def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
description='Build and test models based on dim reductions and provided spectra'
)
subparsers = parser.add_subparsers(dest='subparser_name')
parser.add_argument(
'--metadata_path', type=str, default='.', metavar='PATH',
help='Metadata path to work from, if not ''.'''
)
parser.add_argument(
'--spectra_path', type=str, default='.', metavar='PATH',
help='Spectra path to work from, if not ''.'''
)
parser.add_argument(
'--method', type=str, default='ICA', metavar='METHOD',
help='Dim reduction method to load data for'
)
parser.add_argument(
'--n_jobs', type=int, default=1, metavar='N_JOBS',
help='N_JOBS'
)
parser.add_argument(
'--model', type=str, choices=['ET', 'RF', 'GP', 'KNN', 'SVR'], default='ET',
help='Which model type to use: ET (Extra Trees), RF (Random Forest), GP (Gaussian Process), KNN, or SVR (Support Vector Regression)'
)
parser.add_argument(
'--load_model', action='store_true',
help='Whether or not to load the model from --model_path'
)
parser.add_argument(
'--model_path', type=str, default='model.pkl', metavar='MODEL_PATH',
help='COMPLETE path from which to load a model'
)
parser.add_argument(
'--metadata_flags', type=str, default='', metavar='METADATA_FLAGS',
help='Flags specifying observational metadata pre-processing, e.g. LUNAR_MAG which takes the '\
'magnitude and linearizes it (ignoring that it is an area magnitude)'
)
parser.add_argument(
'--compacted_path', type=str, default=None, metavar='COMPATED_PATH',
help='Path to find compacted/arrayized data; setting this will cause --path, --pattern to be ignored'
)
parser_compare = subparsers.add_parser('compare')
parser_compare.add_argument(
'--folds', type=int, default=3, metavar='TEST_FOLDS',
help='Do k-fold cross validation with specified number of folds. Defaults to 3.'
)
parser_compare.add_argument(
'--iters', type=int, default=50, metavar='HYPER_FIT_ITERS',
help='Number of iterations when fitting hyper-params'
)
parser_compare.add_argument(
'--outputfbk', action='store_true',
help='If set, outputs \'grid_scores_\' data from RandomizedSearchCV'
)
parser_compare.add_argument(
'--save_best', action='store_true',
help='Whether or not to save the (last/best) model built for e.g. --hyper_fit'
)
parser_compare.add_argument(
'--scorer', type=str, choices=['R2', 'MAE', 'MSE', 'LL', 'EXP_VAR', 'MAPED', 'MSEMV'], default='R2',
help='Which scoring method to use to determine ranking of model instances.'
)
parser_compare.add_argument(
'--use_spectra', action='store_true',
help='Whether scoring is done against the DM components or the predicted spectra'
)
parser_compare.add_argument(
'--ivar_cutoff', type=float, default=0.001, metavar='IVAR_CUTOFF',
help='data with inverse variace below cutoff is masked as if ivar==0'
)
parser_compare.add_argument(
'--plot_final_errors', action='store_true',
help='If set, will plot the errors from the final/best model, for the whole dataset, from ' + \
'the best model re-trained on CV folds used for testing.' + \
'Plots all errors on top of each other with low-ish alpha, to give a kind of visual ' + \
'density map of errors.'
