def get_raw_matrix(self, data_tag):
if data_tag == "src":
return FeatureMatrixIO().read_file_to_data_frame(
self.raw_matrix_filepath)
else:
return FeatureMatrixIO().read_file_to_data_frame(
self.raw_matrix_filepath)
def _build_processed_feature_matrix_holdout(self):
fm_io = FeatureMatrixIO()
raw_matrix = fm_io.read_file_to_data_frame(
self._build_raw_matrix_path())
# if outcome_label in self.feat2imputed_dict:
# self.feat2imputed_dict.pop(outcome_label)
#
# processed_matrix = raw_matrix[self.feat2imputed_dict.keys()+[outcome_label]].copy()
'''
TODO: feat2imputed_dict includes the outcome label
'''
processed_matrix = raw_matrix[self.feat2imputed_dict.keys()].copy()
# TODO: tmp solution!
tmp_path = self._build_processed_matrix_path().replace(
"2000", "10000").replace("-holdout", "")
fm_io1 = FeatureMatrixIO()
processed_matrix_previous = fm_io1.read_file_to_data_frame(tmp_path)
processed_matrix = processed_matrix[processed_matrix_previous.columns]
# TODO: tmp solution!
for feat in self.feat2imputed_dict.keys():
processed_matrix[feat] = processed_matrix[feat].fillna(
self.feat2imputed_dict[feat])
fm_io.write_data_frame_to_file(processed_matrix, \
self._build_processed_matrix_path(), None)
def test_write_data_frame_to_file(self):
# Initialize FeatureMatrixIO.
fm_io = FeatureMatrixIO()
# Build paths for test files.
app_dir = os.path.dirname(os.path.abspath(inspect.getfile(inspect.currentframe())))
no_header_file_name = 'test-matrix-no-header.tab'
with_header_file_name = 'test-matrix-with-header.tab'
no_header_file_path = os.path.join(app_dir, no_header_file_name)
with_header_file_path = os.path.join(app_dir, with_header_file_name)
# Read data frames from test files.
matrix_no_header = MANUAL_TEST_CASE['matrix_no_header']
matrix_header = MANUAL_TEST_CASE['custom_header']
# Write data frame without header.
no_header_temp_file_name = 'no-header-temp-file.tab'
self._no_header_temp_file_path = os.path.join(app_dir, no_header_temp_file_name)
fm_io.write_data_frame_to_file(matrix_no_header, self._no_header_temp_file_path)
# Write data frame with header.
with_header_temp_file_name = 'header-temp-file.tab'
self._with_header_temp_file_path = os.path.join(app_dir, with_header_temp_file_name)
fm_io.write_data_frame_to_file(matrix_no_header, self._with_header_temp_file_path, matrix_header)
# Validate output files.
self.assertTrue(filecmp.cmp(no_header_file_path, self._no_header_temp_file_path))
self.assertTrue(filecmp.cmp(with_header_file_path, self._with_header_temp_file_path))
def _analyze_predictors_on_holdout(self):
fm_io = FeatureMatrixIO()
algorithms_to_test = list()
algorithms_to_test.extend(SupervisedClassifier.SUPPORTED_ALGORITHMS)
pipeline_file_name = inspect.getfile(inspect.currentframe())
data_dir = SupervisedLearningPipeline._fetch_data_dir_path(
self, pipeline_file_name)
# for algorithm in SupervisedClassifier.SUPPORTED_ALGORITHMS:
# algorithms_to_test.append('bifurcated-%s' % algorithm)
log.debug('algorithms_to_test: %s' % algorithms_to_test)
for algorithm in algorithms_to_test:
log.info('analyzing %s...' % algorithm)
# If report_dir does not exist, make it.
report_dir = '/'.join([data_dir, algorithm])
pipeline_prefix = '%s-normality-prediction-%s' % (self._var,
algorithm)
predictor_path = self._build_model_dump_path(algorithm)
if os.path.exists(
predictor_path) and 'bifurcated' not in algorithm:
log.debug('Loading model from disk...')
# TODO(sbala): Fix loblib.load so that it works for bifurcated
# supervised classifiers.
self._predictor = joblib.load(predictor_path)
# self._features = self._X_train.columns
status = SupervisedClassifier.TRAINED
SupervisedLearningPipeline._analyze_predictor_holdoutset(
self, report_dir, pipeline_prefix)
def _process_raw_feature_matrix(self):
# Read raw CMM.
self._fm_io = FeatureMatrixIO()
print('Reading raw matrix...')
self._cmm_raw = self._fm_io.read_file_to_data_frame(self._cmm_name_raw)
# Add and remove features to _cmm_processed.
self._fmt = FeatureMatrixTransform()
self._fmt.set_input_matrix(self._cmm_raw)
print('Adding features...')
self._add_features()
print('Imputing data...')
self._impute_data()
self._remove_features()
self._fmt.drop_duplicate_rows()
self._cmm_processed = self._fmt.fetch_matrix()
# Divide _cmm_processed into training and test data.
# This must happen before feature selection so that we don't
# accidentally learn information from the test data.
self._train_test_split()
print('Selecting features...')
self._select_features()
# Write output to new matrix.
train = self._y_train.join(self._X_train)
test = self._y_test.join(self._X_test)
self._cmm_processed = train.append(test)
header = self._build_processed_matrix_header()
self._fm_io.write_data_frame_to_file(self._cmm_processed,
self._cmm_name_processed, header)
def load_imputation_template(lab, dataset_folderpath, lab_type='panel'):
data_lab_folderpath = os.path.join(dataset_folderpath, lab)
imputations = pickle.load(
open(data_lab_folderpath + '/' + "feat2imputed_dict.pkl"))
if len(imputations) < 200: #
'''
only includes selected features
'''
return imputations
if lab_type == 'panel':
ylabel = 'all_components_normal'
else:
ylabel = 'component_normal'
'''
All raw matrix's columns are included. Have to extract final features from processed matrix
'''
fm_io = FeatureMatrixIO()
df_processed = fm_io.read_file_to_data_frame(
data_lab_folderpath + '/' + '%s-normality-matrix-processed.tab' % lab)
df_processed.pop('pat_id')
df_processed.pop(ylabel) # TODO?!
