def test_get_contingency_table(self):
"""Ensures correct output from get_contingency_table."""
this_contingency_table = model_eval.get_contingency_table(
FORECAST_LABELS_THRESHOLD_HALF, OBSERVED_LABELS)
self.assertTrue(
this_contingency_table == CONTINGENCY_TABLE_THRESHOLD_HALF)
def run_evaluation(class_probability_matrix, observed_labels, output_dir_name):
"""Evaluates a set of multiclass probabilistic predictions.
E = number of examples
K = number of classes
:param class_probability_matrix: E-by-K numpy array, where
class_probability_matrix[i, k] = probability that the [i]th example
belongs to the [k]th class. Classes should be mutually exclusive and
collectively exhaustive, so that the sum across each row is 1.0.
:param observed_labels: length-E numpy array of observed labels. Each label
must be an integer from 0...(K - 1).
:param output_dir_name: Name of output directory. Results will be saved
here.
"""
file_system_utils.mkdir_recursive_if_necessary(
directory_name=output_dir_name)
print 'Finding best binarization threshold (front vs. no front)...'
binarization_threshold, best_gerrity_score = (
eval_utils.find_best_binarization_threshold(
class_probability_matrix=class_probability_matrix,
observed_labels=observed_labels,
threshold_arg=model_eval.THRESHOLD_ARG_FOR_UNIQUE_FORECASTS,
criterion_function=eval_utils.get_gerrity_score,
optimization_direction=eval_utils.MAX_OPTIMIZATION_DIRECTION,
forecast_precision_for_thresholds=FORECAST_PRECISION_FOR_THRESHOLDS
))
print(
'Best binarization threshold = {0:.4f} ... corresponding Gerrity score '
'= {1:.4f}').format(binarization_threshold, best_gerrity_score)
print 'Determinizing multiclass probabilities...'
predicted_labels = eval_utils.determinize_probabilities(
class_probability_matrix=class_probability_matrix,
binarization_threshold=binarization_threshold)
contingency_matrix = eval_utils.get_contingency_table(
predicted_labels=predicted_labels,
observed_labels=observed_labels,
num_classes=class_probability_matrix.shape[1])
print 'Multiclass contingency table is shown below:\n{0:s}'.format(
str(contingency_matrix))
print SEPARATOR_STRING
accuracy = eval_utils.get_accuracy(contingency_matrix)
peirce_score = eval_utils.get_peirce_score(contingency_matrix)
heidke_score = eval_utils.get_heidke_score(contingency_matrix)
gerrity_score = eval_utils.get_gerrity_score(contingency_matrix)
print(
'Multiclass accuracy = {0:.4f} ... Peirce score = {1:.4f} ... '
'Heidke score = {2:.4f} ... Gerrity score = {3:.4f}\n').format(
accuracy, peirce_score, heidke_score, gerrity_score)
binary_contingency_dict = model_eval.get_contingency_table(
forecast_labels=(predicted_labels > 0).astype(int),
observed_labels=(observed_labels > 0).astype(int))
print 'Binary contingency table is shown below:\n{0:s}'.format(
str(binary_contingency_dict))
print SEPARATOR_STRING
binary_pod = model_eval.get_pod(binary_contingency_dict)
binary_pofd = model_eval.get_pofd(binary_contingency_dict)
binary_success_ratio = model_eval.get_success_ratio(
binary_contingency_dict)
binary_focn = model_eval.get_focn(binary_contingency_dict)
binary_accuracy = model_eval.get_accuracy(binary_contingency_dict)
binary_csi = model_eval.get_csi(binary_contingency_dict)
binary_frequency_bias = model_eval.get_frequency_bias(
binary_contingency_dict)
print(
'Binary POD = {0:.4f} ... POFD = {1:.4f} ... success ratio = {2:.4f} '
'... FOCN = {3:.4f} ... accuracy = {4:.4f} ... CSI = {5:.4f} ... '
'frequency bias = {6:.4f}\n').format(binary_pod, binary_pofd,
binary_success_ratio, binary_focn,
binary_accuracy, binary_csi,
binary_frequency_bias)
auc_by_class, sklearn_auc_by_class = _plot_roc_curves(
class_probability_matrix=class_probability_matrix,
observed_labels=observed_labels,
output_dir_name=output_dir_name)
print '\n'
aupd_by_class = _plot_performance_diagrams(
class_probability_matrix=class_probability_matrix,
observed_labels=observed_labels,
output_dir_name=output_dir_name)
print '\n'
reliability_by_class, bss_by_class = _plot_attributes_diagrams(
class_probability_matrix=class_probability_matrix,