)
args = parser.parse_args()
obs_metadata = trim_observation_metadata(load_observation_metadata(args.metadata_path, flags=args.metadata_flags))
sources, components, exposures, wavelengths = ICAize.deserialize_data(args.spectra_path, args.method)
source_model, ss, model_args = ICAize.unpickle_model(args.spectra_path, args.method)
comb_flux_arr, comb_exposure_arr, comb_wavelengths = None, None, None
if args.use_spectra:
comb_flux_arr, comb_exposure_arr, comb_ivar_arr, comb_masks, comb_wavelengths = ICAize.load_data(args)
filter_arr = np.in1d(comb_exposure_arr, exposures)
comb_flux_arr = comb_flux_arr[filter_arr]
comb_exposure_arr = comb_exposure_arr[filter_arr]
sorted_inds = np.argsort(comb_exposure_arr)
comb_flux_arr = comb_flux_arr[sorted_inds]
comb_exposure_arr = comb_exposure_arr[sorted_inds]
del comb_ivar_arr
del comb_masks
reduced_obs_metadata = obs_metadata[np.in1d(obs_metadata['EXP_ID'], exposures)]
reduced_obs_metadata.sort('EXP_ID')
sorted_inds = np.argsort(exposures)
reduced_obs_metadata.remove_column('EXP_ID')
md_len = len(reduced_obs_metadata)
var_count = len(reduced_obs_metadata.columns)
X_arr = np.array(reduced_obs_metadata).view('f8').reshape((md_len,-1))
Y_arr = sources[sorted_inds]
if args.load_model:
predictive_model = load_model(args.model_path)
else:
predictive_model = get_model(args.model)
if args.subparser_name == 'compare':
pdist = get_param_distribution_for_model(args.model, args.iters)
scorer = None
if args.scorer == 'R2':
scorer = make_scorer(R2)
elif args.scorer == 'MAE':
if args.use_spectra:
p_MAE_ = partial(MAE, Y_full=Y_arr, flux_arr=comb_flux_arr,
source_model=source_model, ss=ss,
source_model_args=model_args, method=args.method)
scorer = make_scorer(p_MAE_, greater_is_better=False)
else:
scorer = make_scorer(MAE, greater_is_better=False)
elif args.scorer == 'MSE':
if args.use_spectra:
p_MSE_ = partial(MSE, Y_full=Y_arr, flux_arr=comb_flux_arr,
source_model=source_model, ss=ss,
source_model_args=model_args, method=args.method)
scorer = make_scorer(p_MSE_, greater_is_better=False)
else:
scorer = make_scorer(MSE, greater_is_better=False)
elif args.scorer == 'MSEMV':
if args.use_spectra:
p_MSEMV_ = partial(MSEMV, Y_full=Y_arr, flux_arr=comb_flux_arr,
source_model=source_model, ss=ss,
source_model_args=model_args, method=args.method)
scorer = make_scorer(p_MSEMV_, greater_is_better=False)
else:
scorer = make_scorer(MSEMV, greater_is_better=False)
elif args.scorer == 'EXP_VAR':
if args.use_spectra:
p_EXP_VAR_ = partial(EXP_VAR, Y_full=Y_arr, flux_arr=comb_flux_arr,
source_model=source_model, ss=ss,
source_model_args=model_args, method=args.method)
scorer = make_scorer(p_EXP_VAR_)
else:
scorer = make_scorer(EXP_VAR)
elif args.scorer == 'MAPED':
if args.use_spectra:
p_MAPED_ = partial(MAPED, Y_full=Y_arr, flux_arr=comb_flux_arr,
source_model=source_model, ss=ss,
source_model_args=model_args, method=args.method)
scorer = make_scorer(p_MAPED_, greater_is_better=False)
else:
scorer = make_scorer(MAPED, greater_is_better=False)
elif args.scorer == 'LL':
scorer = None
folder = ShuffleSplit(exposures.shape[0], n_iter=args.folds, test_size=1.0/args.folds,
random_state=12345)
if args.model == 'GP':
predictive_model.random_start = args.folds
rcv = GridSearchCV(predictive_model, param_grid=pdist,
error_score=0, cv=3, n_jobs=args.n_jobs,
scoring=scorer)
#random_state=RANDOM_STATE,
#n_iter=args.iters,
else:
rcv = RandomizedSearchCV(predictive_model, param_distributions=pdist,
n_iter=args.iters, cv=folder, n_jobs=args.n_jobs,
scoring=scorer)