processed_columns_stanford = df_processed.columns.values.tolist()
imputations_new = {}
for i, col_selected in enumerate(processed_columns_stanford):
imputations_new[col_selected] = (i, imputations[col_selected])
return imputations_new
def load_raw_matrix(lab, dataset_folderpath):
data_lab_folderpath = os.path.join(dataset_folderpath, lab)
raw_matrix_filepath = os.path.join(data_lab_folderpath,
raw_matrix_template % lab)
fm_io = FeatureMatrixIO()
# TODO: check if raw matrix exists
raw_matrix = fm_io.read_file_to_data_frame(raw_matrix_filepath)
return raw_matrix
def _build_raw_feature_matrix(self):
raw_matrix_path = self._build_raw_matrix_path()
matrix_class = LabNormalityMatrix
SupervisedLearningPipeline._build_raw_feature_matrix(self, matrix_class, \
raw_matrix_path)
if not self._holdOut:
fm_io = FeatureMatrixIO()
matrix = fm_io.read_file_to_data_frame(raw_matrix_path)
self.usedPatIds = set(matrix['pat_id'].values)
def __init__(self, lab_panel, num_episodes, use_cache=None, random_state=None,
timeLimit=None, notUsePatIds=None, holdOut=False, pat_batch_ind=None, includeLastNormality=True):
# self.notUsePatIds = notUsePatIds
self.pat_batch_ind = pat_batch_ind
self._holdOut = holdOut
self.usedPatIds = []
SupervisedLearningPipeline.__init__(self, lab_panel, num_episodes, use_cache, random_state,
timeLimit, notUsePatIds)
# TODO: naming of lab_panel
self._factory = FeatureMatrixFactory()
self._build_raw_feature_matrix()
if LAB_TYPE == 'panel':
self.ylabel = 'all_components_normal'
else:
self.ylabel = 'component_normal'
self.includeLastNormality = includeLastNormality
if self.includeLastNormality:
fm_io = FeatureMatrixIO()
df = fm_io.read_file_to_data_frame('data/'+lab_panel+'/%s-normality-matrix-raw.tab'%lab_panel)
df = df.sort_values(['pat_id', 'order_time']).reset_index(drop=True)
df['last_normality'] = df['order_proc_id'].apply(lambda x:float('nan'))
for i in range(1,df.shape[0]):
if df.ix[i, 'pat_id'] == df.ix[i-1, 'pat_id']:
df.ix[i, 'last_normality'] = df.ix[i-1, self.ylabel]
df.to_csv('data/'+lab_panel+'/%s-normality-matrix-raw.tab'%lab_panel, index=False, sep='\t')
data_lab_folder = self._fetch_data_dir_path(inspect.getfile(inspect.currentframe()))
feat2imputed_dict_path = data_lab_folder + '/feat2imputed_dict.pkl'
if holdOut:
'''
For holdOut evaluation data, produce the raw matrix, pick
features according to the saved feat2imputed_dict.
'''
self.feat2imputed_dict = pickle.load(open(feat2imputed_dict_path, 'r'))
self._build_processed_feature_matrix_holdout()
self._analyze_predictors_on_holdout()
else:
'''
For training/validation data, record the pat_ids,
selected features and their imputed value correspondingly.
'''
pickle.dump(self.usedPatIds, open('data/used_patient_set_%s.pkl'%self._var, 'w'), pickle.HIGHEST_PROTOCOL)
self._build_processed_feature_matrix()
self._build_baseline_results() # TODO: prototype in SLPP
# return
# TODO: find better place to put the dict.pkl
pickle.dump(self.feat2imputed_dict, open(feat2imputed_dict_path, 'w'), pickle.HIGHEST_PROTOCOL)
self._train_and_analyze_predictors()
def _analyze_predictor_holdoutset(self, dest_dir, pipeline_prefix):
slugified_var = '-'.join(self._var.split())
holdout_path = dest_dir + '/../' + '%s-normality-matrix-%d-episodes-processed-holdout.tab' % (
slugified_var, self._num_rows)
fm_io = FeatureMatrixIO()
processed_matrix = fm_io.read_file_to_data_frame(holdout_path)
if self._isLabPanel:
y_holdout = pd.DataFrame(
processed_matrix.pop('all_components_normal'))
else:
y_holdout = pd.DataFrame(processed_matrix.pop('component_normal'))
X_holdout = processed_matrix
analyzer = ClassifierAnalyzer(self._predictor, X_holdout, y_holdout)
train_label = 'holdoutset'
# Build names for output plots and report.
direct_comparisons_name = '%s-direct-compare-results-%s.csv' % (
pipeline_prefix, train_label)
precision_at_k_plot_name = '%s-precision-at-k-plot-%s.png' % (
pipeline_prefix, train_label)
precision_recall_plot_name = '%s-precision-recall-plot-%s.png' % (
pipeline_prefix, train_label)
roc_plot_name = '%s-roc-plot-%s.png' % (pipeline_prefix, train_label)
report_name = '%s-report-%s.tab' % (pipeline_prefix, train_label)
# Build paths.
direct_comparisons_path = '/'.join([dest_dir, direct_comparisons_name])
log.debug('direct_comparisons_path: %s' % direct_comparisons_path)
precision_at_k_plot_path = '/'.join(
[dest_dir, precision_at_k_plot_name])
log.debug('precision_at_k_plot_path: %s' % precision_at_k_plot_path)
precision_recall_plot_path = '/'.join(
[dest_dir, precision_recall_plot_name])
log.debug('precision_recall_plot_path: %s' %
precision_recall_plot_path)
roc_plot_path = '/'.join([dest_dir, roc_plot_name])
log.debug('roc_plot_path: %s' % roc_plot_path)
report_path = '/'.join([dest_dir, report_name])
log.debug('report_path: %s' % report_path)
# Build plot titles.
roc_plot_title = 'ROC (%s)' % pipeline_prefix
precision_recall_plot_title = 'Precision-Recall (%s)' % pipeline_prefix
precision_at_k_plot_title = 'Precision @K (%s)' % pipeline_prefix
# Write output.
analyzer.output_direct_comparisons(direct_comparisons_path)
analyzer.plot_roc_curve(roc_plot_title, roc_plot_path)
analyzer.plot_precision_recall_curve(precision_recall_plot_title,
precision_recall_plot_path)
analyzer.plot_precision_at_k_curve(precision_at_k_plot_title,
precision_at_k_plot_path)
analyzer.write_report(report_path, ci=0.95)
def read_lab_meta_report(lab_panel):
fm_io = FeatureMatrixIO()
data_dir = LabNormalityReport.fetch_data_dir_path()
meta_report_path = data_dir + '/%s/%s-normality-prediction-report.tab' % (lab_panel, lab_panel)
if os.path.exists(meta_report_path):
meta_report = fm_io.read_file_to_data_frame(meta_report_path)
return meta_report
else:
# IF meta_report does not exist, fetch the data on class counts.
algorithm = SupervisedClassifier.REGRESS_AND_ROUND
report_path = data_dir + '/%s/%s/%s-normality-prediction-report.tab' % (lab_panel, algorithm, lab_panel)
algorithm_report = fm_io.read_file_to_data_frame(report_path)
return algorithm_report
def write_matrix(self, dest_path, header=None):
log.info('Writing matrix file...')
