def kfold_prediction(self, k=10):
# generate indices for kfold cross validation
self.num_pred = 0 # number of predictions
prediction = pd.Series(index=self.y.index) # predicted class
onco_prob = pd.Series(index=self.y.index).fillna(0)
tsg_prob = pd.Series(index=self.y.index).fillna(0)
for i in range(self.total_iter):
# randomize for another round
self.x, self.y = futils.randomize(self.x, self.prng)
futils.check_num_classes(self.y) # warn user if not 3 classes
# set up stratified kfold iterator
kf = KFold(n_splits=k)
k_fold = kf.split(self.y)
# obtain predictions from single round of kfold validation
for train_ix, test_ix in k_fold:
# retreive indices from pandas dataframe using row number
tmp_train_ix = self.x.iloc[train_ix].index
tmp_test_ix = self.x.iloc[test_ix].index
if self.is_weighted_sample:
# figure out sample weights
num_train = len(train_ix)
sample_weight = np.zeros(num_train)
onco_ix = np.nonzero(
self.y.loc[tmp_train_ix] == self.onco_num)[0]
tsg_ix = np.nonzero(
self.y.loc[tmp_train_ix] == self.tsg_num)[0]
other_ix = np.nonzero(
self.y.loc[tmp_train_ix] == self.other_num)[0]
sample_weight[onco_ix] = 1. / len(onco_ix)
sample_weight[tsg_ix] = 1. / len(tsg_ix)
sample_weight[other_ix] = 1. / len(other_ix)
# do training with sample weighting
self.clf.fit(self.x.loc[tmp_train_ix].copy(),
self.y.loc[tmp_train_ix].copy(),
sample_weight=sample_weight)
else:
# do training without weighting
self.clf.fit(self.x.loc[tmp_train_ix].copy(),
self.y.loc[tmp_train_ix].copy())
# predict test data in kfold validation
tmp_prob = self.clf.predict_proba(self.x.loc[tmp_test_ix])
onco_prob.loc[tmp_test_ix] += tmp_prob[:, self.onco_num]
tsg_prob.loc[tmp_test_ix] += tmp_prob[:, self.tsg_num]
self.num_pred += 1
# convert number of trees to fraction of trees
onco_prob /= self.num_pred
tsg_prob /= self.num_pred
other_prob = 1 - (onco_prob + tsg_prob)
# return prediction.astype(int), prob
return onco_prob, tsg_prob, other_prob
def kfold_prediction(self, k=10):
# generate indices for kfold cross validation
self.num_pred = 0 # number of predictions
prediction = pd.Series(index=self.y.index) # predicted class
onco_prob = pd.Series(index=self.y.index).fillna(0)
tsg_prob = pd.Series(index=self.y.index).fillna(0)
for i in range(self.total_iter):
# randomize for another round
self.x, self.y = futils.randomize(self.x, self.prng)
futils.check_num_classes(self.y) # warn user if not 3 classes
# set up stratified kfold iterator
k_fold = cross_validation.StratifiedKFold(self.y,
n_folds=k)
# obtain predictions from single round of kfold validation
for train_ix, test_ix in k_fold:
# retreive indices from pandas dataframe using row number
tmp_train_ix = self.x.iloc[train_ix].index
tmp_test_ix = self.x.iloc[test_ix].index
if self.is_weighted_sample:
# figure out sample weights
num_train = len(train_ix)
sample_weight = np.zeros(num_train)
onco_ix = np.nonzero(self.y.ix[tmp_train_ix]==self.onco_num)[0]
tsg_ix = np.nonzero(self.y.ix[tmp_train_ix]==self.tsg_num)[0]
other_ix = np.nonzero(self.y.ix[tmp_train_ix]==self.other_num)[0]
sample_weight[onco_ix] = 1. / len(onco_ix)
sample_weight[tsg_ix] = 1. / len(tsg_ix)
sample_weight[other_ix] = 1. / len(other_ix)
# do training with sample weighting
self.clf.fit(self.x.ix[tmp_train_ix].copy(),
self.y.ix[tmp_train_ix].copy(),
sample_weight=sample_weight)
else:
# do training without weighting
self.clf.fit(self.x.ix[tmp_train_ix].copy(),
self.y.ix[tmp_train_ix].copy())
# predict test data in kfold validation
tmp_prob = self.clf.predict_proba(self.x.ix[tmp_test_ix])
onco_prob.ix[tmp_test_ix] += tmp_prob[:, self.onco_num]
tsg_prob.ix[tmp_test_ix] += tmp_prob[:, self.tsg_num]
self.num_pred += 1
# convert number of trees to fraction of trees
onco_prob /= self.num_pred
tsg_prob /= self.num_pred
other_prob = 1 - (onco_prob + tsg_prob)
# return prediction.astype(int), prob
return onco_prob, tsg_prob, other_prob
def train_cv(self, k=10):
"""Train classifier on entire data set provided, but done in cross-validation."""