# This is going to fit X (metdata) to Y (DM'ed sources). But there are
# really two tests here: how well hyperparams fit/predict the sources
# and how well they fit/predict the actual source spectra. Until I know
# better, I 'm going to need to build a way to test both.
rcv.fit(X_arr, Y_arr)
print(rcv.best_score_)
print(rcv.best_params_)
print(rcv.best_estimator_)
if args.outputfbk:
print("=+"*10 + "=")
for val in rcv.grid_scores_:
print(val)
print("=+"*10 + "=")
if args.save_best:
save_model(rcv.best_estimator_, args.model_path)
if args.plot_final_errors:
for train_inds, test_inds in folder:
rcv.best_estimator_.fit(X_arr[train_inds], Y_arr[train_inds])
predicted = rcv.best_estimator_.predict(X_arr[test_inds])
back_trans_flux = ICAize.inverse_transform(predicted, source_model, ss, args.method, model_args)
diffs = np.abs(comb_flux_arr[test_inds] - back_trans_flux)
#Is there not 'trick' to getting matplotlib to do this without a loop?
for i in range(diffs.shape[0]):
plt.plot(comb_wavelengths, diffs[i, :], 'b-', alpha=0.01)
plt.show()
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
description='Compute PCA/ICA/NMF/etc. components over set of stacked spectra, save those out, and pickle model'
)
subparsers = parser.add_subparsers(dest='subparser_name')
parser.add_argument(
'--pattern', type=str, default='stacked*exp??????.*', metavar='PATTERN',
help='File pattern for stacked sky fibers.'
)
parser.add_argument(
'--path', type=str, default='.', metavar='PATH',
help='Path to work from, if not ''.'''
)
parser.add_argument(
'--compacted_path', type=str, default=None, metavar='COMPATED_PATH',
help='Path to find compacted/arrayized data; setting this will cause --path, --pattern to be ignored'
)
parser.add_argument(
'--method', type=str, default=['ICA'], metavar='METHOD',
choices=['ICA', 'PCA', 'SPCA', 'NMF', 'ISO', 'KPCA', 'FA', 'DL'], nargs='+',
help='Which dim. reduction method to use'
)
parser.add_argument(
'--scale', action='store_true',
help='Should inputs be scaled? Will mean subtract and value scale, but does not scale variace.'
)
parser.add_argument(
'--ivar_cutoff', type=float, default=0.001, metavar='IVAR_CUTOFF',
help='data with inverse variace below cutoff is masked as if ivar==0'
)
parser.add_argument(
'--n_iter', type=int, default=1200, metavar='MAX_ITER',
help='Maximum number of iterations to allow for convergence. For SDSS data 1000 is a safe number of ICA, while SPCA requires larger values e.g. ~2000 to ~2500'
)
parser.add_argument(
'--n_jobs', type=int, default=None, metavar='N_JOBS',
help='N_JOBS'
)
parser_compare = subparsers.add_parser('compare')
parser_compare.add_argument(
'--max_components', type=int, default=50, metavar='COMP_MAX',
help='Max number of components to use/test'
)
parser_compare.add_argument(
'--min_components', type=int, default=0, metavar='COMP_MIN',
help='Min number of compoenents to use/test'
)
parser_compare.add_argument(
'--step_size', type=int, default=5, metavar='COMP_STEP',
help='Step size from comp_min to comp_max'
)
parser_compare.add_argument(
'--comparison', choices=['EXP_VAR', 'R2', 'MSE', 'MAE'], nargs='*', default=['EXP_VAR'],
help='Comparison methods: Explained variance (score), R2 (score), mean sq. error (loss), MEDIAN absolute error (loss)'
)
parser_compare.add_argument(
'--mle_if_avail', action='store_true',
help='In additon to --comparison, include MLE if PCA or FA methods specified'