fm_io = FeatureMatrixIO()
# Get old matrix file.
source_path = self._factory.getMatrixFileName()
# Write to new matrix filee.
matrix_file = open(dest_path, 'w')
# for line in header:
# matrix_file.write('# %s\n' % line)
for line in open(source_path, 'r'):
if line[0] != '#':
matrix_file.write(line)
# Delete old matrix file.
os.remove(source_path)
def test_SupervisedLearner():
from medinfo.ml.SupervisedLearner import SupervisedLearner
import inspect
from medinfo.dataconversion.FeatureMatrixIO import FeatureMatrixIO
class LabNormalityLearner(SupervisedLearner):
def __init__(self, input_matrix, ylabel):
self.working_folderpath = '/'.join(inspect.getfile(inspect.currentframe()).split('/')[:-1])
self.input_matrix = input_matrix
self.ylabel = ylabel
pass
fm_io = FeatureMatrixIO()
processed_matrix = fm_io.read_file_to_data_frame('data-testingSupervisedLearner-panel-10000-episodes/LABA1C/LABA1C-normality-train-matrix-processed.tab')
processed_matrix.pop('pat_id')
lnl = LabNormalityLearner(processed_matrix, 'all_components_normal')
lnl.run()
def test_read_file_to_data_frame(self):
# Initialize FeatureMatrixIO.
fm_io = FeatureMatrixIO()
# Build paths for test files.
app_dir = os.path.dirname(os.path.abspath(inspect.getfile(inspect.currentframe())))
no_header_file_name = 'test-matrix-no-header.tab'
with_header_file_name = 'test-matrix-with-header.tab'
no_header_file_path = os.path.join(app_dir, no_header_file_name)
with_header_file_path = os.path.join(app_dir, with_header_file_name)
# Read files into data frames.
matrix_stripped_header = fm_io.read_file_to_data_frame(with_header_file_path)
matrix_no_header = fm_io.read_file_to_data_frame(no_header_file_path)
# Verify that FeatureMatrixIO correctly stripped the header.
expected_matrix = MANUAL_TEST_CASE['matrix_no_header']
assert_frame_equal(expected_matrix, matrix_stripped_header)
assert_frame_equal(expected_matrix, matrix_no_header)
def test_strip_header(self):
# Initialize FeatureMatrixIO.
fm_io = FeatureMatrixIO()
# Build paths for test files.
app_dir = os.path.dirname(os.path.abspath(inspect.getfile(inspect.currentframe())))
no_header_file_name = 'test-matrix-no-header.tab'
with_header_file_name = 'test-matrix-with-header.tab'
no_header_file_path = os.path.join(app_dir, no_header_file_name)
with_header_file_path = os.path.join(app_dir, with_header_file_name)
# Strip header.
matrix_with_header = fm_io.read_file_to_data_frame(with_header_file_path)
self._stripped_header_file_path = fm_io.strip_header(with_header_file_path)
# Validate matrix data.
expected_matrix = MANUAL_TEST_CASE['matrix_no_header']
actual_matrix = fm_io.read_file_to_data_frame(self._stripped_header_file_path, \
datetime_col_index=1)
assert_frame_equal(expected_matrix, actual_matrix)
def jitter_processed_matrix(lab, pat_num_limit=100):
data_file = "%s-normality-matrix-10000-episodes-processed.tab" % lab
data_path = os.path.join(data_folder, lab, data_file)
fm_io = FeatureMatrixIO()
df = fm_io.read_file_to_data_frame(data_path)
'''
Reset the pat ids
'''
pat_ids = sorted(set(df['pat_id'].values.tolist()))
pat_ids = pat_ids[:pat_num_limit]
pat2pat = {}
for i, pat_id in enumerate(pat_ids):
pat2pat[pat_id] = i
df['pat_id'] = df['pat_id'].apply(lambda x: pat2pat[x]
if x in pat2pat else None)
df = df.dropna()
print np.array_repr(df.values)
print df.columns
def test(test_suite=[]):
import LabNormalityLearner_Config as Config
from medinfo.dataconversion.FeatureMatrixIO import FeatureMatrixIO
fm_io = FeatureMatrixIO()
raw_matrix = fm_io.read_file_to_data_frame(
'LabNormalityLearner_TestData/LABA1C-normality-matrix-raw.tab')
if 'remove' in test_suite:
remover = LNL.FeatureRemover(Config.features_to_remove)
processed_matrix_removed = remover.transform(raw_matrix)
assert raw_matrix.shape[0] < processed_matrix_removed.shape[0]
assert raw_matrix.shape[1] == processed_matrix_removed.shape[1]
if 'impute' in test_suite:
features_to_impute = [
'TBIL.-14_0.max', 'TBIL.-14_0.median', 'TBIL.-14_0.mean',
'TBIL.-14_0.std'
]
#('min', 'max', 'median', 'mean', 'std', 'first', 'last', 'diff', 'slope', 'proximate')
imputation_dict = {}
for feature in features_to_impute:
imputation_dict[feature] = 0
imputer = LNL.FeatureImputer(imputation_dict=imputation_dict)
columns_to_look = [
'pat_id', 'TBIL.-14_0.max', 'TBIL.-14_0.median', 'TBIL.-14_0.mean',
'TBIL.-14_0.std'
]
print 'raw_matrix[columns_to_look].head():', raw_matrix[
columns_to_look].head()
processed_matrix_imputed = imputer.fit_transform(raw_matrix)
print 'processed_matrix_imputed[columns_to_look].head():', processed_matrix_imputed[
columns_to_look].head()
assert processed_matrix_imputed[columns_to_look].isna().any().any(
) == False
assert (raw_matrix['order_proc_id'].values ==
processed_matrix_imputed['order_proc_id'].values).all()
def load_processed_matrix(lab, dataset_folderpath, type='full'):
data_lab_folderpath = os.path.join(dataset_folderpath, lab)
if type == 'train':
matrix_filepath = os.path.join(data_lab_folderpath,
processed_matrix_train_template % lab)
elif type == 'evalu':
matrix_filepath = os.path.join(data_lab_folderpath,
processed_matrix_evalu_template % lab)
else:
matrix_filepath = os.path.join(data_lab_folderpath,
processed_matrix_template % lab)
fm_io = FeatureMatrixIO()
# TODO: check if raw matrix exists
if os.path.exists(matrix_filepath):
matrix = fm_io.read_file_to_data_frame(matrix_filepath)
else:
matrix = fm_io.read_file_to_data_frame(
matrix_filepath.replace('-test', '-evalu'))
return matrix
def _build_processed_feature_matrix(self):
# Define parameters for processing steps.
params = {}
raw_matrix_path = self._build_raw_matrix_path()
processed_matrix_path = self._build_processed_matrix_path(raw_matrix_path)
log.debug('params: %s' % params)
prev_measurement_feature = self._change_params['feature_old']
features_to_add = {'change': [self._change_params]}
features_to_filter_on = [{'feature': prev_measurement_feature,
'value':np.nan}]
imputation_strategies = {
}
features_to_remove = [
'pat_id', 'order_time', 'order_proc_id', 'ord_num_value',
'proc_code', 'abnormal_panel', 'all_components_normal',
'num_normal_components', 'Birth.pre',
'Male.preTimeDays', 'Female.preTimeDays',
'RaceWhiteHispanicLatino.preTimeDays',
'RaceWhiteNonHispanicLatino.preTimeDays',
'RaceHispanicLatino.preTimeDays',
'RaceAsian.preTimeDays',
'RaceBlack.preTimeDays',
'RacePacificIslander.preTimeDays',
'RaceNativeAmerican.preTimeDays',
'RaceOther.preTimeDays',
'RaceUnknown.preTimeDays',
'Death.post',
'Death.postTimeDays',
'num_components'
]
