# Create data file
if verbose:
print 'Creating data file locally'
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X, 'y': y})
# Move to fear if necessary
if not local_computation:
if verbose:
print 'Moving data file to fear'
cblparallel.copy_to_remote(data_file)
# Create a list of MATLAB scripts to assess and optimise parameters for each kernel
if verbose:
print 'Creating scripts'
scripts = [None] * len(models)
for (i, model) in enumerate(models):
parameters = {
'datafile':
data_file.split('/')[-1],
'writefile':
'%(output_file)s', # N.B. cblparallel manages output files
'gpml_path':
cblparallel.gpml_path(local_computation),
'mean_syntax':
model.mean.get_gpml_expression(dimensions=X.shape[1]),
'mean_params':
'[ %s ]' % ' '.join(str(p) for p in model.mean.param_vector),
'kernel_syntax':
model.kernel.get_gpml_expression(dimensions=X.shape[1]),
'kernel_params':
'[ %s ]' % ' '.join(str(p) for p in model.kernel.param_vector),
'lik_syntax':
model.likelihood.get_gpml_expression(dimensions=X.shape[1]),
'lik_params':
'[ %s ]' % ' '.join(str(p) for p in model.likelihood.param_vector),
'inference':
model.likelihood.gpml_inference_method,
'iters':
str(iters),
'seed':
str(np.random.randint(2**31)),
'subset':
'true' if subset else 'false',
'subset_size':
str(subset_size),
'full_iters':
str(full_iters)
}
scripts[i] = gpml.OPTIMIZE_KERNEL_CODE % parameters
#### Need to be careful with % signs
#### For the moment, cblparallel expects no single % signs - FIXME
scripts[i] = re.sub('% ', '%% ', scripts[i])
# Send to cblparallel and save output_files
if verbose:
print 'Sending scripts to cblparallel'
if local_computation:
output_files = cblparallel.run_batch_locally(scripts,
language='matlab',
max_cpu=1.1,
job_check_sleep=5,
submit_sleep=0.1,
max_running_jobs=10,
verbose=verbose)
else:
output_files = cblparallel.run_batch_on_fear(scripts,
language='matlab',
max_jobs=max_jobs,
verbose=verbose,
zip_files=zip_files,
bundle_size=bundle_size)
# Read in results
results = [None] * len(models)
for (i, output_file) in enumerate(output_files):
if verbose:
print 'Reading output file %d of %d' % (i + 1, len(models))
results[i] = GPModel.from_matlab_output(gpml.read_outputs(output_file),
models[i], ndata)
# Tidy up local output files
for (i, output_file) in enumerate(output_files):
if verbose:
print 'Removing output file %d of %d' % (i + 1, len(models))
os.remove(output_file)
# Remove temporary data file (perhaps on the cluster server)
cblparallel.remove_temp_file(data_file, local_computation)
# Return results i.e. list of ScoredKernel objects
return results
def evaluate_models(models, X, y, verbose=True, iters=300, local_computation=False, zip_files=False, max_jobs=500, random_seed=0, subset=False, subset_size=250, full_iters=0, bundle_size=1):
