def test_predictor(predictor_cls):
predictor = predictor_cls()
data_path = Global.path_map('clips_folder')
# arbritary
band_means = np.linspace(0, 200, 66)
band_width = 2
feature_extractor = FFTFeatures(band_means=band_means,
band_width=band_width)
feature_extractor = ARFeatures()
loader = DataLoader(data_path, feature_extractor)
X_list = loader.training_data("Dog_1")
y_list = loader.labels("Dog_1")
print(
XValidation.evaluate(X_list, y_list[0], predictor,
evaluation=accuracy))
# Set the conditioned results for proper evaluation
conditioned = [a * b for (a, b) in zip(y_list[0], y_list[1])]
print(
XValidation.evaluate(X_list,
conditioned,
predictor,
evaluation=accuracy))
def __init__(self, feature_extractors, predictors, patient,
data_path=Global.path_map('clips_folder')):
assert(type(feature_extractors)==type([]))
assert(type(predictors)==type([]))
assert(isinstance(patient, basestring))
self._feature_extractors = feature_extractors
self._predictors = predictors
self._patient = patient
self._data_path = data_path
def test_predictor(predictor_cls, patient_name='Dog_1'):
''' function that loads data for Dog_1 run crossvalidation with ARFeatures
INPUT:
- predictor_cls: a Predictor class (implement)
'''
# instanciating a predictor object from Predictor class
predictor = predictor_cls()
# path to data (here path from within gatsby network)
data_path = Global.path_map('clips_folder')
# creating instance of autoregressive features
#feature_extractor = ARFeatures()
band_means = np.linspace(0, 200, 66)
band_width = 2
FFTFeatures_args = {'band_means': band_means, 'band_width': band_width}
# feature_extractor = MixFeatures([{'name':"ARFeatures",'args':{}},
# {'name':"FFTFeatures",'args':FFTFeatures_args}])
feature_extractor = MixFeatures([{'name': "ARFeatures", 'args': {}}])
# feature_extractor = MixFeatures([{'name':"FFTFeatures",'args':FFTFeatures_args}])
#feature_extractor = ARFeatures()
# loading the data
loader = DataLoader(data_path, feature_extractor)
print(loader.base_dir)
print('\npatient = %s' % patient_name)
X_list = loader.training_data(patient_name)
y_list = loader.labels(patient_name)
# separating the label
early_vs_not = y_list[1] #[a * b for (a, b) in zip(y_list[0], y_list[1])]
seizure_vs_not = y_list[0]
# running cross validation
# conditioned = [a * b for (a, b) in zip(y_list[0], y_list[1])]
print("\ncross validation: seizures vs not")
result = XValidation.evaluate(X_list,
seizure_vs_not,
predictor,
evaluation=auc)
print('cross-validation results: mean = %.3f, sd = %.3f, raw scores = %s' \
% (np.mean(result), np.std(result), result))
print("\ncross validation: early_vs_not")
result = XValidation.evaluate(X_list,
early_vs_not,
predictor,
evaluation=auc)
print('cross-validation results: mean = %.3f, sd = %.3f, raw scores = %s' \
% (np.mean(result), np.std(result), result))
def __init__(self,
patient_name,
base_dir=Global.path_map('clips_folder'),
use_cache=True,
max_train_segments=-1,
max_test_segments=-1):
"""
patient_name: for example, Dog_1
base_dir: path to the directory containing patient folders i.e., directory
containing Dog_1/, Dog_2, ..., Patient_1, Patient_2, ....
use_cache: if True, the loaded data is retained in memory so that the
call to data loading method will return immediately next time.
Require a large amount of memory.
max_XX_segments: maximum segments to load. -1 to use the number of
total segments available. Otherwise, all segments (ictal and interictal)
will be randomly subsampled without replacement.
