def get_submission_targets_and_masks(settings, targets, classifier, classifier_name, pipeline_groups, random_pipelines, random_ratio=0.525, ngen=10, limit=2, random_limit=2):
assert random_limit % limit == 0
random_multiplier = random_limit / limit
quiet = True
random_pipeline = FeatureConcatPipeline(*random_pipelines)
all_pipelines = []
all_pipelines.extend(random_pipelines)
for pg, ratio in pipeline_groups:
all_pipelines.extend(pg)
full_pipeline = FeatureConcatPipeline(*all_pipelines)
run_prepare_data(settings, [(target, full_pipeline, []) for target in targets], test=True)
def get_pipeline_and_feature_masks(target, pipelines, classifier, classifier_name, ratio, ngen):
print target, 'fetching GA pipelines', [p.get_name() for p in pipelines]
pipeline = FeatureConcatPipeline(*pipelines)
score, best_N = process_target(settings, target, pipeline, classifier, classifier_name, ratio=ratio, ngen=ngen, quiet=quiet)
return pipeline, best_N
targets_and_pipelines = []
for target in targets:
# NOTE(mike): All this stuff is a bit nasty. It gets the random-masks and the genetic-masks
# for different pipelines, and then pulls out the mask for each individual pipeline. A single
# FeatureConcatPipeline is then created to represent all the features, and the masks for each
# member of the FCP are merged together to form the single feature mask across the whole FCP.
random_masks = generate_feature_masks(settings, target, random_pipeline, random_limit, random_ratio, random_state=0, quiet=quiet)
# contains a list of pairs, (pipeline, mask)
ga_groups = [get_pipeline_and_feature_masks(target, p, classifier, classifier_name, ratio, ngen) for p, ratio in pipeline_groups]
ga_groups = [(p, masks[0:limit]) for p, masks in ga_groups]
print target, 'extracting GA per-pipeline masks...'
# contains a list of mask dictionaries
ga_dicts = [extract_masks_for_pipeline_and_masks(settings, target, pipeline, masks) for pipeline, masks in ga_groups]
ga_dicts = [mask_dicts * random_multiplier for mask_dicts in ga_dicts]
r_dicts = extract_masks_for_pipeline_and_masks(settings, target, random_pipeline, random_masks)
# this contains a list of dictionaries which maps pipeline names to masks
# e.g. [r_dicts, ga_dicts0, ga_dicts1, ...]
zip_group = [r_dicts] + ga_dicts
print target, 'merging all masks...'
feature_mask_dicts = [merge_dicts(*x) for x in zip(*zip_group)]
feature_masks = []
for feature_mask_dict in feature_mask_dicts:
mask = []
for p in full_pipeline.get_pipelines():
mask.extend(feature_mask_dict[p.get_name()])
feature_masks.append(mask)
targets_and_pipelines.append((target, full_pipeline, feature_masks))
return targets_and_pipelines
def get_pipeline_and_feature_masks(target, pipelines, classifier,
classifier_name, ratio, ngen):
print(target, 'fetching GA pipelines',
[p.get_name() for p in pipelines])
pipeline = FeatureConcatPipeline(*pipelines)
score, best_N = process_target(settings,
target,
pipeline,
classifier,
classifier_name,
ratio=ratio,
ngen=ngen,
quiet=quiet)
return pipeline, best_N
def main():
settings = load_settings()
targets = [
'Dog_1', 'Dog_2', 'Dog_3', 'Dog_4', 'Dog_5', 'Patient_1', 'Patient_2'
]
# The genetic algorithm will be run individually on each pipeline group
pipeline_groups = [
([
Pipeline(InputSource(), Preprocess(), Windower(75), PFD()),
], 0.55),
([
Pipeline(InputSource(), Preprocess(), Windower(75), Hurst()),
], 0.55),
([
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy(
[0.25, 1, 1.75, 2.5, 3.25, 4, 5, 8.5, 12, 15.5, 19.5,
24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5, 6, 15, 24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5, 6, 15])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([6, 15, 24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([2, 3.5, 6])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([3.5, 6, 15])),
Pipeline(InputSource(), Preprocess(), Windower(75), HFD(2)),
], 0.55),
]
make_submission = len(sys.argv) >= 2 and sys.argv[1] == 'submission'
run_ga = not make_submission