# Make data into matrices in case they're unidimensional.
if X.ndim == 1: X = X[:, nax]
if y.ndim == 1: y = y[:, nax]
ndata = y.shape[0]
# Create data file
if verbose:
print 'Creating data file locally'
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X, 'y': y})
# Move to fear if necessary
if not local_computation:
if verbose:
print 'Moving data file to fear'
cblparallel.copy_to_remote(data_file)
# Create a list of MATLAB scripts to assess and optimise parameters for each kernel
if verbose:
print 'Creating scripts'
scripts = [None] * len(models)
for (i, model) in enumerate(models):
parameters = {'datafile': data_file.split('/')[-1],
'writefile': '%(output_file)s', # N.B. cblparallel manages output files
'gpml_path': cblparallel.gpml_path(local_computation),
'mean_syntax': model.mean.get_gpml_expression(dimensions=X.shape[1]),
'mean_params': '[ %s ]' % ' '.join(str(p) for p in model.mean.param_vector),
'kernel_syntax': model.kernel.get_gpml_expression(dimensions=X.shape[1]),
'kernel_params': '[ %s ]' % ' '.join(str(p) for p in model.kernel.param_vector),
'lik_syntax': model.likelihood.get_gpml_expression(dimensions=X.shape[1]),
'lik_params': '[ %s ]' % ' '.join(str(p) for p in model.likelihood.param_vector),
'inference': model.likelihood.gpml_inference_method,
'iters': str(iters),
'seed': str(np.random.randint(2**31)),
'subset': 'true' if subset else 'false',
'subset_size' : str(subset_size),
'full_iters' : str(full_iters)}
scripts[i] = gpml.OPTIMIZE_KERNEL_CODE % parameters
#### Need to be careful with % signs
#### For the moment, cblparallel expects no single % signs - FIXME
scripts[i] = re.sub('% ', '%% ', scripts[i])
# Send to cblparallel and save output_files
if verbose:
print 'Sending scripts to cblparallel'
if local_computation:
output_files = cblparallel.run_batch_locally(scripts, language='matlab', max_cpu=1.1, job_check_sleep=5, submit_sleep=0.1, max_running_jobs=10, verbose=verbose)
else:
output_files = cblparallel.run_batch_on_fear(scripts, language='matlab', max_jobs=max_jobs, verbose=verbose, zip_files=zip_files, bundle_size=bundle_size)
# Read in results
results = [None] * len(models)
for (i, output_file) in enumerate(output_files):
if verbose:
print 'Reading output file %d of %d' % (i + 1, len(models))
results[i] = GPModel.from_matlab_output(gpml.read_outputs(output_file), models[i], ndata)
# Tidy up local output files
for (i, output_file) in enumerate(output_files):
if verbose:
print 'Removing output file %d of %d' % (i + 1, len(models))
os.remove(output_file)
# Remove temporary data file (perhaps on the cluster server)
cblparallel.remove_temp_file(data_file, local_computation)
# Return results i.e. list of ScoredKernel objects
return results
used_script_names = []
for file_name in script_names:
if file_name[-2:] == '.m':
# print 'Reading %s' % file_name
used_script_names.append(file_name)
with open(os.path.join('scripts', file_name)) as script_file:
scripts.append(script_file.read() % {'data_file' : os.path.split(temp_data_file_name)[-1], 'output_file' : '%(output_file)s'})
script_names = used_script_names
# Send to cluster
# print 'Found %d scripts' % len(scripts)
output_files = cblparallel.run_batch_on_fear(scripts, language='matlab', max_jobs=1000, verbose=False, zip_files=False, bundle_size=1)
# print '%d output files returned' % len(output_files)
# Move output
for (src, name) in zip(output_files, script_names):
# print 'Moving %s output' % name.split('.')[0]
dest = os.path.join('outputs', name.split('.')[0] + '.mat')
shutil.move(src, dest)
# print 'Success'
# Delete local data
#for file_name in os.listdir('data'):
# if file_name[-4:] == '.mat':
def evaluate_models(
models,
X,
y,
verbose=True,
iters=300,
local_computation=False,
zip_files=False,
max_jobs=500,
random_seed=0,
subset=False,
subset_size=250,
full_iters=0,
bundle_size=1,
):
# Make data into matrices in case they're unidimensional.
if X.ndim == 1:
X = X[:, nax]
if y.ndim == 1:
y = y[:, nax]
ndata = y.shape[0]
# Create data file
if verbose:
print "Creating data file locally"
data_file = cblparallel.create_temp_file(".mat")
scipy.io.savemat(data_file, {"X": X, "y": y})
# Move to fear if necessary
if not local_computation:
if verbose:
print "Moving data file to fear"
cblparallel.copy_to_remote(data_file)
# Create a list of MATLAB scripts to assess and optimise parameters for each kernel
if verbose:
print "Creating scripts"
scripts = [None] * len(models)
for (i, model) in enumerate(models):
parameters = {
"datafile": data_file.split("/")[-1],
"writefile": "%(output_file)s", # N.B. cblparallel manages output files
"gpml_path": cblparallel.gpml_path(local_computation),
