def m_svm():
if ARGV.one_vs:
svm_model = svm.svm_to_binary(svm.svm_raw())
else:
svm_model = multi.one_against_one(svm.svm_to_binary(svm.svm_raw()))
# return milk.defaultclassifier(mode='slow', multi_strategy='1-vs-1')
learner = milk.supervised.classifier.ctransforms(
# remove nans
supervised.normalise.chkfinite(),
# normalize to [-1,1]
supervised.normalise.interval_normalise(),
# feature selection
featureselection.featureselector(
featureselection.linear_independent_features),
# sda filter
featureselection.sda_filter(),
# same parameter range as 'medium'
supervised.gridsearch(
svm_model,
params = {
'C': 2.0 ** np.arange(-2, 4),
'kernel': [ svm.rbf_kernel(2.0 ** i) for i in xrange(-4, 4) ]
}
)
)
return learner
def test_ecoc_probability():
features,labels = load()
features = features[labels < 5]
labels = labels[labels < 5]
raw = svm.svm_raw(kernel=svm.dot_kernel(), C=1.)
base = ctransforms(raw, svm.svm_sigmoidal_correction())
learner = ecoc_learner(base, probability=True)
model = learner.train(features[::2], labels[::2])
results = list(map(model.apply, features[1::2]))
results = np.array(results)
assert results.shape[1] == len(set(labels))
assert np.mean(results.argmax(1) == labels[1::2]) > .5
def bench_milk(X, y, T, valid):
#
# .. milk ..
#
from milk.supervised import svm
start = datetime.now()
learner = svm.svm_raw(
kernel=svm.rbf_kernel(sigma=1./X.shape[1]), C=1.)
model = learner.train(X,y)
pred = np.sign(map(model.apply, T))
score = np.mean(pred == valid)
return score, datetime.now() - start
def m_svm():
if ARGV.one_vs:
svm_model = svm.svm_to_binary(svm.svm_raw())
else:
svm_model = multi.one_against_one(svm.svm_to_binary(svm.svm_raw()))
# return milk.defaultclassifier(mode='slow', multi_strategy='1-vs-1')
learner = milk.supervised.classifier.ctransforms(
# remove nans
supervised.normalise.chkfinite(),
# normalize to [-1,1]
supervised.normalise.interval_normalise(),
# feature selection
featureselection.featureselector(
featureselection.linear_independent_features),
# sda filter
featureselection.sda_filter(),
# same parameter range as 'medium'
supervised.gridsearch(
svm_model,
params={
'C': 2.0**np.arange(-2, 4),
'kernel': [svm.rbf_kernel(2.0**i) for i in xrange(-4, 4)]
}))
return learner
def test_ecoc_learner():
base = milk.tests.fast_classifier.fast_classifier()
learner = milk.supervised.multi.ecoc_learner(base)
features, labels = load()
nlabels = len(set(labels))
model = learner.train(features[::2],labels[::2])
testl = np.array(model.apply_many(features[1::2]))
assert np.mean(testl == labels[1::2]) > 1./nlabels
assert testl.min() >= 0
assert testl.max() < nlabels
# This failed at one point:
learner = ecoc_learner(svm.svm_to_binary(svm.svm_raw(kernel=svm.dot_kernel(), C=1.)))
model = learner.train(features[:200], labels[:200])
assert (model is not None)