features_to_keep = [
# Keep the # of times it's been ordered in past, even if low info.
'%s.pre' % self._var
]
outcome_label = 'unchanged_yn'
selection_problem = FeatureSelector.CLASSIFICATION
selection_algorithm = FeatureSelector.RECURSIVE_ELIMINATION
percent_features_to_select = 0.05
matrix_class = LabChangeMatrix
pipeline_file_path = inspect.getfile(inspect.currentframe())
data_overview = [
# Overview:
'Overview',
# The outcome label is ___.
'The outcome label is %s.' % outcome_label,
# %s is a boolean indicator which summarizes whether the lab test
'%s is a boolean indicator which summarizes whether the lab test ' % outcome_label,
# result is unchanged compared to the previous measurement.
'result is unchanged compared to the previous measurement.',
# Each row represents a unique lab panel order.
'Each row represents a unique lab panel order.',
# Each row contains fields summarizing the patient's demographics,
"Each row contains fields summarizing the patient's demographics",
# inpatient admit date, prior vitals, and prior lab results.
'inpatient admit date, prior vitals, and prior lab results.',
# Most cells in matrix represent a count statistic for an event's
"Most cells in matrix represent a count statistic for an event's",
# occurrence or a difference between an event's time and index_time.
"occurrence or a difference between an event's time and index_time.",
# Lab panel orders were only included if a previous measurement of
"Lab panel orders were only included if a previous measurement of",
# the same lab panel has been recorded
"the same lab panel has been recorded."
]
# Bundle parameters into single object
params['raw_matrix_path'] = raw_matrix_path
params['processed_matrix_path'] = processed_matrix_path
params['features_to_add'] = features_to_add
params['features_to_keep'] = features_to_keep
params['features_to_filter_on'] = features_to_filter_on
params['imputation_strategies'] = imputation_strategies
params['features_to_remove'] = features_to_remove
params['outcome_label'] = outcome_label
params['selection_problem'] = selection_problem
params['selection_algorithm'] = selection_algorithm
params['percent_features_to_select'] = percent_features_to_select
params['matrix_class'] = matrix_class
params['pipeline_file_path'] = pipeline_file_path
params['data_overview'] = data_overview
# defer to SupervisedLearningPipeline logic by SX
fm_io = FeatureMatrixIO()
log.debug('params: %s' % params)
# If processed matrix exists, and the client has not requested to flush
# the cache, just use the matrix that already exists and return.
processed_matrix_path = params['processed_matrix_path']
if os.path.exists(processed_matrix_path) and not self._flush_cache:
# Assume feature selection already happened, but we still need
# to split the data into training and test data.
processed_matrix = fm_io.read_file_to_data_frame(processed_matrix_path)
'''
Make sure the order of rows is consistent before splitting
'''
processed_matrix.sort_index(inplace=True)
self._train_test_split(processed_matrix, params['outcome_label']) #TODO sxu: when reloading, no pat_id
else:
# Read raw matrix.
raw_matrix = fm_io.read_file_to_data_frame(params['raw_matrix_path'])
# Initialize FMT.
# Add outcome label
raw_fmt = FeatureMatrixTransform()
raw_fmt.set_input_matrix(raw_matrix)
self._filter_on_features(raw_fmt, params['features_to_filter_on'])
self._add_features(raw_fmt, params['features_to_add'])
raw_matrix = raw_fmt.fetch_matrix()
# Divide processed_matrix into training and test data.
# This must happen before feature selection so that we don't
# accidentally learn information from the test data.
# TODO: work on this...
self._train_test_split(raw_matrix, params['outcome_label'])
fmt = FeatureMatrixTransform()
train_df = self._X_train.join(self._y_train)
fmt.set_input_matrix(train_df)
# Remove features.
self._remove_features(fmt, params['features_to_remove'])
# Filter on features
if 'features_to_filter_on' in params:
self._filter_on_features(fmt, params['features_to_filter_on'])
# HACK: When read_csv encounters duplicate columns, it deduplicates
# them by appending '.1, ..., .N' to the column names.
# In future versions of pandas, simply pass mangle_dupe_cols=True
# to read_csv, but not ready as of pandas 0.22.0.
for feature in raw_matrix.columns.values:
if feature[-2:] == ".1":
fmt.remove_feature(feature)
self._removed_features.append(feature)
# Impute data.
self._impute_data(fmt, train_df, params['imputation_strategies'])
# In case any all-null features were created in preprocessing,
# drop them now so feature selection will work
fmt.drop_null_features()
# Build interim matrix.
train_df = fmt.fetch_matrix()
self._y_train = pd.DataFrame(train_df.pop(params['outcome_label']))
self._X_train = train_df
'''
Select X_test columns according to processed X_train
'''
self._X_test = self._X_test[self._X_train.columns]
'''
Impute data according to the same strategy when training
'''
for feat in self._X_test.columns:
self._X_test[feat] = self._X_test[feat].fillna(self.feat2imputed_dict[feat])
self._select_features(params['selection_problem'],
params['percent_features_to_select'],
params['selection_algorithm'],
params['features_to_keep'])
train = self._y_train.join(self._X_train)
test = self._y_test.join(self._X_test)
processed_matrix = train.append(test)
'''
Need to recover the order of rows before writing into disk
'''
processed_matrix.sort_index(inplace=True)
# Write output to new matrix file.
header = self._build_processed_matrix_header(params)
fm_io.write_data_frame_to_file(processed_matrix, \
processed_matrix_path, header)
def get_raw_matrix(self):
return FeatureMatrixIO().read_file_to_data_frame(
self.raw_matrix_filepath)
def apply_src_to_dst(lab,
lab_type,
src_datasource,
dst_datasource,
src_dataset_folderpath,
dst_dataset_folderpath,
output_folderpath,
use_cached_results=True):
'''
What: Use case that transfers model from one institute (src) to another (dst)
Why:
TODO: automatically recognize lab_type
How:
Load inputs:
(1) Read dst raw matrix from dst dataset_folder/lab
(2) Read src imputation template (only includes final features) from src dataset_folder/lab
(3) Read src trained model from src dataset_folder/lab
Process:
(1)
Create a dst imputation template
For each src feature in the src imputation template
map the feature to dst column
if not exists such dst column:
create a new column in the dst raw matrix, fill with the src imputing value
(2)
Feed into process_matrix, pop pat_id and info features, split Xy
(3)
Feed X into classifier, get y_pred, write to direct_comparisons with y_true
Args:
lab:
lab_type:
dataset_folder:
Returns:
'''
print "Transfering %s %s from %s to %s..." % (
lab_type, lab, src_datasource, dst_datasource)
from medinfo.dataconversion.FeatureMatrixIO import FeatureMatrixIO
import pickle
import os
if not os.path.exists(output_folderpath):
os.mkdir(output_folderpath)
output_filepath = os.path.join(output_folderpath, 'direct_comparisons.csv')
if use_cached_results and os.path.exists(output_filepath):