"mean_syntax": model.mean.get_gpml_expression(dimensions=X.shape[1]),
"mean_params": "[ %s ]" % " ".join(str(p) for p in model.mean.param_vector),
"kernel_syntax": model.kernel.get_gpml_expression(dimensions=X.shape[1]),
"kernel_params": "[ %s ]" % " ".join(str(p) for p in model.kernel.param_vector),
"lik_syntax": model.likelihood.get_gpml_expression(dimensions=X.shape[1]),
"lik_params": "[ %s ]" % " ".join(str(p) for p in model.likelihood.param_vector),
"inference": model.likelihood.gpml_inference_method,
"iters": str(iters),
"seed": str(np.random.randint(2 ** 31)),
"subset": "true" if subset else "false",
"subset_size": str(subset_size),
"full_iters": str(full_iters),
}
scripts[i] = gpml.OPTIMIZE_KERNEL_CODE % parameters
#### Need to be careful with % signs
#### For the moment, cblparallel expects no single % signs - FIXME
scripts[i] = re.sub("% ", "%% ", scripts[i])
# Send to cblparallel and save output_files
if verbose:
print "Sending scripts to cblparallel"
if local_computation:
output_files = cblparallel.run_batch_locally(
scripts,
language="matlab",
max_cpu=1.1,
job_check_sleep=5,
submit_sleep=0.1,
max_running_jobs=10,
verbose=verbose,
)
else:
output_files = cblparallel.run_batch_on_fear(
scripts, language="matlab", max_jobs=max_jobs, verbose=verbose, zip_files=zip_files, bundle_size=bundle_size
)
# Read in results
results = [None] * len(models)
for (i, output_file) in enumerate(output_files):
if verbose:
print "Reading output file %d of %d" % (i + 1, len(models))
results[i] = GPModel.from_matlab_output(gpml.read_outputs(output_file), models[i], ndata)
# Tidy up local output files
for (i, output_file) in enumerate(output_files):
if verbose:
print "Removing output file %d of %d" % (i + 1, len(models))
os.remove(output_file)
# Remove temporary data file (perhaps on the cluster server)
cblparallel.remove_temp_file(data_file, local_computation)
# Return results i.e. list of ScoredKernel objects
return results
def covariance_distance(kernels, X, local_computation=True, verbose=True):
'''
Evaluate a distance matrix of kernels, in terms of their covariance matrix evaluated on training inputs
Input:
- kernels - A list of fk.ScoredKernel
- X - A matrix (data_points x dimensions) of input locations
- local_computation - Boolean indicating if computation should be performed on cluster or on local machine
Return:
- A matrix of similarities between the input kernels
'''
assert (len(kernels) > 0
) #### FIXME - This sort of check should happen earlier
# Make data into matrices in case they're unidimensional.
if X.ndim == 1: X = X[:, nax]
# Save temporary data file in standard temporary directory
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X})
# Copy onto cluster server if necessary
if not local_computation:
if verbose:
print 'Moving data file to fear'
cblparallel.copy_to_remote(data_file)
# Construct testing code
code = gpml.DISTANCE_CODE_HEADER % {
'datafile': data_file.split('/')[-1],
'gpml_path': cblparallel.gpml_path(local_computation)
}
for (i, kernel) in enumerate([k.k_opt for k in kernels]):
code = code + gpml.DISTANCE_CODE_COV % {
'iter':
i + 1,
'kernel_family':
kernel.gpml_kernel_expression(),
'kernel_params':
'[ %s ]' % ' '.join(str(p) for p in kernel.param_vector())
}
code = code + gpml.DISTANCE_CODE_FOOTER_HIGH_MEM % {
'writefile': '%(output_file)s'
} # N.B. cblparallel manages output files
code = re.sub('% ', '%% ',
code) # HACK - cblparallel not fond of % signs at the moment
# Run code - either locally or on cluster - returning location of output file
if local_computation:
output_file = cblparallel.run_batch_locally([code],
language='matlab',
max_cpu=1.1,
max_mem=1.1,
job_check_sleep=30,
verbose=verbose,
single_thread=False)[0]
else:
output_file = cblparallel.run_batch_on_fear([code],
language='matlab',
max_jobs=500,
verbose=verbose)[0]
# Read in results from experiment
gpml_result = scipy.io.loadmat(output_file)
distance = gpml_result['sim_matrix']
# Remove temporary files (perhaps on the cluster server)
cblparallel.remove_temp_file(output_file, local_computation)
cblparallel.remove_temp_file(data_file, local_computation)
# Return distance matrix
return distance
def make_predictions(X,
y,
Xtest,
ytest,
best_scored_kernel,
local_computation=False,
max_jobs=500,
verbose=True,
zero_mean=False,
random_seed=0):
'''
Evaluates a kernel on held out data
Input:
- X - A matrix (data_points x dimensions) of input locations
- y - A matrix (data_points x 1) of output values
- Xtest - Held out X data
- ytest - Held out y data
- best_scored_kernel - A Scored Kernel object to be evaluated on the held out data
- ...