observed_labels=observed_labels,
output_dir_name=output_dir_name)
print '\n'
evaluation_file_name = '{0:s}/model_evaluation.p'.format(output_dir_name)
print 'Writing results to: "{0:s}"...\n'.format(evaluation_file_name)
eval_utils.write_evaluation_results(
class_probability_matrix=class_probability_matrix,
observed_labels=observed_labels,
binarization_threshold=binarization_threshold,
accuracy=accuracy,
peirce_score=peirce_score,
heidke_score=heidke_score,
gerrity_score=gerrity_score,
binary_pod=binary_pod,
binary_pofd=binary_pofd,
binary_success_ratio=binary_success_ratio,
binary_focn=binary_focn,
binary_accuracy=binary_accuracy,
binary_csi=binary_csi,
binary_frequency_bias=binary_frequency_bias,
auc_by_class=auc_by_class,
scikit_learn_auc_by_class=sklearn_auc_by_class,
aupd_by_class=aupd_by_class,
reliability_by_class=reliability_by_class,
bss_by_class=bss_by_class,
pickle_file_name=evaluation_file_name)
def _compute_scores(forecast_probabilities,
observed_labels,
num_bootstrap_reps,
output_file_name,
best_prob_threshold=None,
downsampling_dict=None):
"""Computes evaluation scores.
E = number of examples (storm objects)
:param forecast_probabilities: length-E numpy array of forecast event
probabilities.
:param observed_labels: length-E numpy array of observations (1 for event,
0 for non-event).
:param num_bootstrap_reps: Number of bootstrap replicates.
:param output_file_name: Path to output file (will be written by
`model_evaluation.write_evaluation`).
:param best_prob_threshold: Best probability threshold. If None, will be
determined on the fly.
:param downsampling_dict: Dictionary with downsampling fractions. See doc
for `deep_learning_utils.sample_by_class`. If this is None,
downsampling will not be used.
"""
num_examples = len(observed_labels)
num_examples_by_class = numpy.unique(observed_labels,
return_counts=True)[-1]
print('Number of examples by class (no downsampling): {0:s}'.format(
str(num_examples_by_class)))
positive_example_indices = numpy.where(observed_labels == 1)[0]
negative_example_indices = numpy.where(observed_labels == 0)[0]
if downsampling_dict is None:
these_indices = numpy.linspace(0,
num_examples - 1,
num=num_examples,
dtype=int)
else:
these_indices = dl_utils.sample_by_class(
sampling_fraction_by_class_dict=downsampling_dict,
target_name=DUMMY_TARGET_NAME,
target_values=observed_labels,
num_examples_total=num_examples)
this_num_ex_by_class = numpy.unique(observed_labels[these_indices],
return_counts=True)[-1]
print('Number of examples by class (after downsampling): {0:s}'.format(
str(this_num_ex_by_class)))
all_prob_thresholds = model_eval.get_binarization_thresholds(
threshold_arg=model_eval.THRESHOLD_ARG_FOR_UNIQUE_FORECASTS,
forecast_probabilities=forecast_probabilities[these_indices],
forecast_precision=FORECAST_PRECISION)
if best_prob_threshold is None:
best_prob_threshold, best_csi = (
model_eval.find_best_binarization_threshold(
forecast_probabilities=forecast_probabilities[these_indices],
observed_labels=observed_labels[these_indices],
threshold_arg=all_prob_thresholds,
criterion_function=model_eval.get_csi,
optimization_direction=model_eval.MAX_OPTIMIZATION_STRING))
else:
these_forecast_labels = model_eval.binarize_forecast_probs(
forecast_probabilities=forecast_probabilities[these_indices],
binarization_threshold=best_prob_threshold)
this_contingency_dict = model_eval.get_contingency_table(
forecast_labels=these_forecast_labels,
observed_labels=observed_labels[these_indices])
best_csi = model_eval.get_csi(this_contingency_dict)
print(
('Best probability threshold = {0:.4f} ... corresponding CSI = {1:.4f}'
).format(best_prob_threshold, best_csi))
num_examples_by_forecast_bin = model_eval.get_points_in_reliability_curve(
forecast_probabilities=forecast_probabilities[these_indices],
observed_labels=observed_labels[these_indices],
num_forecast_bins=model_eval.DEFAULT_NUM_RELIABILITY_BINS)[-1]
list_of_evaluation_tables = []
for i in range(num_bootstrap_reps):
print(('Computing scores for {0:d}th of {1:d} bootstrap replicates...'