"""
if not os.path.isdir(base_dir):
raise ValueError('%s is not a directory.' % base_dir)
# The followings attributes, once assigned, will not change.
self.base_dir = base_dir
# patient_name = e.g., Dog_1
self.patient_name = patient_name
self.use_cache = use_cache
self.max_train_segments = max_train_segments
self.max_test_segments = max_test_segments
# this will be cached when get_train_data() is called. a list of training
# file names
self.loaded_train_fnames = None
# a list of loaded (Instance, y_seizure, y_early)
self.loaded_train_data = None
# this will be cached when get_test_data() is called. a list of test
# file names
self.loaded_test_fnames = None
self.loaded_test_data = None
# type_labels = a list of {0, 1}. Indicators of a seizure (1 for seizure).
self.type_labels = None
# early_labels = a list of {0, 1}. Indicators of an early seizure
# (1 for an early seizure).
self.early_labels = None
self.params = {
'anti_alias_cutoff': 500.,
'anti_alias_width': 30.,
'anti_alias_attenuation': 40,
'elec_noise_width': 3.,
'elec_noise_attenuation': 60.0,
'elec_noise_cutoff': [59., 61.],
'targetrate': 500
}
def __init__(self,
feature_extractors,
predictors,
patient,
data_path=Global.path_map('clips_folder')):
assert (type(feature_extractors) == type([]))
assert (type(predictors) == type([]))
assert (isinstance(patient, str))
self._feature_extractors = feature_extractors
self._predictors = predictors
self._patient = patient
self._data_path = data_path
def __init__(self, feature_extractor, predictor, patient,
data_path=Global.path_map('clips_folder')):
"""
feature_extractor: an instance of FeatureExtractBase
predictor: an instance of PredictorBase
patient: a string indicating a subject e.g., Dog_1
"""
assert(isinstance(feature_extractor, FeatureExtractBase))
assert(isinstance(predictor, PredictorBase))
assert(isinstance(patient, basestring))
self._feature_extractor = feature_extractor
self._predictor = predictor
self._patient = patient
self._data_path = data_path
def test_predictor(predictor_cls, patient_name='Dog_1'):
''' function that loads data for Dog_1 run crossvalidation with ARFeatures
INPUT:
- predictor_cls: a Predictor class (implement)
'''
# instanciating a predictor object from Predictor class
predictor = predictor_cls()
# path to data (here path from within gatsby network)
data_path = Global.path_map('clips_folder')
# creating instance of autoregressive features
#feature_extractor = ARFeatures()
band_means = np.linspace(0, 200, 66)
band_width = 2
FFTFeatures_args = {'band_means':band_means, 'band_width':band_width}
# feature_extractor = MixFeatures([{'name':"ARFeatures",'args':{}},
# {'name':"FFTFeatures",'args':FFTFeatures_args}])
feature_extractor = MixFeatures([{'name':"ARFeatures",'args':{}}])
# feature_extractor = MixFeatures([{'name':"FFTFeatures",'args':FFTFeatures_args}])
#feature_extractor = ARFeatures()
# loading the data
loader = DataLoader(data_path, feature_extractor)
print loader.base_dir
print '\npatient = %s' % patient_name
X_list = loader.training_data(patient_name)
y_list = loader.labels(patient_name)
# separating the label
early_vs_not = y_list[1] #[a * b for (a, b) in zip(y_list[0], y_list[1])]
seizure_vs_not = y_list[0]
# running cross validation
# conditioned = [a * b for (a, b) in zip(y_list[0], y_list[1])]
print "\ncross validation: seizures vs not"
result = XValidation.evaluate(X_list, seizure_vs_not, predictor, evaluation=auc)
print 'cross-validation results: mean = %.3f, sd = %.3f, raw scores = %s' \
% (np.mean(result), np.std(result), result)
print "\ncross validation: early_vs_not"
result = XValidation.evaluate(X_list, early_vs_not, predictor, evaluation=auc)
print 'cross-validation results: mean = %.3f, sd = %.3f, raw scores = %s' \
% (np.mean(result), np.std(result), result)
def __init__(self,
feature_extractor,
predictor,
patient,
data_path=Global.path_map('clips_folder')):
"""
feature_extractor: an instance of FeatureExtractBase
predictor: an instance of PredictorBase
patient: a string indicating a subject e.g., Dog_1
"""
assert (isinstance(feature_extractor, FeatureExtractBase))
assert (isinstance(predictor, PredictorBase))
assert (isinstance(patient, str))
self._feature_extractor = feature_extractor
self._predictor = predictor
self._patient = patient
self._data_path = data_path
def __init__(self, data_path,patients=None):
"""
Constructor
Parameters:
data_path - / terminated path. This is the base folder
containing e.g., Dog_1/, Dog_2/, ....
patients - a list of patient names e.g., ['Dog_1', 'Patient_2', ...]