# print output_filepath + ' exists!'
return
# TODO: UMich?
if lab_type == 'panel':
# from scripts.LabTestAnalysis.machine_learning.ml_utils import map_panel_from_Stanford_to_UCSF as map_lab
ylabel = 'all_components_normal'
else:
# from scripts.LabTestAnalysis.machine_learning.ml_utils import map_component_from_Stanford_to_UCSF as map_lab
ylabel = 'component_normal'
lab_src = ml_utils.map_lab(lab=lab,
data_source=src_datasource,
lab_type=lab_type)
lab_dst = ml_utils.map_lab(lab=lab,
data_source=dst_datasource,
lab_type=lab_type)
'''
Helper function
'''
fm_io = FeatureMatrixIO()
'''
Data folder
'''
# lab_folder = os.path.join(dataset_folder, lab)
'''
Load raw data from dst
'''
df_raw_dst = SL.load_raw_matrix(lab=lab_dst,
dataset_folderpath=dst_dataset_folderpath)
# imputations_stanford = SL.load_imputation_template(lab=lab_mapped, dataset_folderpath=dst_dataset_folderpath, lab_type=lab_type)
# df_ucsf_raw = fm_io.read_file_to_data_frame(lab_folder + '/' + "%s-normality-matrix-raw.tab"%map_lab_Stanford_to_UCSF[lab])
raw_columns_dst = df_raw_dst.columns.values.tolist()
'''
From test processed, get the patient evalu set
'''
df_processed_evalu_dst = SL.load_processed_matrix(
lab_dst, dst_dataset_folderpath, type='evalu'
) #fm_io.read_file_to_data_frame(lab_folder + '/' + "%s-normality-test-matrix-processed.tab" % map_lab_Stanford_to_UCSF[lab])
patIds_evalu_dst = ml_utils.get_patIds(
df_processed_evalu_dst
) #set(df_ucsf_processed_evalu['pat_id'].values.tolist())
df_raw_evalu_dst = df_raw_dst[df_raw_dst['pat_id'].isin(patIds_evalu_dst)]
assert df_raw_dst.shape[0] > df_raw_evalu_dst.shape[0]
'''
Load imputation template from Stanford
TODO: this is old-versioned template, (1) without column order and (2) a lot of unnecessary columns.
'''
# del impute_dict_old[ylabel] #
'''
Use processed_matrix to select columns
'''
df_processed_src = fm_io.read_file_to_data_frame(
src_dataset_folderpath + '/' + lab_src +
'/%s-normality-matrix-processed.tab' % lab_src)
df_processed_src.pop('pat_id')
df_processed_src.pop(ylabel) # TODO?!
processed_columns_src = df_processed_src.columns.values.tolist()
classifier_src = SL.load_ML_model(
lab=lab_src,
alg='random-forest',
dataset_folderpath=src_dataset_folderpath)
'''
Finding the corresponding UCSF column of each Stanford's processed feature
If this feature exists in UCSF, then good
If not, create dummy feature for UCSF raw matrix!
'''
imputations_filepath = src_dataset_folderpath + '/' + lab_src + '/' + "feat2imputed_dict.pkl"
impute_dict_old = pickle.load(open(imputations_filepath))
if ylabel in impute_dict_old:
del impute_dict_old[ylabel]
impute_dict_new = {}
i = 0
for col_src in processed_columns_src:
col_dst = map_col(col_src, src=src_datasource, dst=dst_datasource)
if col_src in raw_columns_dst:
col_dst = col_src
elif col_dst not in raw_columns_dst:
print "Unknown:", col_src, col_dst
'''
create dummy column for dst
'''
df_raw_evalu_dst[col_dst] = df_raw_evalu_dst['pat_id'].apply(
lambda x: 0)
# if col_dst in impute_dict_new:
# '''
# Different src features map into the same dst feature
# '''
# df_raw_evalu_dst['dummy_%s'%i] = df_raw_evalu_dst['pat_id'].apply(lambda x: 0)
# pass
if col_dst in impute_dict_new:
col_dst = 'dummy_%s' % i
df_raw_evalu_dst[col_dst] = df_raw_evalu_dst['pat_id'].apply(
lambda x: 0)
'''
Use Stanford mean to impute
'''
if col_dst in df_raw_evalu_dst:
# print col_mapped # TODO: XPPT and PPT are the same thing?
pass
# print i, col_selected, col_mapped
'''
40 PCO2A.-14_0.proximate PCO2.-14_0.proximate
41 PCO2V.-14_0.proximate PCO2.-14_0.proximate
'''
impute_dict_new[col_dst] = (i, impute_dict_old[col_src]) #
i += 1
'''
Feature auxillary
'''
features = {'ylabel': ylabel, 'info': ['pat_id']}
df_ucsf_processed_evalu, _ = SL.process_matrix(
df_raw_evalu_dst, features, impute_template=impute_dict_new)
# print "Finished processing!"
# df_ucsf_processed.pop('all_components_normal')
df_ucsf_processed_evalu.pop('pat_id')
'''
Load model
'''