)
parser_compare.add_argument(
'--plot_example_reconstruction', action='store_true',
help='Pick a random spectrum, plot its actual and reconstructed versions'
)
parser_build = subparsers.add_parser('build')
parser_build.add_argument(
'--n_components', type=int, default=40, metavar='N_COMPONENTS',
help='Number of ICA/PCA/etc. components'
)
parser_build.add_argument(
'--n_neighbors', type=int, default=10, metavar='N_NEIGHBORS',
help='Number of neighbots for e.g. IsoMap'
)
args = parser.parse_args()
comb_flux_arr, comb_exposure_arr, comb_ivar_arr, comb_masks, comb_wavelengths = iz.load_data(args)
if 'DL' in args.method:
flux_arr = comb_flux_arr.astype(dtype=np.float64)
else:
flux_arr = comb_flux_arr
scaled_flux_arr = None
ss = None
if args.scale:
ss = skpp.StandardScaler(with_std=False)
scaled_flux_arr = ss.fit_transform(flux_arr)
else:
scaled_flux_arr = flux_arr
if args.subparser_name == 'compare':
fig, ax1 = plt.subplots()
ax2 = ax1.twinx()
for method in args.method:
model = iz.get_model(method, max_iter=args.n_iter, random_state=iz.random_state, n_jobs=args.n_jobs)
scores = {}
mles_and_covs = args.mle_if_avail and (method == 'FA' or method == 'PCA')
n_components = np.arange(args.min_components, args.max_components+1, args.step_size)
for n in n_components:
print("Cross validating for n=" + str(n) + " on method " + method)
model.n_components = n
comparisons = iz.score_via_CV(args.comparison,
flux_arr if method == 'NMF' else scaled_flux_arr,
model, method, n_jobs=args.n_jobs, include_mle=mles_and_covs,
modeler=_iter_modeler, scorer=_iter_scorer)
for key, val in comparisons.items():
if key in scores:
scores[key].append(val)
else:
scores[key] = [val]
if mles_and_covs:
#ax2.axhline(cov_mcd_score(scaled_flux_arr, args.scale), color='violet', label='MCD Cov', linestyle='--')
ax2.axhline(cov_lw_score(scaled_flux_arr, args.scale), color='orange', label='LW Cov', linestyle='--')
for key, score_list in scores.items():
if key != 'mle':
ax1.plot(n_components, score_list, label=method + ':' + key + ' scores')
else:
ax2.plot(n_components, score_list, '-.', label=method + ' mle scores')
ax1.set_xlabel('nb of components')
ax1.set_ylabel('CV scores', figure=fig)
ax1.legend(loc='lower left')
ax2.legend(loc='lower right')
plt.show()
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
description='Compute PCA/ICA/NMF/etc. components over set of stacked spectra, save those out, and pickle model'
)
parser.add_argument(
'--pattern', type=str, default='stacked*exp??????.*', metavar='PATTERN',
help='File pattern for stacked sky fibers.'
)
parser.add_argument(
'--path', type=str, default='.', metavar='PATH',
help='Path to work from, if not ''.'''
)
parser.add_argument(
'--compacted_path', type=str, default=None, metavar='COMPATED_PATH',
help='Path to find compacted/arrayized data; setting this will cause --path, --pattern to be ignored'
)
parser.add_argument(
'--n_components', type=int, default=40, metavar='N_COMPONENTS',
help='Number of ICA/PCA/etc. components'
)
parser.add_argument(
'--method', type=str, default='ICA', metavar='METHOD',
choices=['ICA', 'PCA', 'SPCA', 'NMF', 'ISO', 'KPCA', 'FA', 'DL'],
help='Which dim. reduction method to use'
)
parser.add_argument(
'--scale', action='store_true',
help='Should inputs variance be scaled? Defaults to mean subtract and value scale, but w/out this does not scale variance.'