"""
if not os.path.isdir(data_path):
raise ValueError('%s is not a directory.'%data_path)
self.data_path = Global.path_map('clips_folder')
#self.data_path = '/nfs/data3/kaggle_seizure/clips/'
if patients == None:
self.patients = ["Dog_%d" % i for i in range(1, 5)] + ["Patient_%d" % i for i in range(1, 9)]
else:
self.patients = [patients] # will only work for single subject here...
def Xval_on_single_patient(predictor_cls,
feature_extractor,
patient_name="Dog_1",
preprocess=True):
"""
Single patient cross validation
Returns 2 lists of cross validation performances
:param predictor_cls:
:param feature_extractor
:param patient_name:
:return:
"""
# predictor_cls is a handle to an instance of PredictorBase
# Instantiate the predictor
predictor = predictor_cls()
base_dir = Global.path_map('clips_folder')
base_dir = '/nfs/data3/kaggle_seizure/clips/'
loader = DataLoader(base_dir, feature_extractor)
X_list, y_seizure, y_early = loader.blocks_for_Xvalidation(
patient_name, preprocess=preprocess)
#X_train,y_seizure, y_early = loader.training_data(patient_name)
#y_train = [y_seizure,y_early]
#X_list,y_list = train_test_split(X_train,y_train)
# running cross validation
print(patient_name)
print("\ncross validation: seizures vs not")
result_seizure = XValidation.evaluate(X_list,
y_seizure,
predictor,
evaluation=auc)
print('cross-validation results: mean = %.3f, sd = %.3f, raw scores = %s' \
% (np.mean(result_seizure), np.std(result_seizure), result_seizure))
print("\ncross validation: early_vs_not")
result_early = XValidation.evaluate(X_list,
y_early,
predictor,
evaluation=auc)
print('cross-validation results: mean = %.3f, sd = %.3f, raw scores = %s' \
% (np.mean(result_early), np.std(result_early), result_early))
return result_seizure, result_early
def test_predictor(predictor_cls):
predictor = predictor_cls()
data_path = Global.path_map('clips_folder')
# arbritary
band_means = np.linspace(0, 200, 66)
band_width = 2
feature_extractor = FFTFeatures(band_means=band_means, band_width=band_width)
feature_extractor = ARFeatures()
loader = DataLoader(data_path, feature_extractor)
X_list = loader.training_data("Dog_1")
y_list = loader.labels("Dog_1")
print XValidation.evaluate(X_list, y_list[0], predictor, evaluation=accuracy)
# Set the conditioned results for proper evaluation
conditioned = [a * b for (a, b) in zip(y_list[0], y_list[1])]
print XValidation.evaluate(X_list, conditioned, predictor, evaluation=accuracy)
def __init__(self, data_path, patients=None):
"""
Constructor
Parameters:
data_path - / terminated path. This is the base folder
containing e.g., Dog_1/, Dog_2/, ....
patients - a list of patient names e.g., ['Dog_1', 'Patient_2', ...]
"""
if not os.path.isdir(data_path):
raise ValueError('%s is not a directory.' % data_path)
self.data_path = Global.path_map('clips_folder')
#self.data_path = '/nfs/data3/kaggle_seizure/clips/'
if patients == None:
self.patients = ["Dog_%d" % i for i in range(1, 5)
] + ["Patient_%d" % i for i in range(1, 9)]
else:
self.patients = [patients
] # will only work for single subject here...
