def run_kfold_svm(method, fv, fold=1):
print("---run svm for method:%s of fv:%s fold:%d---" % (method, fv, fold))
train_items, val_items, test_items = load_kfold_fv(method, fv, fold)
train_gene, train_fv, train_label = zip(*train_items)
val_gene, val_fv, val_label = zip(*val_items)
test_gene, test_fv, test_label = zip(*test_items)
scaler = StandardScaler()
scaler.fit(train_fv)
train_fv = np.stack(scaler.transform(train_fv))
val_fv = np.stack(scaler.transform(val_fv))
test_fv = np.stack(scaler.transform(test_fv))
train_label = np.stack(train_label)
val_label = np.stack(val_label)
test_label = np.stack(test_label)
best_classifier, best_f1 = run_classifier(train_fv, train_label, val_fv,
val_label, 'linear')
tdir = os.path.join("result/%s_%s" % (method, fv))
if not os.path.exists(tdir):
os.mkdir(tdir)
joblib.dump(best_classifier, "%s/fold%d_linear.joblib" % (tdir, fold))
C_list = [0.001, 0.01, 0.1, 1, 10, 100, 1000]
gamma_list = ['auto', 0.0001, 0.001, 0.01, 0.1, 1]
best_k = 'linear'
best_C = None
best_gamma = None
for C in C_list:
for gamma in gamma_list:
classifier, val_f1 = run_classifier(train_fv, train_label, val_fv,
val_label, 'rbf', C, gamma)
if val_f1 > best_f1:
best_f1 = val_f1
best_classifier = classifier
best_k = 'rbf'
best_C = C
best_gamma = gamma
name = 'fold%d_rbf_C%s_g%s' % (fold, C, gamma)
pth = os.path.join(tdir, "%s.joblib" % name)
joblib.dump(classifier, pth)
test_pd = best_classifier.predict(test_fv)
path = os.path.join(tdir, "fold%d.txt" % fold)
from contextlib import redirect_stdout
with open(path, 'w') as f:
with redirect_stdout(f):
print("\n---test res---", "K:", best_k, "C:", best_C, 'gamma',
best_gamma)
npmetrics.print_metrics(test_label, test_pd)
def run_classifier(train_fv,
train_label,
val_fv,
val_label,
k,
C=1.0,
gamma='auto'):
estimator = SVC(kernel=k, C=C, gamma=gamma)
classifier = OneVsRestClassifier(estimator, n_jobs=-1)
classifier.fit(train_fv, train_label)
val_pd = classifier.predict(val_fv)
val_f1 = npmetrics.label_f1_macro(val_label, val_pd)
print("\n---val---", "k:", k, 'C:', C, 'gamma', gamma, 'f1:', val_f1)
npmetrics.print_metrics(val_label, val_pd)
return classifier, val_f1
def svm(fv='fv0'):
train_items, val_items, test_items = load_fv(fv)
print("train items", len(train_items))
print("val items", len(val_items))
print("test items", len(test_items))
train_gene, train_fv, train_label = zip(*train_items)
val_gene, val_fv, val_label = zip(*val_items)
test_gene, test_fv, test_label = zip(*test_items)
print("-------run svm for bov--------", fv)
scaler = StandardScaler()
scaler.fit(train_fv)
train_fv = np.stack(scaler.transform(train_fv))
val_fv = np.stack(scaler.transform(val_fv))
test_fv = np.stack(scaler.transform(test_fv))
train_label = np.stack(train_label)
val_label = np.stack(val_label)
test_label = np.stack(test_label)
kernels = ['linear', 'rbf', 'poly', 'sigmoid']
class_weights = ['balanced', None]
best_f1 = 0.0
best_classifier = None
best_k = None
best_b = None
for k in kernels:
for b in class_weights:
estimator = SVC(kernel=k, class_weight=b)
classifier = OneVsRestClassifier(estimator, n_jobs=-1)
classifier.fit(train_fv, train_label)
val_pd = classifier.predict(val_fv)
val_f1 = npmetrics.label_f1_macro(val_label, val_pd)
print("\n---svm for bov---", "k:", k, 'b:', b, 'f1:', val_f1)
npmetrics.print_metrics(val_label, val_pd)
if val_f1 > best_f1:
best_f1 = val_f1
best_classifier = classifier
best_k = k
best_b = b
test_pd = best_classifier.predict(test_fv)
print("\n---svm for bov test result---", "k:", best_k, "b:", best_b)
npmetrics.print_metrics(test_label, test_pd)
def run_svm(method, fv):
print("---run svm for method:%s of fv:%s---" % (method, fv))
train_items, val_items, test_items = load_fv(method, fv)
train_gene, train_fv, train_label = zip(*train_items)
val_gene, val_fv, val_label = zip(*val_items)
test_gene, test_fv, test_label = zip(*test_items)
scaler = StandardScaler()
scaler.fit(train_fv)
train_fv = np.stack(scaler.transform(train_fv))
val_fv = np.stack(scaler.transform(val_fv))
test_fv = np.stack(scaler.transform(test_fv))
train_label = np.stack(train_label)
val_label = np.stack(val_label)
test_label = np.stack(test_label)
best_classifier, best_f1 = run_classifier(train_fv, train_label, val_fv,
val_label, 'linear')
joblib.dump(best_classifier,
"../result/%s_%s_linear.joblib" % (method, fv))
C_list = [0.001, 0.01, 0.1, 1, 10, 100, 1000]
gamma_list = ['auto', 0.0001, 0.001, 0.01, 0.1, 1]
best_k = 'linear'
best_C = None
best_gamma = None
for C in C_list:
for gamma in gamma_list:
classifier, val_f1 = run_classifier(train_fv, train_label, val_fv,
val_label, 'rbf', C, gamma)
name = '%s_%s_rbf_C%s_g%s' % (method, fv, C, gamma)
pth = os.path.join("../result/%s.joblib" % name)
joblib.dump(classifier, pth)
if val_f1 > best_f1:
best_f1 = val_f1
best_classifier = classifier
best_k = 'rbf'
best_C = C
best_gamma = gamma
test_pd = best_classifier.predict(test_fv)
print("\n---test res---", "K:", best_k, "C:", best_C, 'gamma', best_gamma)
npmetrics.print_metrics(test_label, test_pd)