)
parser.add_argument(
'--no_scale', action='store_true',
help='Suppresses all scaling'
)
parser.add_argument(
'--ivar_cutoff', type=float, default=0.001, metavar='IVAR_CUTOFF',
help='data with inverse variace below cutoff is masked as if ivar==0'
)
parser.add_argument(
'--n_iter', type=int, default=1200, metavar='MAX_ITER',
help='Maximum number of iterations to allow for convergence. For SDSS data 1000 is a safe number of ICA, while SPCA requires larger values e.g. ~2000 to ~2500'
)
parser.add_argument(
'--n_jobs', type=int, default=None, metavar='N_JOBS',
help='N_JOBS'
)
args = parser.parse_args()
comb_flux_arr, comb_exposure_arr, comb_ivar_arr, comb_masks, comb_wavelengths = iz.load_data(args)
model = iz.get_model(args.method, n=args.n_components, n_neighbors=None, max_iter=args.n_iter, random_state=iz.random_state, n_jobs=args.n_jobs)
ss = None
if args.no_scale:
scaled_flux_arr = comb_flux_arr
else:
ss = skpp.StandardScaler(with_std=False)
if args.scale:
ss = skpp.StandardScaler(with_std=True)
scaled_flux_arr = ss.fit_transform(comb_flux_arr)
#Heavily copied from J. Vanderplas/astroML bayesian_blocks.py
N = comb_wavelengths.size
step = args.n_components * 4
edges = np.concatenate([comb_wavelengths[:1:step],
0.5 * (comb_wavelengths[1::step] + comb_wavelengths[:-1:step]),
comb_wavelengths[-1::step]])
block_length = comb_wavelengths[-1::step] - edges
# arrays to store the best configuration
nn_vec = np.ones(N/step) * step
best = np.zeros(N, dtype=float)
last = np.zeros(N, dtype=int)
for R in range(N/step):
print("R: " + str(R))
width = block_length[:R + 1] - block_length[R + 1]
count_vec = np.cumsum(nn_vec[:R + 1][::-1])[::-1]
#width = nn_vec[:R + 1] - nn_vec[R + 1]
#count_vec = np.cumsum(nn_vec[:R + 1][::-1])[::-1]
#print(width)
#print(count_vec)
#raw_input("Pausing... ")
fit_vec = map(lambda n: iz.score_via_CV(['LL'], scaled_flux_arr[:, :n], model, ss, args.method, folds=3, n_jobs=args.n_jobs), count_vec)
fit_vec = [d["mle"] for d in fit_vec]
#print(fit_vec)
fit_vec[1:] += best[:R]
#print(fit_vec)
i_max = np.argmax(fit_vec)
last[R] = i_max
best[R] = fit_vec[i_max]
#print(best)
change_points = np.zeros(N/step, dtype=int)
i_cp = N/step
ind = N/step
while True:
i_cp -= 1
change_points[i_cp] = ind
if ind == 0:
break
ind = last[ind - 1]
change_points = change_points[i_cp:]
print(edges[change_points])
'''
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
description=
'Compute PCA/ICA/NMF/etc. components over set of stacked spectra, save those out, and pickle model'
)
parser.add_argument('--pattern',
type=str,
default='stacked*exp??????.*',
metavar='PATTERN',
help='File pattern for stacked sky fibers.')
parser.add_argument('--path',
type=str,
default='.',
metavar='PATH',
help='Path to work from, if not '
'.'
'')
parser.add_argument(
'--compacted_path',
type=str,
default=None,
metavar='COMPATED_PATH',
help=
'Path to find compacted/arrayized data; setting this will cause --path, --pattern to be ignored'
)
parser.add_argument('--n_components',
type=int,
default=40,
metavar='N_COMPONENTS',
help='Number of ICA/PCA/etc. components')
parser.add_argument(
'--method',
type=str,
default='ICA',
metavar='METHOD',
choices=['ICA', 'PCA', 'SPCA', 'NMF', 'ISO', 'KPCA', 'FA', 'DL'],
help='Which dim. reduction method to use')
parser.add_argument(
'--scale',
action='store_true',
help=
'Should inputs variance be scaled? Defaults to mean subtract and value scale, but w/out this does not scale variance.'