).format(i + 1, num_bootstrap_reps))
if num_bootstrap_reps == 1:
if downsampling_dict is None:
these_indices = numpy.linspace(0,
num_examples - 1,
num=num_examples,
dtype=int)
else:
these_indices = dl_utils.sample_by_class(
sampling_fraction_by_class_dict=downsampling_dict,
target_name=DUMMY_TARGET_NAME,
target_values=observed_labels,
num_examples_total=num_examples)
else:
if len(positive_example_indices) > 0:
these_positive_indices = bootstrapping.draw_sample(
positive_example_indices)[0]
else:
these_positive_indices = numpy.array([], dtype=int)
these_negative_indices = bootstrapping.draw_sample(
negative_example_indices)[0]
these_indices = numpy.concatenate(
(these_positive_indices, these_negative_indices))
if downsampling_dict is not None:
these_subindices = dl_utils.sample_by_class(
sampling_fraction_by_class_dict=downsampling_dict,
target_name=DUMMY_TARGET_NAME,
target_values=observed_labels[these_indices],
num_examples_total=num_examples)
these_indices = these_indices[these_subindices]
if downsampling_dict is not None:
this_num_ex_by_class = numpy.unique(observed_labels[these_indices],
return_counts=True)[-1]
print('Number of examples by class: {0:s}'.format(
str(this_num_ex_by_class)))
this_evaluation_table = model_eval.run_evaluation(
forecast_probabilities=forecast_probabilities[these_indices],
observed_labels=observed_labels[these_indices],
best_prob_threshold=best_prob_threshold,
all_prob_thresholds=all_prob_thresholds,
climatology=numpy.mean(observed_labels[these_indices]))
list_of_evaluation_tables.append(this_evaluation_table)
if i == num_bootstrap_reps - 1:
print(SEPARATOR_STRING)
else:
print(MINOR_SEPARATOR_STRING)
if i == 0:
continue
list_of_evaluation_tables[-1] = list_of_evaluation_tables[-1].align(
list_of_evaluation_tables[0], axis=1)[0]
evaluation_table = pandas.concat(list_of_evaluation_tables,
axis=0,
ignore_index=True)
print('Writing results to: "{0:s}"...'.format(output_file_name))
model_eval.write_evaluation(
pickle_file_name=output_file_name,
forecast_probabilities=forecast_probabilities,
observed_labels=observed_labels,
best_prob_threshold=best_prob_threshold,
all_prob_thresholds=all_prob_thresholds,
num_examples_by_forecast_bin=num_examples_by_forecast_bin,
downsampling_dict=downsampling_dict,
evaluation_table=evaluation_table)
def run_evaluation(forecast_probabilities, observed_labels, output_dir_name):
"""Evaluates forecast-observation pairs from any forecasting method.
