import numpy as np
import sklearn
from sklearn.linear_model import ARDRegression, LinearRegression
from sklearn.metrics import roc_curve, auc
import sklearn.linear_model
import azimuth.util
import azimuth.metrics as ranking_metrics
import azimuth.predict
import numbers
def ARDRegression_on_fold(feature_sets, train, test, y, y_all, X, dim, dimsum, learn_options):
'''
'''
clf = ARDRegression()
clf.fit(X[train], y[train][:, 0])
y_pred = clf.predict(X[test])[:, None]
return y_pred, clf
def train_linreg_model(alpha, l1r, learn_options, fold, X, y, y_all):
'''
fold is something like train_inner (boolean array specifying what is in the fold)
'''
if learn_options["penalty"] == "L2":
clf = sklearn.linear_model.Ridge(alpha=alpha, fit_intercept=learn_options["fit_intercept"], normalize=learn_options['normalize_features'], copy_X=True, max_iter=None, tol=0.001, solver='auto')
weights = get_weights(learn_options, fold, y, y_all)
clf.fit(X[fold], y[fold], sample_weight=weights)
elif learn_options["penalty"] == 'EN' or learn_options["penalty"] == 'L1':
if learn_options["loss"] == "squared":
clf = sklearn.linear_model.ElasticNet(alpha=alpha, l1_ratio=l1r, fit_intercept=learn_options["fit_intercept"], normalize=learn_options['normalize_features'], max_iter=3000)
elif learn_options["loss"] == "huber":
clf = sklearn.linear_model.SGDRegressor('huber', epsilon=0.7, alpha=alpha,
l1_ratio=l1r, fit_intercept=learn_options["fit_intercept"], n_iter=10,
penalty='elasticnet', shuffle=True, normalize=learn_options['normalize_features'])
clf.fit(X[fold], y[fold])
return clf
def logreg_on_fold(feature_sets, train, test, y, y_all, X, dim, dimsum, learn_options):
'''
(L1/L2 penalized) logistic reggresion using scikitlearn
'''
assert len(np.unique(y)) <= 2, "if using logreg need binary targets"
assert learn_options["weighted"] is None, "cannot do weighted Log reg"
assert learn_options['feature_select'] is False, "cannot do feature selection yet in logistic regression--see linreg_on_fold to implement"
cv, n_folds = set_up_inner_folds(learn_options, y_all.iloc[train])
assert learn_options['penalty'] == "L1" or learn_options['penalty'] == "L2", "can only use L1 or L2 with logistic regression"
tol = 0.00001#0.0001
performance = np.zeros((len(learn_options["alpha"]), 1))
# degenerate_pred = np.zeros((len(learn_options["alpha"])))
for train_inner, test_inner in cv:
for i, alpha in enumerate(learn_options["alpha"]):
clf = sklearn.linear_model.LogisticRegression(penalty=learn_options['penalty'].lower(), dual=False, fit_intercept=learn_options["fit_intercept"], class_weight=learn_options["class_weight"], tol=tol, C=1.0/alpha)
clf.fit(X[train][train_inner], y[train][train_inner].flatten())
#tmp_pred = clf.predict(X[train][test_inner])
tmp_pred = clf.predict_proba(X[train][test_inner])[:,1]
if learn_options["training_metric"] == "AUC":
fpr, tpr, _ = roc_curve(y_all[learn_options["ground_truth_label"]][train][test_inner], tmp_pred)
assert ~np.any(np.isnan(fpr)), "found nan fpr"
assert ~np.any(np.isnan(tpr)), "found nan tpr"
tmp_auc = auc(fpr, tpr)
performance[i] += tmp_auc
else:
raise Exception("can only use AUC metric for cv with classification")
performance /= n_folds
max_score_ind = np.where(performance == np.nanmax(performance))
assert max_score_ind != len(performance), "enlarge alpha range as hitting max boundary"