)
parser.add_argument('--no_scale',
action='store_true',
help='Suppresses all scaling')
parser.add_argument(
'--ivar_cutoff',
type=float,
default=0.001,
metavar='IVAR_CUTOFF',
help='data with inverse variace below cutoff is masked as if ivar==0')
parser.add_argument(
'--n_iter',
type=int,
default=1200,
metavar='MAX_ITER',
help=
'Maximum number of iterations to allow for convergence. For SDSS data 1000 is a safe number of ICA, while SPCA requires larger values e.g. ~2000 to ~2500'
)
parser.add_argument('--n_jobs',
type=int,
default=None,
metavar='N_JOBS',
help='N_JOBS')
args = parser.parse_args()
comb_flux_arr, comb_exposure_arr, comb_ivar_arr, comb_masks, comb_wavelengths = iz.load_data(
args)
model = iz.get_model(args.method,
n=args.n_components,
n_neighbors=None,
max_iter=args.n_iter,
random_state=iz.random_state,
n_jobs=args.n_jobs)
ss = None
if args.no_scale:
scaled_flux_arr = comb_flux_arr
else:
ss = skpp.StandardScaler(with_std=False)
if args.scale:
ss = skpp.StandardScaler(with_std=True)
scaled_flux_arr = ss.fit_transform(comb_flux_arr)
#Heavily copied from J. Vanderplas/astroML bayesian_blocks.py
N = comb_wavelengths.size
step = args.n_components * 4
edges = np.concatenate([
comb_wavelengths[:1:step],
0.5 * (comb_wavelengths[1::step] + comb_wavelengths[:-1:step]),
comb_wavelengths[-1::step]
])
block_length = comb_wavelengths[-1::step] - edges
# arrays to store the best configuration
nn_vec = np.ones(N / step) * step
best = np.zeros(N, dtype=float)
last = np.zeros(N, dtype=int)
for R in range(N / step):
print("R: " + str(R))
width = block_length[:R + 1] - block_length[R + 1]
count_vec = np.cumsum(nn_vec[:R + 1][::-1])[::-1]
#width = nn_vec[:R + 1] - nn_vec[R + 1]
#count_vec = np.cumsum(nn_vec[:R + 1][::-1])[::-1]
#print(width)
#print(count_vec)
#raw_input("Pausing... ")
fit_vec = map(
lambda n: iz.score_via_CV(['LL'],
scaled_flux_arr[:, :n],
model,
ss,
args.method,
folds=3,
n_jobs=args.n_jobs), count_vec)
fit_vec = [d["mle"] for d in fit_vec]
#print(fit_vec)
fit_vec[1:] += best[:R]
#print(fit_vec)
i_max = np.argmax(fit_vec)
last[R] = i_max
best[R] = fit_vec[i_max]
#print(best)
change_points = np.zeros(N / step, dtype=int)
i_cp = N / step
ind = N / step
while True:
i_cp -= 1
change_points[i_cp] = ind
if ind == 0:
break
ind = last[ind - 1]
change_points = change_points[i_cp:]
print(edges[change_points])
'''
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
description=
'Build and test models based on dim reductions and provided spectra')
subparsers = parser.add_subparsers(dest='subparser_name')
parser.add_argument('--metadata_path',
type=str,
default='.',
metavar='PATH',
help='Metadata path to work from, if not '
'.'
'')
parser.add_argument('--spectra_path',
type=str,
default='.',
metavar='PATH',
help='Spectra path to work from, if not '
'.'
'')
parser.add_argument('--method',
type=str,
default='ICA',
metavar='METHOD',
help='Dim reduction method to load data for')
parser.add_argument('--n_jobs',
type=int,
default=1,
metavar='N_JOBS',
help='N_JOBS')
parser.add_argument(
'--model',
type=str,
choices=['ET', 'RF', 'GP', 'KNN', 'SVR'],
default='ET',
help=
'Which model type to use: ET (Extra Trees), RF (Random Forest), GP (Gaussian Process), KNN, or SVR (Support Vector Regression)'
)
parser.add_argument(
'--load_model',
action='store_true',
help='Whether or not to load the model from --model_path')
parser.add_argument('--model_path',
type=str,
default='model.pkl',
metavar='MODEL_PATH',
help='COMPLETE path from which to load a model')
parser.add_argument(
'--metadata_flags', type=str, default='', metavar='METADATA_FLAGS',
help='Flags specifying observational metadata pre-processing, e.g. LUNAR_MAG which takes the '\
'magnitude and linearizes it (ignoring that it is an area magnitude)'
)
parser.add_argument(
'--compacted_path',
type=str,
default=None,
metavar='COMPATED_PATH',
help=
'Path to find compacted/arrayized data; setting this will cause --path, --pattern to be ignored'
)
parser_compare = subparsers.add_parser('compare')
parser_compare.add_argument(
'--folds',
type=int,
default=3,
metavar='TEST_FOLDS',
help=
'Do k-fold cross validation with specified number of folds. Defaults to 3.'
