def main():
parser = argparse.ArgumentParser()
parser.add_argument('-d', '--distribution', choices=['normal', 't'], default='t')
parser.add_argument('-x', '--dimension-x', type=int, default=DIMENSION_X_DEFAULT)
parser.add_argument('-y', '--dimension-y', type=int, default=DIMENSION_Y_DEFAULT)
parser.add_argument('-N', '--num-samples', type=int, default=NUM_SAMPLES_DEFAULT, nargs='+')
parser.add_argument('-M', '--average-iterations', type=int, default=AVERAGE_ITERATIONS_DEFAULT)
parser.add_argument('-s', '--seed', type=int, default=1)
args = parser.parse_args()
dx = args.dimension_x
dy = args.dimension_y
Ns = args.num_samples
M = args.average_iterations
T = 200
dist = args.distribution
dataset = SyntheticDataset(dx=dx, dy=dy, Ns=Ns, M=M, dist=dist, seed=args.seed)
metric = plots.ConditionalRegressionNMSEErrorMetric(T=T, dataset=dataset, output_path='results-camsap-synthetic.pickle')
estimator_objects = [
regressors.common.HuberRegressor(name='ROMER-Huber'),
regressors.conditional.ConditionalRegressor(estimators.conditional.gauss_loss.NewtonConditionalEstimator(newton_tol=1e-6), name='GCRF'),
#regressors.conditional.ConditionalRegressor(estimators.conditional.general_loss.MMNewtonConditionalEstimator(
# loss=losses.tyler(dy), tolerance=1e-6, max_iters=25, newton_tol=1e-6), name='ROMER-Tyler'),
regressors.conditional.ConditionalRegressor(estimators.conditional.general_loss.MMNewtonConditionalEstimator(
loss=losses.generalized_gaussian(0.5, 1), tolerance=1e-6, max_iters=25, newton_tol=1e-6), name='ROMER-GG beta=0.5'),
regressors.conditional.ConditionalRegressor(estimators.conditional.general_loss.MMNewtonConditionalEstimator(
loss=losses.multivariate_t(dy, T_DIST_NU), tolerance=1e-6, max_iters=25, newton_tol=1e-6), name='ROMER-T-distribution')
]
plots.plot_variables_vs_N(estimator_objects, Ns, M, metric)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-p',
'--dimension',
type=int,
default=DIMENSION_DEFAULT)
parser.add_argument('-N',
'--num-samples',
type=int,
default=[60, 70, 80, 100, 150, 250, 500, 850],
nargs='+')
parser.add_argument('-M',
'--average-iterations',
type=int,
default=AVERAGE_ITERATIONS_DEFAULT)
parser.add_argument('-b',
'--beta',
type=float,
default=[1.0, 0.5, 0.2],
nargs='+')
args = parser.parse_args()
p = args.dimension
Ns = args.num_samples
M = args.average_iterations
T = 200
TYLER_MAX_ITERS = 1000
TYLER_NEWTON_STEPS = 750
GAUSSIAN_NEWTON_STEPS = 750
for beta in args.beta:
dataset = SyntheticDataset(p=p, Ns=Ns, M=M, beta=beta)
metric = JointEstimationDistanceErrorMetric(T=T, dataset=dataset)
estimator_objects = [
#estimators.joint.general_loss.MMNewtonJointEstimator(
# loss=losses.tyler(dataset.get_dimension()), tolerance=1e-6, max_iters=TYLER_MAX_ITERS, newton_num_steps=TYLER_NEWTON_STEPS,
# newton_tol=1e-6, name='Tyler'),
# estimators.joint.gauss_loss.NewtonJointEstimator(
# newton_num_steps=GAUSSIAN_NEWTON_STEPS, newton_tol=1e-6, name='GMRF Newton'),
# estimators.joint.general_loss.MMJointEstimator(
# estimators.joint.gauss_loss.InvestJointEstimator(), loss=losses.tyler(dataset.get_dimension()),
# tolerance=1e-6, max_iters=TYLER_MAX_ITERS, name='Tyler'),
#estimators.joint.gauss_loss.InvestJointEstimator(name='GMRF'),
estimators.joint.general_loss.MMNewtonJointEstimator(
loss=losses.generalized_gaussian(beta, 1),
tolerance=1e-6,
max_iters=TYLER_MAX_ITERS,
newton_num_steps=TYLER_NEWTON_STEPS,
newton_tol=1e-6,
name='GG'),
estimators.joint.gauss_loss.SampleCovarianceJointEstimator(
name='Sample covariance')
]
plots.plot_variables_vs_N(estimator_objects, Ns, M, metric, show=False)
plots.show()
def main():
#M = 5
T = 50
parser = argparse.ArgumentParser()
parser.add_argument('-N',
'--num-samples',
type=int,
nargs='+',
default=NUM_SAMPLES_DEFAULT)
parser.add_argument('-M',
'--average-iterations',
type=int,
default=AVERAGE_ITERATIONS_DEFAULT)
args = parser.parse_args()
M = args.average_iterations
Ns = args.num_samples
#Ns = list(range(30, 401, 10))
#Ns = list(range(30, 401, 20))
dataset = FloodsDataset(Ns=Ns,
M=M,
test_size=200,
mode=FloodsDataset.Mode.DP_TO_D)
dataset_full = dataset.structured_full()