# print "Finished Loading!"
# print classifier.description() # TODO: why is this step so slow?!
# print classifier.predict_probability(df_ucsf_processed)
# print classifier_src._params_random_forest()['decision_features']
X_evalu, y_evalu = SL.split_Xy(data_matrix=df_ucsf_processed_evalu,
outcome_label=ylabel)
SL.predict(X_evalu,
y_evalu,
classifier_src,
output_filepath=output_filepath)
labs['predictable_CV'] = (labs['percent_predictably_positive'].astype('float') * labs['annual_median_charge_volume ($)'].astype('float') / 1000).map('${:,.0f}'.format)
labs['predictable_CV[-0.95]'] = (labs['percent_predictably_positive_0.95_lower_ci'].astype('float') * labs['annual_median_charge_volume ($)'].astype('float') / 1000).map('${:,.0f}'.format)
labs['predictable_CV[+0.95]'] = (labs['percent_predictably_positive_0.95_upper_ci'].astype('float') * labs['annual_median_charge_volume ($)'].astype('float') / 1000).map('${:,.0f}'.format)
summary = DataFrame()
summary['lab'] = labs['label']
summary['charge'] = labs['median_charge'].astype('float').map('${:,.0f}'.format)
summary['volume'] = labs['volume'].floordiv(6).astype('float').map('{:,.0f}'.format)
summary['normal rate'] = labs['normality']
summary['[email protected]'] = labs['[email protected]'] + ' [' + \
labs['[email protected][-0.95]'] + ', ' + \
labs['[email protected][+0.95]'] + ']'
summary['predictable CV ($1,000s)'] = labs['predictable_CV'] + ' [' + \
labs['predictable_CV[-0.95]'] + ', ' + \
labs['predictable_CV[+0.95]'] + ']'
return summary
if __name__ == '__main__':
fm_io = FeatureMatrixIO()
summary_table = LabNormalityReport.build_lab_performance_summary_table()
fm_io.write_data_frame_to_file(summary_table, 'lab-performance-summary.tab')
summary = LabNormalityReport.build_algorithm_performance_summary_table()
fm_io.write_data_frame_to_file(summary, 'algorithm-performance-summary.tab')
LabNormalityReport.plot_predictable_and_expensive_charges()
summary = LabNormalityReport.build_lab_predictability_summary_report()
fm_io.write_data_frame_to_file(summary, 'predictable-labs.tab')
summary = LabNormalityReport.build_lab_predictability_summary_report(all=True)
fm_io.write_data_frame_to_file(summary, 'all-labs.tab')
def apply_Stanford_to_UCSF():
from medinfo.dataconversion.FeatureMatrixIO import FeatureMatrixIO
import pickle
'''
Data folder
'''
dataset_folder = "data-apply-Stanford-to-UCSF-10000-episodes/LABMGN/"
'''
Helper function
'''
fm_io = FeatureMatrixIO()
'''
Load raw data from UCSF
'''
df_ucsf_raw = fm_io.read_file_to_data_frame(
dataset_folder + "Magnesium, Serum - Plasma-normality-matrix-raw.tab")
raw_columns_ucsf = df_ucsf_raw.columns.values.tolist()
'''
Load imputation template from Stanford
TODO: this is old-versioned template, (1) without column order and (2) a lot of unnecessary columns.
'''
impute_dict_old = pickle.load(
open(dataset_folder + "feat2imputed_dict.pkl"))
del impute_dict_old['all_components_normal'] # TODO?!
'''
Use processed_matrix to select columns
'''
df_stanford_processed = fm_io.read_file_to_data_frame(
dataset_folder + 'LABMGN-normality-matrix-processed.tab')
df_stanford_processed.pop('pat_id')
df_stanford_processed.pop('all_components_normal') # TODO?!
processed_columns_stanford = df_stanford_processed.columns.values.tolist()
'''
Finding the corresponding UCSF column of each Stanford's processed feature
If this feature exists in UCSF, then good
If not, create dummy feature for UCSF raw matrix!
'''
impute_dict_new = {}
for i, col_selected in enumerate(processed_columns_stanford):
col_mapped = map_col_Stanford_to_UCSF(col_selected)
if col_mapped in raw_columns_ucsf:
impute_dict_new[col_mapped] = (i, impute_dict_old[col_selected])
else:
'''
Features Unknown to Stanford
'''
df_ucsf_raw[col_mapped] = df_ucsf_raw['pat_id'].apply(lambda x: 0)
impute_dict_new[col_mapped] = (i, 0
) # TODO: better strategy later?
print "Unknown:", col_mapped
'''
Feature auxillary
'''
features = {'ylabel': 'all_components_normal', 'info': ['pat_id']}
df_ucsf_processed, _ = SL.process_matrix(df_ucsf_raw,
features,
impute_template=impute_dict_new)
print "Finished processing!"
# df_ucsf_processed.pop('all_components_normal')
df_ucsf_processed.pop('pat_id')
'''
Load model
'''
classifier = joblib.load(dataset_folder +
'LABMGN-normality-random-forest-model.pkl')
print "Finished Loading!"
# print classifier.description() # TODO: why is this step so slow?!
# print classifier.predict_probability(df_ucsf_processed)
print classifier._params_random_forest()['decision_features']
X_evalu, y_evalu = SL.split_Xy(data_matrix=df_ucsf_processed,
outcome_label='all_components_normal')
SL.predict(X_evalu,
y_evalu,
classifier,
output_filepath=dataset_folder + 'direct_comparisons.csv')
def __init__(self):
self._fm_io = FeatureMatrixIO()
def get_train_and_evalu_raw_matrices(lab,
data_lab_folderpath,
random_state,
train_size=0.75,
columnToSplitOn='pat_id'):
'''
If train and eval exist, direct get from disk
Avoided saving as 2 raw matrices, too much space!
elif raw matrix exists, get from dist and split
else, get from SQL
Args:
raw_matrix_filepath:
random_state:
use_cached:
Returns:
'''
raw_matrix_filepath = os.path.join(data_lab_folderpath,
raw_matrix_template % lab)
fm_io = FeatureMatrixIO()
# TODO: check if raw matrix exists
raw_matrix = fm_io.read_file_to_data_frame(raw_matrix_filepath)
pat_split_filepath = os.path.join(data_lab_folderpath, pat_split_filename)
'''
Old pipeline style
'''
if os.path.exists(pat_split_filepath):
pat_split_df = pd.read_csv(pat_split_filepath)
pat_ids_train = pat_split_df[pat_split_df['in_train'] ==
1]['pat_id'].values.tolist()
raw_matrix_train = raw_matrix[raw_matrix['pat_id'].isin(pat_ids_train)]
pat_ids_evalu = pat_split_df[pat_split_df['in_train'] ==
0]['pat_id'].values.tolist()
raw_matrix_evalu = raw_matrix[raw_matrix['pat_id'].isin(pat_ids_evalu)]
else:
raw_matrix_train, raw_matrix_evalu = split_rows(
raw_matrix,
train_size=train_size,
columnToSplitOn=columnToSplitOn,
random_state=random_state)
pat_ids_train = set(raw_matrix_train['pat_id'].values.tolist())
pat_split_df = raw_matrix[['pat_id']].copy()
pat_split_df['in_train'] = pat_split_df['pat_id'].apply(
lambda x: 1 if x in pat_ids_train else 0)
# pat_split_df.to_csv(pat_split_filepath, index=False)
assert set(raw_matrix_train['pat_id'].values.tolist()) & set(
raw_matrix_evalu['pat_id'].values.tolist()) == set([])
assert raw_matrix_train.shape[0] + raw_matrix_evalu.shape[
0] == raw_matrix.shape[0]
return raw_matrix_train, raw_matrix_evalu
def _build_processed_feature_matrix(self, params):