)
parser_compare.add_argument(
'--iters',
type=int,
default=50,
metavar='HYPER_FIT_ITERS',
help='Number of iterations when fitting hyper-params')
parser_compare.add_argument(
'--outputfbk',
action='store_true',
help='If set, outputs \'grid_scores_\' data from RandomizedSearchCV')
parser_compare.add_argument(
'--save_best',
action='store_true',
help=
'Whether or not to save the (last/best) model built for e.g. --hyper_fit'
)
parser_compare.add_argument(
'--scorer',
type=str,
choices=['R2', 'MAE', 'MSE', 'LL', 'EXP_VAR', 'MAPED', 'MSEMV'],
default='R2',
help=
'Which scoring method to use to determine ranking of model instances.')
parser_compare.add_argument(
'--use_spectra',
action='store_true',
help=
'Whether scoring is done against the DM components or the predicted spectra'
)
parser_compare.add_argument(
'--ivar_cutoff',
type=float,
default=0.001,
metavar='IVAR_CUTOFF',
help='data with inverse variace below cutoff is masked as if ivar==0')
parser_compare.add_argument(
'--plot_final_errors', action='store_true',
help='If set, will plot the errors from the final/best model, for the whole dataset, from ' + \
'the best model re-trained on CV folds used for testing.' + \
'Plots all errors on top of each other with low-ish alpha, to give a kind of visual ' + \
'density map of errors.'
)
args = parser.parse_args()
obs_metadata = trim_observation_metadata(
load_observation_metadata(args.metadata_path,
flags=args.metadata_flags))
sources, components, exposures, wavelengths = ICAize.deserialize_data(
args.spectra_path, args.method)
source_model, ss, model_args = ICAize.unpickle_model(
args.spectra_path, args.method)
comb_flux_arr, comb_exposure_arr, comb_wavelengths = None, None, None
if args.use_spectra:
comb_flux_arr, comb_exposure_arr, comb_ivar_arr, comb_masks, comb_wavelengths = ICAize.load_data(
args)
filter_arr = np.in1d(comb_exposure_arr, exposures)
comb_flux_arr = comb_flux_arr[filter_arr]
comb_exposure_arr = comb_exposure_arr[filter_arr]
sorted_inds = np.argsort(comb_exposure_arr)
comb_flux_arr = comb_flux_arr[sorted_inds]
comb_exposure_arr = comb_exposure_arr[sorted_inds]
del comb_ivar_arr
del comb_masks
reduced_obs_metadata = obs_metadata[np.in1d(obs_metadata['EXP_ID'],
exposures)]
reduced_obs_metadata.sort('EXP_ID')
sorted_inds = np.argsort(exposures)
reduced_obs_metadata.remove_column('EXP_ID')
md_len = len(reduced_obs_metadata)
var_count = len(reduced_obs_metadata.columns)
X_arr = np.array(reduced_obs_metadata).view('f8').reshape((md_len, -1))
Y_arr = sources[sorted_inds]
if args.load_model:
predictive_model = load_model(args.model_path)
else:
predictive_model = get_model(args.model)
if args.subparser_name == 'compare':
pdist = get_param_distribution_for_model(args.model, args.iters)
scorer = None
if args.scorer == 'R2':
scorer = make_scorer(R2)
elif args.scorer == 'MAE':
if args.use_spectra:
p_MAE_ = partial(MAE,
Y_full=Y_arr,
flux_arr=comb_flux_arr,
source_model=source_model,