# generate indices for kfold cross validation
self.num_pred = 0 # number of predictions
self.test_fold_df = pd.DataFrame(
{l + 1: 0
for l in range(self.total_iter)}, index=self.x.index)
for i in range(self.total_iter):
# randomize for another round
self.x, self.y = futils.randomize(self.x, self.prng)
futils.check_num_classes(self.y) # warn user if not 3 classes
# set up stratified kfold iterator
kf = KFold(n_splits=k)
k_fold = kf.split(self.y)
# obtain predictions from single round of kfold validation
for nfold, (train_ix, test_ix) in enumerate(k_fold):
# retreive indices from pandas dataframe using row number
tmp_train_ix = self.x.iloc[train_ix].index
# save which genes are in the test fold
tmp_test_ix = self.x.iloc[test_ix].index
self.test_fold_df.loc[tmp_test_ix, i + 1] = nfold + 1
if self.is_weighted_sample:
# figure out sample weights
num_train = len(train_ix)
sample_weight = np.zeros(num_train)
onco_ix = np.nonzero(
self.y.loc[tmp_train_ix] == self.onco_num)[0]
tsg_ix = np.nonzero(
self.y.loc[tmp_train_ix] == self.tsg_num)[0]
other_ix = np.nonzero(
self.y.loc[tmp_train_ix] == self.other_num)[0]
sample_weight[onco_ix] = 1. / len(onco_ix)
sample_weight[tsg_ix] = 1. / len(tsg_ix)
sample_weight[other_ix] = 1. / len(other_ix)
# do training with sample weighting
self.clf.fit(self.x.loc[tmp_train_ix].copy(),
self.y.loc[tmp_train_ix].copy(),
sample_weight=sample_weight)
else:
# do training without weighting
self.clf.fit(self.x.loc[tmp_train_ix].copy(),
self.y.loc[tmp_train_ix].copy())
self.clf.append_fold_result(
) # add the training result from each fold
self.clf.append_cv_result(
) # add the training result for a single CV to the R variable
self.num_pred += 1
self.clf.set_cv_fold(self.test_fold_df)
def train_cv(self, k=10):
"""Train classifier on entire data set provided, but done in cross-validation."""
# generate indices for kfold cross validation
self.num_pred = 0 # number of predictions
self.test_fold_df = pd.DataFrame({l+1: 0 for l in range(self.total_iter)}, index=self.x.index)
for i in range(self.total_iter):
# randomize for another round
self.x, self.y = futils.randomize(self.x, self.prng)
futils.check_num_classes(self.y) # warn user if not 3 classes
# set up stratified kfold iterator
k_fold = cross_validation.StratifiedKFold(self.y,
n_folds=k)
# obtain predictions from single round of kfold validation
for nfold, (train_ix, test_ix) in enumerate(k_fold):
# retreive indices from pandas dataframe using row number
tmp_train_ix = self.x.iloc[train_ix].index
# save which genes are in the test fold
tmp_test_ix = self.x.iloc[test_ix].index
self.test_fold_df.loc[tmp_test_ix, i+1] = nfold + 1
if self.is_weighted_sample:
# figure out sample weights
num_train = len(train_ix)
sample_weight = np.zeros(num_train)
onco_ix = np.nonzero(self.y.ix[tmp_train_ix]==self.onco_num)[0]
tsg_ix = np.nonzero(self.y.ix[tmp_train_ix]==self.tsg_num)[0]
other_ix = np.nonzero(self.y.ix[tmp_train_ix]==self.other_num)[0]
sample_weight[onco_ix] = 1. / len(onco_ix)
sample_weight[tsg_ix] = 1. / len(tsg_ix)
sample_weight[other_ix] = 1. / len(other_ix)
# do training with sample weighting
self.clf.fit(self.x.ix[tmp_train_ix].copy(),
self.y.ix[tmp_train_ix].copy(),
sample_weight=sample_weight)
else:
# do training without weighting
self.clf.fit(self.x.ix[tmp_train_ix].copy(),
self.y.ix[tmp_train_ix].copy())
self.clf.append_fold_result() # add the training result from each fold
self.clf.append_cv_result() # add the training result for a single CV to the R variable
self.num_pred += 1
self.clf.set_cv_fold(self.test_fold_df)
def train(self):
"""Train classifier on entire data set provided."""
self.x, self.y = futils.randomize(self.x, self.prng)
futils.check_num_classes(self.y) # warn user if not 3 classes
self.clf.fit(self.x, self.y)
def kfold_validation(self, k=10):
"""Records the performance in terms of ROC and PR AUC for cross-validation.