def visualise():
data_path = Global.path_map('clips_folder')
# arbritary
band_means = np.linspace(0, 200, 66)
band_width = 2
feature_extractor = FFTFeatures(band_means=band_means, band_width=band_width)
loader = DataLoader(data_path, feature_extractor)
X_list = loader.training_data("Dog_1")
y_list = loader.labels("Dog_1")[0]
plt.figure()
for i in range(len(X_list)):
X = X_list[i]
y_seizure = y_list[i]
_, _, V = np.linalg.svd(X, full_matrices=True)
plt.plot(V[0][y_seizure == 0], V[1][y_seizure == 0], 'bo')
plt.plot(V[0][y_seizure == 1], V[1][y_seizure == 1], 'ro')
plt.show()
def visualise():
data_path = Global.path_map('clips_folder')
# arbritary
band_means = np.linspace(0, 200, 66)
band_width = 2
feature_extractor = FFTFeatures(band_means=band_means, band_width=band_width)
loader = DataLoader(data_path, feature_extractor)
X_list = loader.training_data("Dog_1")
y_list = loader.labels("Dog_1")[0]
plt.figure()
for i in range(len(X_list)):
X = X_list[i]
y_seizure = y_list[i]
_, _, V = np.linalg.svd(X, full_matrices=True)
plt.plot(V[0][y_seizure == 0], V[1][y_seizure == 0], 'bo')
plt.plot(V[0][y_seizure == 1], V[1][y_seizure == 1], 'ro')
plt.show()
import numpy as np
from seizures.submission import SubmissionFile
from seizures.prediction.SVMPredictor import SVMPredictor
from seizures.features.MixFeatures import MixFeatures
from seizures.Global import Global
# Example script to generate submission file
#data_path = "/nfs/data3/kaggle_seizure/scratch/Stiched_data/Dog_1/"
data_path = Global.get_subject_folder('Dog_1')
# Define Predictor
predictor_seizure = SVMPredictor
predictor_early = SVMPredictor
# Define Features
band_means = np.linspace(0, 200, 66)
band_width = 2
FFTFeatures_args = {'band_means':band_means, 'band_width':band_width}
#feature_extractor = MixFeatures([{'name':"ARFeatures",'args':{}},
# {'name':"FFTFeatures",'args':FFTFeatures_args}])
feature_extractor = MixFeatures([{'name':"ARFeatures",'args':{}},
{'name':"PLVFeatures",'args':{}}])
submissionfile = SubmissionFile(data_path)
# Load training data
# Learn classifiers
# Make final file
submissionfile.generate_submission(predictor_seizure, predictor_early,
'''
Created on 28 Jun 2014
@author: heiko
'''
from seizures.features.RandomFeatures import RandomFeatures
from seizures.prediction.RandomPredictor import RandomPredictor
from seizures.submission.SubmissionFile import SubmissionFile
from seizures.Global import Global
if __name__ == '__main__':
predictor = RandomPredictor()
extractor = RandomFeatures()
test_files = ["Dog_1_test_segment_1.mat"]
data_path = Global.path_map('clips_folder')
submission = SubmissionFile(data_path)
predictor_seizure = predictor
predictor_early = predictor
submission.generate_submission(predictor_seizure,
predictor_early,
extractor,
test_filenames=test_files)
'''
Created on 28 Jun 2014
@author: heiko
'''
from seizures.features.RandomFeatures import RandomFeatures
from seizures.prediction.RandomPredictor import RandomPredictor
from seizures.submission.SubmissionFile import SubmissionFile
from seizures.Global import Global
if __name__ == '__main__':
predictor = RandomPredictor()
extractor = RandomFeatures()
test_files = ["Dog_1_test_segment_1.mat"]
data_path = Global.path_map('clips_folder')
submission = SubmissionFile(data_path)
predictor_seizure = predictor
predictor_early = predictor
submission.generate_submission(predictor_seizure, predictor_early,
extractor, test_filenames=test_files)
def generate_submission(self,
predictor_seizure,
predictor_early,
feature_extractor,
output_fname="output.csv",
test_filenames=None,
preprocess=True):
"""
Generates a submission file for a given pair of predictors, which will
be trained on all training data per patient/dog instance.