# params is a dict defining the details of how the raw feature matrix
# should be transformed into the processed matrix. Given the sequence
# of steps will be identical across all pipelines, sbala decided to
# pack all the variability into this dict. It's not ideal because the
# dict has 10+ values, but that seems better than forcing all pipelines
# to reproduce the logic of the processing steps.
# Principle: Minimize overridden function calls.
# params['features_to_add'] = features_to_add
# params['features_to_filter_on'] (optional) = features_to_filter_on
# params['imputation_strategies'] = imputation_strategies
# params['features_to_remove'] = features_to_remove
# params['outcome_label'] = outcome_label
# params['selection_problem'] = selection_problem
# params['selection_algorithm'] = selection_algorithm
# params['percent_features_to_select'] = percent_features_to_select
# params['matrix_class'] = matrix_class
# params['pipeline_file_path'] = pipeline_file_path
# TODO(sbala): Determine which fields should have defaults.
fm_io = FeatureMatrixIO()
log.debug('params: %s' % params)
# If processed matrix exists, and the client has not requested to flush
# the cache, just use the matrix that already exists and return.
processed_matrix_path = params['processed_matrix_path']
if os.path.exists(processed_matrix_path) and not self._flush_cache:
# Assume feature selection already happened, but we still need
# to split the data into training and test data.
processed_matrix = fm_io.read_file_to_data_frame(
processed_matrix_path)
self._train_test_split(processed_matrix, params['outcome_label'])
else:
# Read raw matrix.
raw_matrix = fm_io.read_file_to_data_frame(
params['raw_matrix_path'])
# Initialize FMT.
fmt = FeatureMatrixTransform()
fmt.set_input_matrix(raw_matrix)
# Add features.
self._add_features(fmt, params['features_to_add'])
# Remove features.
self._remove_features(fmt, params['features_to_remove'])
# Filter on features
if 'features_to_filter_on' in params:
self._filter_on_features(fmt, params['features_to_filter_on'])
# HACK: When read_csv encounters duplicate columns, it deduplicates
# them by appending '.1, ..., .N' to the column names.
# In future versions of pandas, simply pass mangle_dupe_cols=True
# to read_csv, but not ready as of pandas 0.22.0.
for feature in raw_matrix.columns.values:
if feature[-2:] == ".1":
fmt.remove_feature(feature)
self._removed_features.append(feature)
# Impute data.
self._impute_data(fmt, raw_matrix, params['imputation_strategies'])
# In case any all-null features were created in preprocessing,
# drop them now so feature selection will work
fmt.drop_null_features()
# Build interim matrix.
processed_matrix = fmt.fetch_matrix()
# Divide processed_matrix into training and test data.
# This must happen before feature selection so that we don't
# accidentally learn information from the test data.
self._train_test_split(processed_matrix, params['outcome_label'])
self._select_features(params['selection_problem'],
params['percent_features_to_select'],
params['selection_algorithm'],
params['features_to_keep'])
train = self._y_train.join(self._X_train)
test = self._y_test.join(self._X_test)
processed_matrix = train.append(test)
# Write output to new matrix file.
header = self._build_processed_matrix_header(params)
fm_io.write_data_frame_to_file(processed_matrix, \
processed_matrix_path, header)
def _train_and_analyze_predictors(self):
log.info('Training and analyzing predictors...')
problem = SupervisedLearningPipeline.CLASSIFICATION
meta_report = None
fm_io = FeatureMatrixIO()
# Build paths for output.
pipeline_file_name = inspect.getfile(inspect.currentframe())
data_dir = self._fetch_data_dir_path(pipeline_file_name)
# Test BifurcatedSupervisedClassifier and SupervisedClassifier.
algorithms_to_test = list()
algorithms_to_test.extend(SupervisedClassifier.SUPPORTED_ALGORITHMS)
for algorithm in SupervisedClassifier.SUPPORTED_ALGORITHMS:
pass # TODO:(raikens) something in the BifurcatedSupervisedClassifier pipeline is crashing
#algorithms_to_test.append('bifurcated-%s' % algorithm)
log.debug('algorithms_to_test: %s' % algorithms_to_test)
# Train and analyse algorithms.
for algorithm in algorithms_to_test:
log.info('Training and analyzing %s...' % algorithm)
# If report_dir does not exist, make it.
report_dir = '/'.join([data_dir, algorithm])
if not os.path.exists(report_dir):
os.makedirs(report_dir)
log.debug('report_dir: %s' % report_dir)
# Define hyperparams.
hyperparams = {}
hyperparams['algorithm'] = algorithm
hyperparams[
'hyperparam_strategy'] = SupervisedClassifier.EXHAUSTIVE_SEARCH
hyperparams['max_iter'] = 1024
# If bifurcated algorithm, define bifurcator.
if 'bifurcated' in algorithm:
# bifrucator = LAB.pre == 0
hyperparams['bifurcator'] = '%s.pre' % self._var
hyperparams[
'bifurcation_strategy'] = BifurcatedSupervisedClassifier.EQUAL
hyperparams['bifurcation_value'] = 0
hyperparams['bifurcated'] = True
# Train classifier.
predictor_path = self._build_model_dump_path(algorithm)
if os.path.exists(
predictor_path) and 'bifurcated' not in algorithm:
log.debug('Loading model from disk...')
# TODO(sbala): Fix loblib.load so that it works for bifurcated
# supervised classifiers.
self._predictor = joblib.load(predictor_path)
self._features = self._X_train.columns
status = SupervisedClassifier.TRAINED
else:
status = SupervisedLearningPipeline._train_predictor(
self, problem, [0, 1], hyperparams)
# If failed to train, write an error report.
y_train_counts = self._y_train[
self._y_train.columns[0]].value_counts()
y_test_counts = self._y_test[
self._y_test.columns[0]].value_counts()
if status == SupervisedClassifier.INSUFFICIENT_SAMPLES:
# Skip all analysis and reporting.
# This will be true for all algorithms, so just return.
# Build error report.
algorithm_report = DataFrame(
{
'lab_panel': [self._var],
'algorithm': [algorithm],
'error': [status],
'y_train.value_counts()': [y_train_counts.to_dict()],
'y_test.value_counts()': [y_test_counts.to_dict()]
},
columns=[
'lab_panel', 'algorithm', 'error',
'y_train.value_counts()', 'y_test.value_counts()'
])
header = [
'LabChangePredictionPipeline("%s", %d)' %
(self._var, self._num_rows)
]
# Write error report.
fm_io.write_data_frame_to_file(algorithm_report, \
'/'.join([report_dir, '%s-change-prediction-report.tab' % (self._var)]), \
header)
# If successfully trained, append to a meta report.
elif status == SupervisedClassifier.TRAINED:
pipeline_prefix = '%s-change-prediction-%s' % (self._var,
algorithm)
SupervisedLearningPipeline._analyze_predictor(
self, report_dir, pipeline_prefix)
if meta_report is None:
meta_report = fm_io.read_file_to_data_frame('/'.join(
[report_dir,
'%s-report.tab' % pipeline_prefix]))
else:
algorithm_report = fm_io.read_file_to_data_frame('/'.join(
[report_dir,
'%s-report.tab' % pipeline_prefix]))
log.debug('algorithm_report: %s' % algorithm_report)
meta_report = meta_report.append(algorithm_report)
# Write predictor to disk.
predictor = SupervisedLearningPipeline.predictor(self)
predictor_path = self._build_model_dump_path(algorithm)
joblib.dump(predictor, predictor_path)