Return:
- A dictionary of results from the MATLAB script containing:
- loglik - an array of log likelihoods of test data
- predictions - an array of mean predictions for the held out data
- actuals - ytest
- model - I'm not sure FIXME
- timestamp - A time stamp of some sort
'''
# Make data into matrices in case they're unidimensional.
if X.ndim == 1: X = X[:, nax]
if y.ndim == 1: y = y[:, nax]
ndata = y.shape[0]
# Save temporary data file in standard temporary directory
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {
'X': X,
'y': y,
'Xtest': Xtest,
'ytest': ytest
})
# Copy onto cluster server if necessary
if not local_computation:
if verbose:
print 'Moving data file to fear'
cblparallel.copy_to_remote(data_file)
# Create prediction code
parameters = {
'datafile':
data_file.split('/')[-1],
'writefile':
'%(output_file)s',
'gpml_path':
cblparallel.gpml_path(local_computation),
'kernel_family':
best_scored_kernel.k_opt.gpml_kernel_expression(),
'kernel_params':
'[ %s ]' %
' '.join(str(p) for p in best_scored_kernel.k_opt.param_vector()),
'noise':
str(best_scored_kernel.noise),
'iters':
str(30),
'seed':
str(random_seed)
}
if zero_mean:
code = gpml.PREDICT_AND_SAVE_CODE_ZERO_MEAN % parameters
else:
code = gpml.PREDICT_AND_SAVE_CODE % parameters
code = re.sub('% ', '%% ',
code) # HACK - cblparallel currently does not like % signs
# Evaluate code - potentially on cluster
if local_computation:
temp_results_file = cblparallel.run_batch_locally([code],
language='matlab',
max_cpu=1.1,
max_mem=1.1,
verbose=verbose)[0]
else:
temp_results_file = cblparallel.run_batch_on_fear([code],
language='matlab',
max_jobs=max_jobs,
verbose=verbose)[0]
results = scipy.io.loadmat(temp_results_file)
# Remove temporary files (perhaps on the cluster server)
cblparallel.remove_temp_file(temp_results_file, local_computation)
cblparallel.remove_temp_file(data_file, local_computation)
# Return dictionary of MATLAB results
return results
def evaluate_kernels(kernels,
X,
y,
verbose=True,
noise=None,
iters=300,
local_computation=False,
zip_files=False,
max_jobs=500,
zero_mean=False,
random_seed=0):
'''
Sets up the kernel optimisation and nll calculation experiments, returns the results as scored kernels
Input:
- kernels - A list of kernels (i.e. not scored kernels)
- X - A matrix (data_points x dimensions) of input locations
- y - A matrix (data_points x 1) of output values
- ...
Return:
- A list of ScoredKernel objects
'''
# Make data into matrices in case they're unidimensional.
if X.ndim == 1: X = X[:, nax]
if y.ndim == 1: y = y[:, nax]
ndata = y.shape[0]
# Set default noise using a heuristic.
if noise is None:
noise = np.log(np.var(y) / 10)
# Create data file
if verbose:
print 'Creating data file locally'
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X, 'y': y}) # Save regression data
# Move to fear if necessary
if not local_computation:
if verbose:
print 'Moving data file to fear'
cblparallel.copy_to_remote(data_file)
# Create a list of MATLAB scripts to assess and optimise parameters for each kernel
if verbose:
print 'Creating scripts'
scripts = [None] * len(kernels)
for (i, kernel) in enumerate(kernels):
x = kernel.param_vector()
parameters = {
'datafile':
data_file.split('/')[-1],
'writefile':
'%(output_file)s', # N.B. cblparallel manages output files
'gpml_path':
cblparallel.gpml_path(local_computation),
'kernel_family':
kernel.gpml_kernel_expression(),
'kernel_params':
'[ %s ]' % ' '.join(str(p) for p in x if type(p) != list),
'eff_dimensions':
'[ %s ]' % ';'.join(str(p) for p in x if type(p) == list),
'dim_positions':
'[ %s ]' %
' '.join(str(i) for i in range(len(x)) if type(x[i]) == list),
'noise':
str(noise),
'iters':
str(iters),
'seed':
str(random_seed)
}
print parameters['kernel_params'], parameters[
'eff_dimensions'], parameters['dim_positions']
if zero_mean:
scripts[i] = gpml.OPTIMIZE_KERNEL_CODE_ZERO_MEAN % parameters
else:
scripts[i] = gpml.OPTIMIZE_KERNEL_CODE % parameters
#### Need to be careful with % signs
#### For the moment, cblparallel expects no single % signs - FIXME
scripts[i] = re.sub('% ', '%% ', scripts[i])