Parameters:
predictor_seizure - Instance of PredictorBase, fixed parameters
predictor_early - Instance of PredictorBase, fixed parameters
feature_extractor - Instance of FeatureExtractBase, to extract test features
output_fname - Optional filename for result submission file
test_filename - Optional list of filenames to produce results on,
default is to use all
"""
# make sure given objects are valid
assert (isinstance(predictor_seizure, PredictorBase))
assert (isinstance(predictor_early, PredictorBase))
assert (isinstance(feature_extractor, FeatureExtractBase))
test_filenames = SubmissionFile.get_submission_filenames()
train_filenames = SubmissionFile.get_train_filenames()
all_result_lines = []
for patient in self.patients:
result_lines = []
loader = DataLoader(self.data_path, feature_extractor)
# X_train is n x d
X_train, y_seizure, y_early = loader.training_data(
patient, preprocess=preprocess)
print X_train.shape
print y_seizure.shape
# train both models
#print
print "Training seizure for " + patient
predictor_seizure.fit(X_train, y_seizure)
print "Training early for " + patient
predictor_early.fit(X_train, y_early)
pred_seizure = predictor_seizure.predict(X_train)
pred_early = predictor_early.predict(X_train)
print 'Results on training data'
print 'seizure\tearly\tp(seizure)\tp(early)'
for y_1, y_2, p_1, p_2 in izip(y_seizure, y_early, pred_seizure,
pred_early):
print '%d\t%d\t%.3f\t%.3f' % (y_1, y_2, p_1, p_2)
# find out filenames that correspond to patient/dog
test_fnames_patient = []
for fname in test_filenames:
if patient in fname:
test_fnames_patient += [fname]
# now predict on all test points
loader = DataLoader(self.data_path, feature_extractor)
# X_test: n x d matrix
X_test = loader.test_data(patient, preprocess=preprocess)
test_fnames_patient = loader.files
for ifname in range(len(test_fnames_patient)):
fname = test_fnames_patient[ifname]
# X is one instance
X = X_test[ifname, :]
# [0] to extract probability out of the ndarray
pred_seizure = predictor_seizure.predict(X)[0]
pred_early = predictor_early.predict(X)[0]
name = fname.split("/")[-1]
result_lines.append(",".join(
[name, str(pred_seizure),
str(pred_early)]))
csv_fname = patient + '_' + output_fname + '.csv'
csv_path = Global.get_child_result_folder(csv_fname)
print "Storing results to", csv_fname
f = open(csv_path, "w")
f.write("clip,seizure,early\n")
for line in result_lines:
f.write(line + '\n')
f.close()
all_result_lines.append(result_lines)
def generate_submission(self, predictor_seizure, predictor_early,
feature_extractor, output_fname="output.csv",
test_filenames=None, preprocess=True):
"""
Generates a submission file for a given pair of predictors, which will
be trained on all training data per patient/dog instance.
Parameters:
predictor_seizure - Instance of PredictorBase, fixed parameters
predictor_early - Instance of PredictorBase, fixed parameters
feature_extractor - Instance of FeatureExtractBase, to extract test features
output_fname - Optional filename for result submission file
test_filename - Optional list of filenames to produce results on,
default is to use all
"""
# make sure given objects are valid
assert(isinstance(predictor_seizure, PredictorBase))
assert(isinstance(predictor_early, PredictorBase))
assert(isinstance(feature_extractor, FeatureExtractBase))
test_filenames = SubmissionFile.get_submission_filenames()
train_filenames = SubmissionFile.get_train_filenames()
all_result_lines = []
for patient in self.patients:
result_lines = []
loader = DataLoader(self.data_path, feature_extractor)
# X_train is n x d
X_train,y_seizure, y_early = loader.training_data(patient,preprocess=preprocess)
print X_train.shape
print y_seizure.shape
# train both models
#print
print "Training seizure for " + patient
predictor_seizure.fit(X_train, y_seizure)
print "Training early for " + patient
predictor_early.fit(X_train, y_early)
pred_seizure = predictor_seizure.predict(X_train)
pred_early = predictor_early.predict(X_train)
print 'Results on training data'
print 'seizure\tearly\tp(seizure)\tp(early)'
for y_1, y_2, p_1, p_2 in izip(y_seizure, y_early, pred_seizure, pred_early):
print '%d\t%d\t%.3f\t%.3f' % (y_1, y_2, p_1, p_2)
# find out filenames that correspond to patient/dog
test_fnames_patient = []
for fname in test_filenames:
if patient in fname:
test_fnames_patient += [fname]
# now predict on all test points
loader = DataLoader(self.data_path, feature_extractor)
# X_test: n x d matrix
X_test = loader.test_data(patient,preprocess=preprocess)
test_fnames_patient = loader.files
for ifname in range(len(test_fnames_patient)):
fname = test_fnames_patient[ifname]
# X is one instance
X = X_test[ifname,:]
# [0] to extract probability out of the ndarray
pred_seizure = predictor_seizure.predict(X)[0]
pred_early = predictor_early.predict(X)[0]
name = fname.split("/")[-1]
result_lines.append(",".join([name, str(pred_seizure), str(pred_early)]))
csv_fname = patient + '_' + output_fname + '.csv'
csv_path = Global.get_child_result_folder(csv_fname)
print "Storing results to", csv_fname
f = open(csv_path, "w")
f.write("clip,seizure,early\n")
for line in result_lines:
f.write(line + '\n')
f.close()
all_result_lines.append(result_lines)