# After building per-algorithm reports, write to meta report.
# Note that if there were insufficient samples to build any of the
# algorithms, then meta_report will still be None.
if meta_report is not None:
header = [
'LabChangePredictionPipeline("%s", %d)' %
(self._var, self._num_rows)
]
fm_io.write_data_frame_to_file(meta_report, \
'/'.join([data_dir, '%s-change-prediction-report.tab' % self._var]), header)
def _build_processed_feature_matrix(self, params):
# params is a dict defining the details of how the raw feature matrix
# should be transformed into the processed matrix. Given the sequence
# of steps will be identical across all pipelines, sbala decided to
# pack all the variability into this dict. It's not ideal because the
# dict has 10+ values, but that seems better than forcing all pipelines
# to reproduce the logic of the processing steps.
# Principle: Minimize overridden function calls.
# params['features_to_add'] = features_to_add
# params['features_to_filter_on'] (optional) = features_to_filter_on
# params['imputation_strategies'] = imputation_strategies
# params['features_to_remove'] = features_to_remove
# params['outcome_label'] = outcome_label
# params['selection_problem'] = selection_problem
# params['selection_algorithm'] = selection_algorithm
# params['percent_features_to_select'] = percent_features_to_select
# params['matrix_class'] = matrix_class
# params['pipeline_file_path'] = pipeline_file_path
# TODO(sbala): Determine which fields should have defaults.
fm_io = FeatureMatrixIO()
log.debug('params: %s' % params)
# If processed matrix exists, and the client has not requested to flush
# the cache, just use the matrix that already exists and return.
processed_matrix_path = params['processed_matrix_path']
if os.path.exists(processed_matrix_path) and not self._flush_cache:
# Assume feature selection already happened, but we still need
# to split the data into training and test data.
processed_matrix = fm_io.read_file_to_data_frame(
processed_matrix_path)
# processed_matrix['pat_id'] = processed_matrix['pat_id'].apply(lambda x: str(x))
self._train_test_split(processed_matrix, params['outcome_label'])
'''
Pandas dataframe may automatically convert bigint to float (and round the last
few digits), which may damage the uniqueness of pat_ids.
'''
# processed_matrix['pat_id'] = processed_matrix['pat_id'].apply(lambda x: str(x))
else:
# Read raw matrix.
raw_matrix = fm_io.read_file_to_data_frame(
params['raw_matrix_path'])
# raw_matrix['pat_id'] = raw_matrix['pat_id'].apply(lambda x: str(x))
# Initialize FMT.
# Divide processed_matrix into training and test data.
# This must happen before feature selection so that we don't
# accidentally learn information from the test data.
patIds_df = raw_matrix['pat_id'].copy()
self._train_test_split(raw_matrix, params['outcome_label'])
# ##
# folder_path = '/'.join(params['raw_matrix_path'].split('/')[:-1])
# self._X_train.join(self._y_train).to_csv(folder_path + '/' + 'train_raw.csv', index=False)
# self._X_test.join(self._y_test).to_csv(folder_path + '/' + 'test_raw.csv', index=False)
#
# '''
# Mini-test that there are no overlapping patients
# '''
# assert bool(set(self._X_train['pat_id'].values) & set(self._X_test['pat_id'].values)) == False
# ##
fmt = FeatureMatrixTransform()
train_df = self._X_train.join(self._y_train)
fmt.set_input_matrix(train_df)
# Add features.
self._add_features(fmt, params['features_to_add'])
# Filter on features
if 'features_to_filter_on' in params:
self._filter_on_features(fmt, params['features_to_filter_on'])
# HACK: When read_csv encounters duplicate columns, it deduplicates
# them by appending '.1, ..., .N' to the column names.
# In future versions of pandas, simply pass mangle_dupe_cols=True
# to read_csv, but not ready as of pandas 0.22.0.
for feature in raw_matrix.columns.values:
if feature[-2:] == ".1":
fmt.remove_feature(feature)
self._removed_features.append(feature)
# Impute data.
if params['imputation_strategies'] == {'sxu_new_imputation'}:
train_df = fmt.fetch_matrix()
means = {}
for column in train_df.columns.values.tolist():
# column_tail = column.split('.')[-1].strip()
if train_df[column].dtype == 'float64':
means[column] = train_df[column].mean()
train_df = fmt.do_impute_sx(train_df, means)
fmt.set_input_matrix(train_df)
self._X_test = fmt.do_impute_sx(self._X_test, means)
self._remove_features(fmt, params['features_to_remove'])
else:
self._remove_features(fmt, params['features_to_remove'])
self._impute_data(fmt, train_df,
params['imputation_strategies'])
# Remove features.
'''
Moved here, since still need pat_id for imputation!
'''
# self._remove_features(fmt, params['features_to_remove'])
# In case any all-null features were created in preprocessing,
# drop them now so feature selection will work
fmt.drop_null_features()
# Build interim matrix.
train_df = fmt.fetch_matrix()
self._y_train = pd.DataFrame(train_df.pop(params['outcome_label']))
self._X_train = train_df
'''
Select X_test columns according to processed X_train
'''
self._X_test = self._X_test[self._X_train.columns]
if not params['imputation_strategies'] == {'sxu_new_imputation'}:
for feat in self._X_test.columns:
self._X_test[feat] = self._X_test[feat].fillna(
self.feat2imputed_dict[feat])
self._select_features(params['selection_problem'],
params['percent_features_to_select'],
params['selection_algorithm'],
params['features_to_keep'])
'''
The join is based on index by default.
Will remove 'pat_id' (TODO sxu: more general in the future) later in train().
'''
self._X_train = self._X_train.join(patIds_df, how='left')
self._X_test = self._X_test.join(patIds_df, how='left')
# print set(self._X_train['pat_id'].values.tolist()) & set(self._X_test['pat_id'].values.tolist())
train = self._y_train.join(self._X_train)
test = self._y_test.join(self._X_test)
processed_trainMatrix_path = processed_matrix_path.replace(
"matrix", "train-matrix")
train.to_csv(processed_trainMatrix_path, sep='\t', index=False)
processed_testMatrix_path = processed_matrix_path.replace(
"matrix", "test-matrix")
test.to_csv(processed_testMatrix_path, sep='\t', index=False)
processed_matrix = train.append(test)
'''
Recover the order of rows before writing into disk,
where the index info will be missing.
'''
processed_matrix.sort_index(inplace=True)
# Write output to new matrix file.
header = self._build_processed_matrix_header(params)
fm_io.write_data_frame_to_file(processed_matrix, \
processed_matrix_path, header)
'''
Pop out pat_id from the feature matrices.
Also check whether there is pat_id leakage.
'''
self._patIds_train = self._X_train.pop('pat_id').values.tolist()
self._patIds_test = self._X_test.pop('pat_id').values.tolist()
assert not (set(self._patIds_train) & set(self._patIds_test))