# in the unlikely event of tied scores, take the first one.
if len(max_score_ind[0]) > 1:
max_score_ind = [max_score_ind[0][0], max_score_ind[1][0]]
best_alpha = learn_options["alpha"][max_score_ind[0]]
best_alpha = best_alpha[0]
if not isinstance(best_alpha, numbers.Number):
raise Exception("best_alpha must be a number but is %s" % type(best_alpha))
print "\t\tbest alpha is %f from range=%s" % (best_alpha, learn_options["alpha"][[0, -1]])
max_perf = np.nanmax(performance)
if max_perf < 0.0:
raise Exception("performance is negative")
print "\t\tbest performance is %f" % np.nanmax(performance)
clf = sklearn.linear_model.LogisticRegression(penalty=learn_options['penalty'],
dual=False, fit_intercept=learn_options["fit_intercept"], class_weight=learn_options["class_weight"], tol=tol, C=1.0/best_alpha)
clf.fit(X[train], y[train].flatten())
# debugging check that get samed paramter estimation when have no regularization and use
# either data with only that feature on, or all data), AND WITH NO INTERCEPT
if False:
# grab only feature "GA3"
keep_ind = np.where(feature_sets['mutletpos'].columns=="GA3")[0]
print "%s, %s" % (str(clf.intercept_ ), str(clf.coef_[0, keep_ind]))
clf.fit(X[train][:,keep_ind], y[train].flatten())
print "%s, %s" % (str(clf.intercept_ ), str(clf.coef_))
import ipdb; ipdb.set_trace()
#y_pred = clf.predict(X[test])
y_pred = clf.predict_proba(X[test])[:,1]
y_pred = y_pred[:, None]
#fpr, tpr, _ = roc_curve(y, y_pred); tmp_auc = auc(fpr, tpr)
#import ipdb; ipdb.set_trace()
return y_pred, clf
def linreg_on_fold(feature_sets, train, test, y, y_all, X, dim, dimsum, learn_options):
'''
linreg using scikitlearn, using more standard regression models with penalization requiring
nested-cross-validation
'''
if learn_options["weighted"] is not None and (learn_options["penalty"] != "L2" or learn_options["method"] != "linreg"):
raise NotImplementedError("weighted prediction not implemented for any methods by L2 at the moment")
if not learn_options.has_key("fit_intercept"):
learn_options["fit_intercept"] = True
if not learn_options.has_key('normalize_features'):
learn_options['normalize_features'] = True
cv, n_folds = set_up_inner_folds(learn_options, y_all.iloc[train])
if learn_options['penalty'] == "L1":
l1_ratio = [1.0]
elif learn_options['penalty'] == "L2":
l1_ratio = [0.0]
elif learn_options['penalty'] == "EN": # elastic net
l1_ratio = np.linspace(0.0, 1.0, 20)
performance = np.zeros((len(learn_options["alpha"]), len(l1_ratio)))
degenerate_pred = np.zeros((len(learn_options["alpha"])))
for train_inner, test_inner in cv:
for i, alpha in enumerate(learn_options["alpha"]):
for j, l1r in enumerate(l1_ratio):
clf = train_linreg_model(alpha, l1r, learn_options, train_inner, X[train], y[train], y_all.iloc[train])
if learn_options["feature_select"]:
clf, tmp_pred = feature_select(clf, learn_options, test_inner, train_inner, X[train], y[train])
else:
tmp_pred = clf.predict(X[train][test_inner])
if learn_options["training_metric"] == "AUC":
fpr, tpr, _ = roc_curve(y_all[learn_options["ground_truth_label"]][train][test_inner], tmp_pred)
assert ~np.any(np.isnan(fpr)), "found nan fpr"
assert ~np.any(np.isnan(tpr)), "found nan tpr"
tmp_auc = auc(fpr, tpr)
performance[i, j] += tmp_auc
elif learn_options['training_metric'] == 'spearmanr':
spearman = azimuth.util.spearmanr_nonan(y_all[learn_options['ground_truth_label']][train][test_inner], tmp_pred.flatten())[0]
performance[i, j] += spearman
elif learn_options['training_metric'] == 'score':
performance[i, j] += clf.score(X[test_inner], y_all[learn_options['ground_truth_label']][train][test_inner])
elif learn_options["training_metric"] == "NDCG":
assert "thresh" not in learn_options["ground_truth_label"], "for NDCG must not use thresholded ranks, but pure ranks"
# sorted = tmp_pred[np.argsort(y_all[ground_truth_label].values[test_inner])[::-1]].flatten()
# sortedgt = np.sort(y_all[ground_truth_label].values[test_inner])[::-1].flatten()
# tmp_perf = ranking_metrics.ndcg_at_k_ties(sorted, learn_options["NDGC_k"], sortedgt)
tmp_truth = y_all[learn_options["ground_truth_label"]].values[train][test_inner].flatten()
tmp_perf = ranking_metrics.ndcg_at_k_ties(tmp_truth, tmp_pred.flatten(), learn_options["NDGC_k"])