Params
------
k : int (10)
Number of cross-validation folds
"""
self.num_pred = 0 # number of predictions
for i in range(self.total_iter):
# randomize for another round
self.x, self.y = futils.randomize(self.x, self.prng)
futils.check_num_classes(self.y) # warn user if not 3 classes
# initialize predicted results variables
num_genes = len(self.y)
onco_pred = np.zeros(num_genes)
onco_prob = np.zeros(num_genes)
tsg_pred = np.zeros(num_genes)
tsg_prob = np.zeros(num_genes)
overall_pred = np.zeros(num_genes)
# set up stratified kfold iterator
k_fold = cross_validation.StratifiedKFold(self.y, n_folds=k)
# evaluate k-fold cross validation
for train_ix, test_ix in k_fold:
if self.is_weighted_sample:
# weight classes by using sample weights
num_train = len(train_ix)
sample_weight = np.zeros(num_train)
onco_ix = np.nonzero(self.y[train_ix] == self.onco_num)[0]
tsg_ix = np.nonzero(self.y[train_ix] == self.tsg_num)[0]
other_ix = np.nonzero(
self.y[train_ix] == self.other_num)[0]
sample_weight[onco_ix] = 1. / len(onco_ix)
sample_weight[tsg_ix] = 1. / len(tsg_ix)
sample_weight[other_ix] = 1. / len(other_ix)
# do training
self.clf.fit(self.x.iloc[train_ix].copy(),
self.y.iloc[train_ix].copy(),
sample_weight=sample_weight)
else:
# do training without sample weights
self.clf.fit(self.x.iloc[train_ix].copy(),
self.y.iloc[train_ix].copy())
# do prediction
y_pred = self.clf.predict(self.x.iloc[test_ix])
proba_ = self.clf.predict_proba(self.x.iloc[test_ix])
# update information
overall_pred[
test_ix] = y_pred # prediction including all classes
onco_pred[test_ix] = (y_pred == self.onco_num).astype(
int) # predicted oncogenes
onco_prob[test_ix] = proba_[:, self.
onco_num] # predicted oncogenes
tsg_pred[test_ix] = (y_pred == self.tsg_num).astype(
int) # predicted oncogenes
tsg_prob[test_ix] = proba_[:,
self.tsg_num] # predicted oncogenes
# update information
true_onco = (self.y == self.onco_num).astype(int)
self._update_onco_metrics(true_onco, onco_pred, onco_prob)
true_tsg = (self.y == self.tsg_num).astype(int) # true oncogenes
self._update_tsg_metrics(true_tsg, tsg_pred, tsg_prob)
self._update_metrics(self.y, overall_pred, onco_prob, tsg_prob)
self.num_pred += 1
self._on_finish() # update info for kfold cross-validation
def train(self):
"""Train classifier on entire data set provided."""
self.x, self.y = futils.randomize(self.x, self.prng)
futils.check_num_classes(self.y) # warn user if not 3 classes
self.clf.fit(self.x, self.y)
def kfold_validation(self, k=10):
"""Records the performance in terms of ROC and PR AUC for cross-validation.
Params
------
k : int (10)
Number of cross-validation folds
"""
self.num_pred = 0 # number of predictions
for i in range(self.total_iter):
# randomize for another round
self.x, self.y = futils.randomize(self.x, self.prng)
futils.check_num_classes(self.y) # warn user if not 3 classes
# initialize predicted results variables
num_genes = len(self.y)
onco_pred = np.zeros(num_genes)
onco_prob = np.zeros(num_genes)
tsg_pred = np.zeros(num_genes)
tsg_prob = np.zeros(num_genes)
overall_pred = np.zeros(num_genes)
# set up stratified kfold iterator
k_fold = cross_validation.StratifiedKFold(self.y,
n_folds=k)
# evaluate k-fold cross validation
for train_ix, test_ix in k_fold:
if self.is_weighted_sample:
# weight classes by using sample weights
num_train = len(train_ix)
sample_weight = np.zeros(num_train)
onco_ix = np.nonzero(self.y[train_ix]==self.onco_num)[0]
tsg_ix = np.nonzero(self.y[train_ix]==self.tsg_num)[0]
other_ix = np.nonzero(self.y[train_ix]==self.other_num)[0]
sample_weight[onco_ix] = 1. / len(onco_ix)
sample_weight[tsg_ix] = 1. / len(tsg_ix)
sample_weight[other_ix] = 1. / len(other_ix)
# do training
self.clf.fit(self.x.iloc[train_ix].copy(),
self.y.iloc[train_ix].copy(),
sample_weight=sample_weight)
else:
# do training without sample weights
self.clf.fit(self.x.iloc[train_ix].copy(),
self.y.iloc[train_ix].copy())
# do prediction
y_pred = self.clf.predict(self.x.iloc[test_ix])
proba_ = self.clf.predict_proba(self.x.iloc[test_ix])
# update information
overall_pred[test_ix] = y_pred # prediction including all classes
onco_pred[test_ix] = (y_pred==self.onco_num).astype(int) # predicted oncogenes
onco_prob[test_ix] = proba_[:, self.onco_num] # predicted oncogenes
tsg_pred[test_ix] = (y_pred==self.tsg_num).astype(int) # predicted oncogenes
tsg_prob[test_ix] = proba_[:, self.tsg_num] # predicted oncogenes
# update information
true_onco = (self.y==self.onco_num).astype(int)
self._update_onco_metrics(true_onco,
onco_pred,
onco_prob)
true_tsg = (self.y==self.tsg_num).astype(int) # true oncogenes
self._update_tsg_metrics(true_tsg,
tsg_pred,
tsg_prob)
self._update_metrics(self.y,
overall_pred,
onco_prob,
tsg_prob)
self.num_pred += 1
self._on_finish() # update info for kfold cross-validation