# This classifier is used in the genetic algorithm
ga_classifier, ga_classifier_name = make_svm(gamma=0.0079, C=2.7)
if run_ga:
quiet = False
summaries = []
for ngen in [10]:
for pipelines, ratio in pipeline_groups:
out = []
for target in targets:
print 'Running target', target
run_prepare_data_for_cross_validation(settings, [target],
pipelines,
quiet=True)
pipeline = FeatureConcatPipeline(*pipelines)
score, best_N = process_target(settings,
target,
pipeline,
ga_classifier,
ga_classifier_name,
ratio=ratio,
ngen=ngen,
quiet=quiet)
print target, score, [
np.sum(mask) for mask in best_N[0:10]
]
out.append((target, score, pipeline, best_N))
scores = np.array([score for _, score, _, _ in out])
summary = get_score_summary(
'%s ngen=%d' % (ga_classifier_name, ngen), scores,
np.mean(scores), targets)
summaries.append((summary, np.mean(scores)))
print summary
print_results(summaries)
if make_submission:
random_pipelines = [
Pipeline(InputSource(), Preprocess(), Windower(75),
Correlation('none')),
Pipeline(InputSource(), Preprocess(), Windower(75),
FreqCorrelation(1, None, 'none')),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
FreqBinning(winning_bins, 'mean'), Log10(), FlattenChannels()),
]
# These classifiers are used to make the final predictions
final_classifiers = [
# make_svm(gamma=0.0079, C=2.7),
make_svm(gamma=0.0068, C=2.0),
# make_svm(gamma=0.003, C=150.0),
# make_lr(C=0.04),
# make_simple_lr(),
]
targets_and_pipelines = get_submission_targets_and_masks(
settings, targets, ga_classifier, ga_classifier_name,
pipeline_groups, random_pipelines)
for classifier, classifier_name in final_classifiers:
run_make_submission(settings, targets_and_pipelines, classifier,
classifier_name)
def get_submission_targets_and_masks(settings,
targets,
classifier,
classifier_name,
pipeline_groups,
random_pipelines,
random_ratio=0.525,
ngen=10,
limit=2,
random_limit=2):
assert random_limit % limit == 0
random_multiplier = random_limit / limit
quiet = True
random_pipeline = FeatureConcatPipeline(*random_pipelines)
all_pipelines = []
all_pipelines.extend(random_pipelines)
for pg, ratio in pipeline_groups:
all_pipelines.extend(pg)
full_pipeline = FeatureConcatPipeline(*all_pipelines)
run_prepare_data(settings,
[(target, full_pipeline, []) for target in targets],
test=True)
def get_pipeline_and_feature_masks(target, pipelines, classifier,
classifier_name, ratio, ngen):
print target, 'fetching GA pipelines', [
p.get_name() for p in pipelines
]
pipeline = FeatureConcatPipeline(*pipelines)
score, best_N = process_target(settings,
target,
pipeline,
classifier,
classifier_name,
ratio=ratio,
ngen=ngen,
quiet=quiet)
return pipeline, best_N
targets_and_pipelines = []
for target in targets:
# NOTE(mike): All this stuff is a bit nasty. It gets the random-masks and the genetic-masks
# for different pipelines, and then pulls out the mask for each individual pipeline. A single
# FeatureConcatPipeline is then created to represent all the features, and the masks for each
# member of the FCP are merged together to form the single feature mask across the whole FCP.
random_masks = generate_feature_masks(settings,
target,
random_pipeline,
random_limit,
random_ratio,
random_state=0,
quiet=quiet)
# contains a list of pairs, (pipeline, mask)
ga_groups = [
get_pipeline_and_feature_masks(target, p, classifier,
classifier_name, ratio, ngen)
for p, ratio in pipeline_groups
]
ga_groups = [(p, masks[0:limit]) for p, masks in ga_groups]
print target, 'extracting GA per-pipeline masks...'
# contains a list of mask dictionaries
ga_dicts = [
extract_masks_for_pipeline_and_masks(settings, target, pipeline,
masks)
for pipeline, masks in ga_groups
]
ga_dicts = [mask_dicts * random_multiplier for mask_dicts in ga_dicts]
r_dicts = extract_masks_for_pipeline_and_masks(settings, target,
random_pipeline,
random_masks)
# this contains a list of dictionaries which maps pipeline names to masks
# e.g. [r_dicts, ga_dicts0, ga_dicts1, ...]
zip_group = [r_dicts] + ga_dicts
print target, 'merging all masks...'