performance[i, j] += tmp_perf
degenerate_pred_tmp = len(np.unique(tmp_pred)) < len(tmp_pred)/2.0
degenerate_pred[i] += degenerate_pred_tmp
# sanity checking metric wrt ties, etc.
# rmse = np.sqrt(np.mean((tmp_pred - tmp_truth)**2))
tmp_pred_r, tmp_truth_r = ranking_metrics.rank_data(tmp_pred, tmp_truth)
# rmse_r = np.sqrt(np.mean((tmp_pred_r-tmp_truth_r)**2))
performance /= n_folds
max_score_ind = np.where(performance == np.nanmax(performance))
assert max_score_ind != len(performance), "enlarge alpha range as hitting max boundary"
# assert degenerate_pred[max_score_ind[0][0]]==0, "found degenerate predictions at max score"
# in the unlikely event of tied scores, take the first one.
if len(max_score_ind[0]) > 1:
max_score_ind = [max_score_ind[0][0], max_score_ind[1][0]]
best_alpha, best_l1r = learn_options["alpha"][max_score_ind[0]], l1_ratio[max_score_ind[1]]
print "\t\tbest alpha is %f from range=%s" % (best_alpha, learn_options["alpha"][[0, -1]])
if learn_options['penalty'] == "EN":
print "\t\tbest l1_ratio is %f from range=%s" % (best_l1r, l1_ratio[[0, -1]])
max_perf = np.nanmax(performance)
if max_perf < 0.0:
raise Exception("performance is negative")
print "\t\tbest performance is %f" % max_perf
clf = train_linreg_model(best_alpha, l1r, learn_options, train, X, y, y_all)
if learn_options["feature_select"]:
raise Exception("untested in a long time, should double check")
clf, y_pred = feature_select(clf, learn_options, test, train, X, y)
else:
y_pred = clf.predict(X[test])
if learn_options["penalty"] != "L2":
y_pred = y_pred[:, None]
return y_pred, clf
def feature_select(clf, learn_options, test_inner, train_inner, X, y):
assert not learn_options["weighted"] is not None, "cannot currently do feature selection with weighted regression"
assert learn_options["loss"] is not "huber", "won't use huber loss function with feature selection"
non_zero_coeff = (clf.coef_ != 0.0)
if non_zero_coeff.sum() > 0:
clf = LinearRegression()
clf.fit(X[train_inner][:, non_zero_coeff.flatten()], y[train_inner])
tmp_pred = clf.predict(X[test_inner][:, non_zero_coeff.flatten()])
else:
tmp_pred = np.ones_like(test_inner)
return clf, tmp_pred
def get_weights(learn_options, fold, y, y_all):
'''
fold is an object like train_inner which is boolean for which indexes are in the fold
'''
weights = None
if learn_options["weighted"] == "variance":
weights = 1.0/y_all["variance"].values[fold]
elif learn_options["weighted"] == "ndcg":
# DCG: r[0] + np.sum(r[1:] / np.log2(np.arange(2, r.size + 1)))
N = len(fold)
r = np.ones(N)
discount = np.concatenate((np.array([r[0]]), r[1:] / np.log2(np.arange(2, r.size + 1))))[::1]
ind = np.argsort(y[fold], axis=0).flatten()
weights = np.ones(len(ind))
weights[ind] = discount
elif learn_options["weighted"] == "rank":
N = len(y[fold])
inverse_ranks = (np.arange(N) + 1.0)[::-1]
ind = np.argsort(y[fold], axis=0).flatten()
weights = np.ones(len(ind))
weights[ind] = inverse_ranks
elif learn_options["weighted"] == "score":
N = len(y[fold])
score = y[fold] + np.abs(np.min(y[fold]))
ind = np.argsort(y[fold], axis=0).flatten()
weights = np.ones(len(ind))
weights[ind] = score
elif learn_options["weighted"] == "random":
N = len(y[fold])
weights = np.random.rand(N)
elif learn_options["weighted"] is not None:
raise Exception("invalid weighted type, %s" % learn_options["weighted"])
# plt.plot(weights, y[train_inner],'.')
return weights
def set_up_inner_folds(learn_options, y):
label_encoder = sklearn.preprocessing.LabelEncoder()
label_encoder.fit(y['Target gene'].values)
gene_classes = label_encoder.transform(y['Target gene'].values)
n_genes = len(np.unique(gene_classes))
if learn_options['ignore_gene_level_for_inner_loop'] or learn_options["cv"] == "stratified" or n_genes==1:
if 'n_folds' not in learn_options.keys():
n_folds = len(np.unique(gene_classes))
else:
n_folds = learn_options['n_folds']
cv = sklearn.cross_validation.StratifiedKFold(gene_classes, n_folds=n_folds, shuffle=True)
elif learn_options["cv"] == "gene":
gene_list = np.unique(y['Target gene'].values)
cv = []
for gene in gene_list:
cv.append(azimuth.predict.get_train_test(gene, y))
n_folds = len(cv)
return cv, n_folds
