def run_cross_validation(settings, targets, pipelines, mask_range, split_ratios, classifiers):
pool = Pool(settings.N_jobs)
for i, pipeline in enumerate(pipelines):
for j, (classifier, classifier_name) in enumerate(classifiers):
for k, target in enumerate(targets):
pool.apply_async(cross_validation_score, [settings, target, pipeline, classifier, classifier_name], {'quiet': True})
for split_num, split_ratio in enumerate(split_ratios):
masks = generate_feature_masks(settings, target, pipeline, np.max(mask_range), split_ratio, random_state=0, quiet=True)
for mask_num, mask in enumerate(masks):
progress_str = 'P=%d/%d C=%d/%d T=%d/%d S=%d/%d M=%d/%d' % (i+1, len(pipelines), j+1, len(classifiers), k+1, len(targets), split_num+1, len(split_ratios), mask_num+1, len(masks))
cross_validation_score(settings, target, pipeline, classifier, classifier_name, feature_mask=mask, quiet=True, return_data=False, pool=pool, progress_str=progress_str)
pool.close()
pool.join()
print 'Finished cross validation mp'
summaries = []
for p_num, pipeline in enumerate(pipelines):
for classifier, classifier_name in classifiers:
scores_full = []
scores_masked = [[[] for y in mask_range] for x in split_ratios]
for i, target in enumerate(targets):
run_prepare_data_for_cross_validation(settings, [target], [pipeline], quiet=True)
data = cross_validation_score(settings, target, pipeline, classifier, classifier_name, pool=None, quiet=True)
scores_full.append(data.mean_score)
for split_index, split_ratio in enumerate(split_ratios):
masks = generate_feature_masks(settings, target, pipeline, np.max(mask_range), split_ratio, random_state=0, quiet=True)
for mask_index, num_masks in enumerate(mask_range):
predictions = []
y_cvs = None
for mask in masks[0:num_masks]:
data = cross_validation_score(settings, target, pipeline, classifier, classifier_name, feature_mask=mask, pool=None, quiet=True)
predictions.append(data.mean_predictions)
if y_cvs is None:
y_cvs = data.y_cvs
else:
for y_cv_1, y_cv_2 in zip(y_cvs, data.y_cvs):
assert np.alltrue(y_cv_1 == y_cv_2)
predictions = np.mean(predictions, axis=0)
scores = [roc_auc_score(y_cv, p) for p, y_cv in zip(predictions, y_cvs)]
score = np.mean(scores)
scores_masked[split_index][mask_index].append(score)
summary = get_score_summary('%s p=%d full' % (classifier_name, p_num), scores_full, np.mean(scores_full), targets)
summaries.append((summary, np.mean(scores_full)))
for split_index, split_ratio in enumerate(split_ratios):
for mask_index, num_masks in enumerate(mask_range):
scores = scores_masked[split_index][mask_index]
summary = get_score_summary('%s p=%d split_ratio=%s masks=%d' % (classifier_name, p_num, split_ratio, num_masks), scores, np.mean(scores), targets)
summaries.append((summary, np.mean(scores)))
print summary
print_results(summaries)
def run_cross_validation(settings, targets, pipelines, mask_range, split_ratios, classifiers):
pool = Pool(settings.N_jobs)
for i, pipeline in enumerate(pipelines):
for j, (classifier, classifier_name) in enumerate(classifiers):
for k, target in enumerate(targets):
pool.apply_async(cross_validation_score, [settings, target, pipeline, classifier, classifier_name], {'quiet': True})
for split_num, split_ratio in enumerate(split_ratios):
masks = generate_feature_masks(settings, target, pipeline, np.max(mask_range), split_ratio, random_state=0, quiet=True)
for mask_num, mask in enumerate(masks):
progress_str = 'P=%d/%d C=%d/%d T=%d/%d S=%d/%d M=%d/%d' % (i+1, len(pipelines), j+1, len(classifiers), k+1, len(targets), split_num+1, len(split_ratios), mask_num+1, len(masks))
cross_validation_score(settings, target, pipeline, classifier, classifier_name, feature_mask=mask, quiet=True, return_data=False, pool=pool, progress_str=progress_str)
pool.close()
pool.join()
print('Finished cross validation mp')
summaries = []
for p_num, pipeline in enumerate(pipelines):
for classifier, classifier_name in classifiers:
scores_full = []
scores_masked = [[[] for y in mask_range] for x in split_ratios]
for i, target in enumerate(targets):
run_prepare_data_for_cross_validation(settings, [target], [pipeline], quiet=True)
data = cross_validation_score(settings, target, pipeline, classifier, classifier_name, pool=None, quiet=True)
scores_full.append(data.mean_score)
for split_index, split_ratio in enumerate(split_ratios):
masks = generate_feature_masks(settings, target, pipeline, np.max(mask_range), split_ratio, random_state=0, quiet=True)
for mask_index, num_masks in enumerate(mask_range):
predictions = []
y_cvs = None
for mask in masks[0:num_masks]:
data = cross_validation_score(settings, target, pipeline, classifier, classifier_name, feature_mask=mask, pool=None, quiet=True)
predictions.append(data.mean_predictions)
if y_cvs is None:
y_cvs = data.y_cvs
else:
for y_cv_1, y_cv_2 in zip(y_cvs, data.y_cvs):
assert np.alltrue(y_cv_1 == y_cv_2)
predictions = np.mean(predictions, axis=0)
scores = [roc_auc_score(y_cv, p) for p, y_cv in zip(predictions, y_cvs)]
score = np.mean(scores)
scores_masked[split_index][mask_index].append(score)
summary = get_score_summary('%s p=%d full' % (classifier_name, p_num), scores_full)
summaries.append((summary, np.mean(scores_full)))
for split_index, split_ratio in enumerate(split_ratios):
for mask_index, num_masks in enumerate(mask_range):
scores = scores_masked[split_index][mask_index]
summary = get_score_summary('%s p=%d split_ratio=%s masks=%d' % (classifier_name, p_num, split_ratio, num_masks), scores)
summaries.append((summary, np.mean(scores)))
print(summary)
print_results(summaries)
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 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)