Specifically, this method does the following:
- creates ROC (receiver operating characteristic) curve
- creates performance diagram
- creates attributes diagram
- saves each of the aforelisted figures to a .jpg file
- computes many performance metrics and saves them to a Pickle file
:param forecast_probabilities: length-N numpy array of forecast event
probabilities.
:param observed_labels: length-N numpy array of observed labels (1 for
"yes", 0 for "no").
:param output_dir_name: Name of output directory.
"""
file_system_utils.mkdir_recursive_if_necessary(
directory_name=output_dir_name)
# TODO(thunderhoser): Make binarization threshold an input argument to this
# method.
(binarization_threshold, best_csi
) = model_eval.find_best_binarization_threshold(
forecast_probabilities=forecast_probabilities,
observed_labels=observed_labels,
threshold_arg=model_eval.THRESHOLD_ARG_FOR_UNIQUE_FORECASTS,
criterion_function=model_eval.get_csi,
optimization_direction=model_eval.MAX_OPTIMIZATION_DIRECTION,
unique_forecast_precision=FORECAST_PRECISION_FOR_THRESHOLDS)
print (
'Best binarization threshold = {0:.4f} ... corresponding CSI = {1:.4f}'
).format(binarization_threshold, best_csi)
print 'Binarizing forecast probabilities...'
forecast_labels = model_eval.binarize_forecast_probs(
forecast_probabilities=forecast_probabilities,
binarization_threshold=binarization_threshold)
print 'Creating contingency table...'
contingency_table_as_dict = model_eval.get_contingency_table(
forecast_labels=forecast_labels, observed_labels=observed_labels)
print '{0:s}\n'.format(str(contingency_table_as_dict))
print 'Computing performance metrics...'
pod = model_eval.get_pod(contingency_table_as_dict)
pofd = model_eval.get_pofd(contingency_table_as_dict)
success_ratio = model_eval.get_success_ratio(contingency_table_as_dict)
focn = model_eval.get_focn(contingency_table_as_dict)
accuracy = model_eval.get_accuracy(contingency_table_as_dict)
csi = model_eval.get_csi(contingency_table_as_dict)
frequency_bias = model_eval.get_frequency_bias(contingency_table_as_dict)
peirce_score = model_eval.get_peirce_score(contingency_table_as_dict)
heidke_score = model_eval.get_heidke_score(contingency_table_as_dict)
print (
'POD = {0:.4f} ... POFD = {1:.4f} ... success ratio = {2:.4f} ... '
'FOCN = {3:.4f} ... accuracy = {4:.4f} ... CSI = {5:.4f} ... frequency '
'bias = {6:.4f} ... Peirce score = {7:.4f} ... Heidke score = {8:.4f}\n'
).format(pod, pofd, success_ratio, focn, accuracy, csi, frequency_bias,
peirce_score, heidke_score)
auc, scikit_learn_auc = _create_roc_curve(
forecast_probabilities=forecast_probabilities,
observed_labels=observed_labels, output_dir_name=output_dir_name)
print '\n'
bss_dict = _create_attributes_diagram(
forecast_probabilities=forecast_probabilities,
observed_labels=observed_labels, output_dir_name=output_dir_name)
print '\n'
aupd = _create_performance_diagram(
forecast_probabilities=forecast_probabilities,
observed_labels=observed_labels, output_dir_name=output_dir_name)
print '\n'
evaluation_file_name = '{0:s}/model_evaluation.p'.format(output_dir_name)
print 'Writing results to: "{0:s}"...'.format(evaluation_file_name)
model_eval.write_results(
forecast_probabilities=forecast_probabilities,
observed_labels=observed_labels,
binarization_threshold=binarization_threshold, pod=pod, pofd=pofd,
success_ratio=success_ratio, focn=focn, accuracy=accuracy, csi=csi,
frequency_bias=frequency_bias, peirce_score=peirce_score,
heidke_score=heidke_score, auc=auc, scikit_learn_auc=scikit_learn_auc,
aupd=aupd, bss_dict=bss_dict, pickle_file_name=evaluation_file_name)