dataset_time = dataset.structured_time()
dataset_timespace = dataset.structured_timespace()
loss = losses.tyler(dataset.get_dimension())
TYLER_MAX_ITERS = 12
TYLER_NEWTON_STEPS = 25
GAUSSIAN_NEWTON_STEPS = 200
dx = dataset.get_dimension_x()
dy = dataset.get_dimension_y()
plots.plot_variables_vs_N(
[
# regressors.joint.JointRegressor(dx, dy, estimators.joint.general_loss.MMNewtonJointEstimator(
# loss=loss, tolerance=1e-6, max_iters=TYLER_MAX_ITERS, newton_num_steps=TYLER_NEWTON_STEPS,
# newton_tol=1e-6
# ), name='Tyler time', dataset=dataset_time, initial_N=30),
# regressors.joint.JointRegressor(dx, dy, estimators.joint.gauss_loss.NewtonJointEstimator(
# newton_num_steps=GAUSSIAN_NEWTON_STEPS, newton_tol=1e-6
# ), name='GMRF time', dataset=dataset_time, initial_N=30),
#
# regressors.joint.JointRegressor(dx, dy, estimators.joint.general_loss.MMNewtonJointEstimator(
# loss=loss, tolerance=1e-6, max_iters=TYLER_MAX_ITERS, newton_num_steps=TYLER_NEWTON_STEPS,
# newton_tol=1e-6
# ), name='Tyler time-space', dataset=dataset_timespace, initial_N=60),
# regressors.joint.JointRegressor(dx, dy, estimators.joint.gauss_loss.NewtonJointEstimator(
# newton_num_steps=GAUSSIAN_NEWTON_STEPS, newton_tol=1e-6
# ), name='GMRF time-space', dataset=dataset_timespace, initial_N=50),
regressors.joint.JointRegressor(
dx,
dy,
estimators.joint.log_loss.TylerJointEstimator(),
name='Tyler',
dataset=dataset_full,
initial_N=180),
regressors.common.LinearRegressor(
dx=dx, dy=dy, name='GMRF', dataset=dataset_full,
initial_N=240),
regressors.joint.JointRegressor(
dx,
dy,
estimators.joint.general_loss.MMNewtonJointEstimator(
losses.generalized_gaussian(0.5, 1)),
name='GG 0.5',
dataset=dataset_full,
initial_N=180),
regressors.joint.JointRegressor(
dx,
dy,
estimators.joint.general_loss.MMNewtonJointEstimator(
losses.generalized_gaussian(0.2, 1)),
name='GG 0.2',
dataset=dataset_full,
initial_N=180),
regressors.joint.JointRegressor(
dx,
dy,
estimators.joint.general_loss.MMNewtonJointEstimator(
losses.multivariate_t(dx + dy, 4)),
name='T 2.5',
dataset=dataset_full,
initial_N=180),
],
dataset.get_Ns(),
dataset.get_averaging(),
plots.metrics.JointRegressionNMSEErrorMetric(T),
independent_variable='Training set size')
def main():
#M = 5
T = 20 #50
parser = argparse.ArgumentParser()
parser.add_argument('-d', '--divisions', type=int, default=5) #default=75)
parser.add_argument('-s', '--shuffling', type=int,
default=2) #default=100)
parser.add_argument('-N',
'--num-samples',
type=int,
default=list(range(200, 801, 50)) + [900],
nargs='+')
args = parser.parse_args()
#Ns = list(range(200, 801, 50)) + [900]
#Ns = [200, 300, 400, 600, 700, 900]
Ns = [200, 500, 800]
#Ns = args.num_samples
#dataset = StocksDataset(Ns=Ns, divisions=args.divisions, shuffling=args.shuffling, num_observed=30, num_hidden=7)
dataset = StocksDataset(Ns=Ns,
divisions=args.divisions,
shuffling=args.shuffling,
num_observed=105,
num_hidden=15)
#dataset = StocksDataset(Ns=Ns, divisions=args.divisions, shuffling=args.shuffling, num_observed=60, num_hidden=15)
#dataset = StocksDataset(Ns=Ns, divisions=args.divisions, shuffling=args.shuffling)
#TYLER_MAX_ITERS = 12
#TYLER_NEWTON_STEPS = 25
TYLER_MAX_ITERS = 100
TYLER_NEWTON_STEPS = 750
GAUSSIAN_NEWTON_STEPS = 750
dx = dataset.get_dimension_x()
dy = dataset.get_dimension_y()
plots.plot_variables_vs_N(
[
regressors.joint.JointRegressor(
dx,
dy,
estimators.joint.general_loss.MMNewtonJointEstimator(
loss=losses.tyler(dataset.get_dimension()),
tolerance=1e-6,
max_iters=TYLER_MAX_ITERS,
newton_num_steps=TYLER_NEWTON_STEPS,
newton_tol=1e-6),
name='Tyler'),
regressors.joint.JointRegressor(
dx,
dy,
estimators.joint.gauss_loss.NewtonJointEstimator(
newton_num_steps=GAUSSIAN_NEWTON_STEPS, newton_tol=1e-6),
name='GMRF'),
regressors.joint.JointRegressor(
dx,
dy,
estimators.joint.general_loss.MMNewtonJointEstimator(
loss=losses.generalized_gaussian_d(0.5, 1, dx + dy),
tolerance=1e-6,
max_iters=TYLER_MAX_ITERS,
newton_num_steps=GAUSSIAN_NEWTON_STEPS,
newton_tol=1e-6),
name='Laplace'),
regressors.conditional.ConditionalRegressor(
estimators.conditional.general_loss.