# Send to cblparallel and save output_files
if verbose:
print 'Sending scripts to cblparallel'
if local_computation:
output_files = cblparallel.run_batch_locally(scripts,
language='matlab',
max_cpu=1.1,
job_check_sleep=5,
submit_sleep=0.1,
max_running_jobs=10,
verbose=verbose)
else:
output_files = cblparallel.run_batch_on_fear(scripts,
language='matlab',
max_jobs=max_jobs,
verbose=verbose,
zip_files=zip_files)
# Read in results
results = [None] * len(kernels)
for (i, output_file) in enumerate(output_files):
if verbose:
print 'Reading output file %d of %d: %s' % (i + 1, len(kernels),
output_file)
results[i] = ScoredKernel.from_matlab_output(
gpml.read_outputs(output_file), kernels[i].family(), ndata)
# Tidy up local output files
for (i, output_file) in enumerate(output_files):
if verbose:
print 'Removing output file %d of %d' % (i + 1, len(kernels))
os.remove(output_file)
# Remove temporary data file (perhaps on the cluster server)
cblparallel.remove_temp_file(data_file, local_computation)
# Return results i.e. list of ScoredKernel objects
return results
def evaluate_kernels(kernels, X, y, verbose=True, noise=None, iters=300, local_computation=False, zip_files=False, max_jobs=500, zero_mean=False, random_seed=0):
'''
Sets up the kernel optimisation and nll calculation experiments, returns the results as scored kernels
Input:
- kernels - A list of kernels (i.e. not scored kernels)
- X - A matrix (data_points x dimensions) of input locations
- y - A matrix (data_points x 1) of output values
- ...
Return:
- A list of ScoredKernel objects
'''
# Make data into matrices in case they're unidimensional.
if X.ndim == 1: X = X[:, nax]
if y.ndim == 1: y = y[:, nax]
ndata = y.shape[0]
# Set default noise using a heuristic.
if noise is None:
noise = np.log(np.var(y)/10)
# Create data file
if verbose:
print 'Creating data file locally'
data_file = cblparallel.create_temp_file('.mat')
scipy.io.savemat(data_file, {'X': X, 'y': y}) # Save regression data
# Move to fear if necessary
if not local_computation:
if verbose:
print 'Moving data file to fear'
cblparallel.copy_to_remote(data_file)
# Create a list of MATLAB scripts to assess and optimise parameters for each kernel
if verbose:
print 'Creating scripts'
scripts = [None] * len(kernels)
for (i, kernel) in enumerate(kernels):
parameters = {'datafile': data_file.split('/')[-1],
'writefile': '%(output_file)s', # N.B. cblparallel manages output files
'gpml_path': cblparallel.gpml_path(local_computation),
'kernel_family': kernel.gpml_kernel_expression(),
'kernel_params': '[ %s ]' % ' '.join(str(p) for p in kernel.param_vector()),
'noise': str(noise),
'iters': str(iters),
'seed': str(random_seed)}
if zero_mean:
scripts[i] = gpml.OPTIMIZE_KERNEL_CODE_ZERO_MEAN % parameters
else:
scripts[i] = gpml.OPTIMIZE_KERNEL_CODE % parameters
#### Need to be careful with % signs
#### For the moment, cblparallel expects no single % signs - FIXME
scripts[i] = re.sub('% ', '%% ', scripts[i])
# Send to cblparallel and save output_files
if verbose:
print 'Sending scripts to cblparallel'
if local_computation:
output_files = cblparallel.run_batch_locally(scripts, language='matlab', max_cpu=1.1, job_check_sleep=5, submit_sleep=0.1, max_running_jobs=10, verbose=verbose)
else:
output_files = cblparallel.run_batch_on_fear(scripts, language='matlab', max_jobs=max_jobs, verbose=verbose, zip_files=zip_files)
# Read in results
results = [None] * len(kernels)
for (i, output_file) in enumerate(output_files):
if verbose:
print 'Reading output file %d of %d' % (i + 1, len(kernels))
results[i] = ScoredKernel.from_matlab_output(gpml.read_outputs(output_file), kernels[i].family(), ndata)
# Tidy up local output files
for (i, output_file) in enumerate(output_files):
if verbose:
print 'Removing output file %d of %d' % (i + 1, len(kernels))
os.remove(output_file)
# Remove temporary data file (perhaps on the cluster server)
cblparallel.remove_temp_file(data_file, local_computation)
# Return results i.e. list of ScoredKernel objects
return results