ss=ss,
source_model_args=model_args,
method=args.method)
scorer = make_scorer(p_MAE_, greater_is_better=False)
else:
scorer = make_scorer(MAE, greater_is_better=False)
elif args.scorer == 'MSE':
if args.use_spectra:
p_MSE_ = partial(MSE,
Y_full=Y_arr,
flux_arr=comb_flux_arr,
source_model=source_model,
ss=ss,
source_model_args=model_args,
method=args.method)
scorer = make_scorer(p_MSE_, greater_is_better=False)
else:
scorer = make_scorer(MSE, greater_is_better=False)
elif args.scorer == 'MSEMV':
if args.use_spectra:
p_MSEMV_ = partial(MSEMV,
Y_full=Y_arr,
flux_arr=comb_flux_arr,
source_model=source_model,
ss=ss,
source_model_args=model_args,
method=args.method)
scorer = make_scorer(p_MSEMV_, greater_is_better=False)
else:
scorer = make_scorer(MSEMV, greater_is_better=False)
elif args.scorer == 'EXP_VAR':
if args.use_spectra:
p_EXP_VAR_ = partial(EXP_VAR,
Y_full=Y_arr,
flux_arr=comb_flux_arr,
source_model=source_model,
ss=ss,
source_model_args=model_args,
method=args.method)
scorer = make_scorer(p_EXP_VAR_)
else:
scorer = make_scorer(EXP_VAR)
elif args.scorer == 'MAPED':
if args.use_spectra:
p_MAPED_ = partial(MAPED,
Y_full=Y_arr,
flux_arr=comb_flux_arr,
source_model=source_model,
ss=ss,
source_model_args=model_args,
method=args.method)
scorer = make_scorer(p_MAPED_, greater_is_better=False)
else:
scorer = make_scorer(MAPED, greater_is_better=False)
elif args.scorer == 'LL':
scorer = None
folder = ShuffleSplit(exposures.shape[0],
n_iter=args.folds,
test_size=1.0 / args.folds,
random_state=12345)
if args.model == 'GP':
predictive_model.random_start = args.folds
rcv = GridSearchCV(predictive_model,
param_grid=pdist,
error_score=0,
cv=3,
n_jobs=args.n_jobs,
scoring=scorer)
#random_state=RANDOM_STATE,
#n_iter=args.iters,
else:
rcv = RandomizedSearchCV(predictive_model,
param_distributions=pdist,
n_iter=args.iters,
cv=folder,
n_jobs=args.n_jobs,
scoring=scorer)
# This is going to fit X (metdata) to Y (DM'ed sources). But there are
# really two tests here: how well hyperparams fit/predict the sources
# and how well they fit/predict the actual source spectra. Until I know
# better, I 'm going to need to build a way to test both.
rcv.fit(X_arr, Y_arr)
print(rcv.best_score_)
print(rcv.best_params_)
print(rcv.best_estimator_)
if args.outputfbk:
print("=+" * 10 + "=")
for val in rcv.grid_scores_:
print(val)
print("=+" * 10 + "=")
if args.save_best:
save_model(rcv.best_estimator_, args.model_path)
if args.plot_final_errors:
for train_inds, test_inds in folder:
rcv.best_estimator_.fit(X_arr[train_inds], Y_arr[train_inds])
predicted = rcv.best_estimator_.predict(X_arr[test_inds])
back_trans_flux = ICAize.inverse_transform(
predicted, source_model, ss, args.method, model_args)
diffs = np.abs(comb_flux_arr[test_inds] - back_trans_flux)
#Is there not 'trick' to getting matplotlib to do this without a loop?
for i in range(diffs.shape[0]):
plt.plot(comb_wavelengths, diffs[i, :], 'b-', alpha=0.01)
plt.show()