MMNewtonConditionalEstimator(
loss=losses.tyler(dataset.get_dimension()),
tolerance=1e-6,
max_iters=TYLER_MAX_ITERS,
newton_num_steps=TYLER_NEWTON_STEPS,
newton_tol=1e-6),
name='Tyler conditional'),
regressors.conditional.ConditionalRegressor(
estimators.conditional.gauss_loss.NewtonConditionalEstimator(
newton_num_steps=GAUSSIAN_NEWTON_STEPS, newton_tol=1e-6),
name='GCRF'),
regressors.conditional.ConditionalRegressor(
estimators.conditional.general_loss.
MMNewtonConditionalEstimator(
loss=losses.generalized_gaussian_d(0.5, 1, dx + dy),
tolerance=1e-6,
max_iters=TYLER_MAX_ITERS,
newton_num_steps=TYLER_NEWTON_STEPS,
newton_tol=1e-6),
name='Laplace conditional'),
# regressors.joint.JointRegressor(dx, dy, estimators.joint.log_loss.TylerJointEstimator(),
# name='Unstructured Tyler'),
# regressors.joint.JointRegressor(dx, dy, estimators.joint.general_loss.MMNewtonJointEstimator(
# loss=losses.generalized_gaussian(0.2, 1), tolerance=1e-6, max_iters=15,
# newton_num_steps=400, newton_tol=1e-6
# ), name='GG m=1 b=0.2'),
#regressors.joint.JointRegressor(dx, dy, estimators.joint.gauss_loss.InvestJointEstimator(), name='GMRF invest')
],
dataset.get_Ns(),
dataset.get_averaging(),
StocksMSEErrorMetric(T, divisions=args.divisions, dataset=dataset),
independent_variable='Training samples')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-N',
'--num-samples',
type=int,
nargs='+',
default=NUM_SAMPLES_DEFAULT)
parser.add_argument('-M',
'--average-iterations',
type=int,
default=AVERAGE_ITERATIONS_DEFAULT)
args = parser.parse_args()
T = 20
M = args.average_iterations
Ns = args.num_samples
dataset = FloodsDataset(Ns=Ns,
M=10,
test_size=200,
mode=FloodsDataset.Mode.DP_TO_D)
dataset = dataset.structured_full()
plots.plot_variables_vs_N(
[
regressors.common.LinearRegressor(name='Linear Regression'),
# regressors.conditional.ConditionalRegressor(estimators.conditional.general_loss.MMNewtonConditionalEstimator(
# loss=losses.tyler(dataset.get_dimension_y()),
# tolerance=1e-6, max_iters=TYLER_MAX_ITERS, newton_num_steps=TYLER_NEWTON_STEPS, newton_tol=1e-6
# ), name='ROMER-Tyler'),
regressors.conditional.ConditionalRegressor(
estimators.conditional.general_loss.
MMNewtonConditionalEstimator(
loss=losses.generalized_gaussian(0.8, 1),
tolerance=1e-6,
max_iters=TYLER_MAX_ITERS,
newton_num_steps=TYLER_NEWTON_STEPS,
newton_tol=1e-6),
name='ROMER-GG m=1 beta=0.8'),
regressors.conditional.ConditionalRegressor(
estimators.conditional.general_loss.
MMNewtonConditionalEstimator(
loss=losses.generalized_gaussian(0.5, 1),
tolerance=1e-6,
max_iters=TYLER_MAX_ITERS,
newton_num_steps=TYLER_NEWTON_STEPS,
newton_tol=1e-6),
name='ROMER-GG m=1 beta=0.5'),
regressors.conditional.ConditionalRegressor(
estimators.conditional.general_loss.
MMNewtonConditionalEstimator(
loss=losses.multivariate_t(dataset.get_dimension(), 4),
tolerance=1e-6,
max_iters=TYLER_MAX_ITERS,
newton_num_steps=TYLER_NEWTON_STEPS,
newton_tol=1e-6),
name='ROMER-T-distribution')
],
dataset.get_Ns(),
dataset.get_averaging(),
plots.metrics.ConditionalRegressionNMSEErrorMetric(
T,
dataset=dataset,
output_path='results-camsap-floods-dense.pickle'),
independent_variable='Training set size')