feature_mask_dicts = [merge_dicts(*x) for x in zip(*zip_group)]
feature_masks = []
for feature_mask_dict in feature_mask_dicts:
mask = []
for p in full_pipeline.get_pipelines():
mask.extend(feature_mask_dict[p.get_name()])
feature_masks.append(mask)
targets_and_pipelines.append((target, full_pipeline, feature_masks))
return targets_and_pipelines
def main():
settings = load_settings()
pipelines = [
FeatureConcatPipeline(
Pipeline(InputSource(), Preprocess(), Windower(75),
Correlation('none')),
Pipeline(InputSource(), Preprocess(), Windower(75),
FreqCorrelation(1, None, 'none')),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
FreqBinning(winning_bins, 'mean'), Log10(), FlattenChannels()),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy(
[0.25, 1, 1.75, 2.5, 3.25, 4, 5, 8.5, 12, 15.5, 19.5,
24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5, 6, 15, 24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5, 6, 15])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([6, 15, 24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([2, 3.5, 6])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([3.5, 6, 15])),
Pipeline(InputSource(), Preprocess(), Windower(75), HFD(2)),
Pipeline(InputSource(), Preprocess(), Windower(75), PFD()),
Pipeline(InputSource(), Preprocess(), Windower(75), Hurst()),
),
]
targets = [
'Dog_1', 'Dog_2', 'Dog_3', 'Dog_4', 'Dog_5', 'Patient_1', 'Patient_2'
]
classifiers = [
make_svm(gamma=0.0079, C=2.7),
make_svm(gamma=0.0068, C=2.0),
make_svm(gamma=0.003, C=150.0),
make_lr(C=0.04),
make_simple_lr(),
]
make_submission = len(sys.argv) >= 2 and sys.argv[1] == 'submission'
do_cv = not make_submission
if do_cv:
mask_range = [3]
split_ratios = [0.4, 0.525, 0.6]
run_prepare_data_for_cross_validation(settings, targets, pipelines)
run_cross_validation(settings, targets, pipelines, mask_range,
split_ratios, classifiers)
if make_submission:
num_masks = 10
split_ratio = 0.525
classifiers = [
# make_svm(gamma=0.0079, C=2.7),
make_svm(gamma=0.0068, C=2.0),
# make_svm(gamma=0.003, C=150.0),
# make_lr(C=0.04),
# make_simple_lr(),
]
targets_and_pipelines = []
pipeline = pipelines[0]
for classifier, classifier_name in classifiers:
for i, target in enumerate(targets):
run_prepare_data(settings, [target], [pipeline], test=True)
feature_masks = generate_feature_masks(settings,
target,
pipeline,
num_masks,
split_ratio,
random_state=0,
quiet=True)
targets_and_pipelines.append((target, pipeline, feature_masks,
classifier, classifier_name))
run_make_submission(settings, targets_and_pipelines, split_ratio)
def main():
settings = load_settings()
targets = [
'Dog_1', 'Dog_2', 'Dog_3', 'Dog_4', 'Dog_5', 'Patient_1', 'Patient_2'
]
pipelines = [
FeatureConcatPipeline(
Pipeline(InputSource(), Preprocess(), Windower(75),
Correlation('none')),
Pipeline(InputSource(), Preprocess(), Windower(75),
FreqCorrelation(1, None, 'none')),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
FreqBinning(winning_bins, 'mean'), Log10(), FlattenChannels()),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy(
[0.25, 1, 1.75, 2.5, 3.25, 4, 5, 8.5, 12, 15.5, 19.5,
24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5, 6, 15, 24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5, 6, 15])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([6, 15, 24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([2, 3.5, 6])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([3.5, 6, 15])),
Pipeline(InputSource(), Preprocess(), Windower(75), HFD(2)),
Pipeline(InputSource(), Preprocess(), Windower(75), PFD()),
Pipeline(InputSource(), Preprocess(), Windower(75), Hurst()),
),
]
classifiers = [
make_svm(gamma=0.0079, C=2.7),
make_svm(gamma=0.0068, C=2.0),
make_svm(gamma=0.003, C=150.0),
make_lr(C=0.04),
make_simple_lr(),
]
submission_pipelines = [
FeatureConcatPipeline(
Pipeline(InputSource(), Preprocess(), Windower(75),
Correlation('none')),
Pipeline(InputSource(), Preprocess(), Windower(75),
FreqCorrelation(1, None, 'none')),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
FreqBinning(winning_bins, 'mean'), Log10(), FlattenChannels()),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy(
[0.25, 1, 1.75, 2.5, 3.25, 4, 5, 8.5, 12, 15.5, 19.5,
24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5, 6, 15, 24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5, 6, 15])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([0.25, 2, 3.5])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([6, 15, 24])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([2, 3.5, 6])),
Pipeline(
InputSource(Preprocess(), Windower(75), FFT(), Magnitude()),
PIBSpectralEntropy([3.5, 6, 15])),
Pipeline(InputSource(), Preprocess(), Windower(75), HFD(2)),
Pipeline(InputSource(), Preprocess(), Windower(75), PFD()),
Pipeline(InputSource(), Preprocess(), Windower(75), Hurst()),
),
]
submission_classifiers = [
make_simple_lr(),
]
if len(sys.argv) >= 2 and sys.argv[1] == 'submission':
run_make_submission(settings, targets, submission_classifiers,
submission_pipelines)
else:
run_cross_validation(settings, targets, classifiers, pipelines)