def do_comparison(X_train, y_train, X_test, y_test, dataset):
# evaluate both svms on a given datasets, generate plots
Cs = 10.**np.arange(-4, 1)
multisvm = MultiSVM()
svm = OneSlackSSVM(MultiClassClf(), tol=0.01)
accs_pystruct, times_pystruct = eval_on_data(X_train,
y_train,
X_test,
y_test,
svm,
Cs=Cs)
accs_svmstruct, times_svmstruct = eval_on_data(X_train,
y_train,
X_test,
y_test,
multisvm,
Cs=Cs)
plot_curves(times_svmstruct,
times_pystruct,
Cs=Cs,
title="learning time (s) %s" % dataset,
filename="times_%s.pdf" % dataset)
plot_curves(accs_svmstruct,
accs_pystruct,
Cs=Cs,
title="accuracy %s" % dataset,
filename="accs_%s.pdf" % dataset)
def test_class_weights_rescale_C():
# check that our crammer-singer implementation with class weights and
# rescale_C=True is the same as LinearSVC's c-s class_weight implementation
raise SkipTest("class weight test needs update")
from sklearn.svm import LinearSVC
X, Y = make_blobs(n_samples=210,
centers=3,
random_state=1,
cluster_std=3,
shuffle=False)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X, Y = X[:170], Y[:170]
weights = len(Y) / (np.bincount(Y) * len(np.unique(Y)))
pbl_class_weight = MultiClassClf(n_features=3,
n_classes=3,
class_weight=weights,
rescale_C=True)
svm_class_weight = OneSlackSSVM(pbl_class_weight, C=10, tol=1e-5)
svm_class_weight.fit(X, Y)
try:
linearsvm = LinearSVC(multi_class='crammer_singer',
fit_intercept=False,
class_weight='balanced',
C=10)
linearsvm.fit(X, Y)
assert_array_almost_equal(svm_class_weight.w, linearsvm.coef_.ravel(),
3)
except TypeError:
# travis has a really old sklearn version that doesn't support
# class_weight in LinearSVC
pass
def crammer_singer_classifier(X_train_bias, y_train, num_classes, n_jobs=2, C=1):
model = MultiClassClf(n_features=X_train_bias.shape[1], n_classes=num_classes)
# n-slack cutting plane ssvm
n_slack_svm = NSlackSSVM(model, n_jobs=n_jobs, verbose=0,
check_constraints=False, C=C,
batch_size=100, tol=1e-2)
n_slack_svm.fit(X_train_bias, y_train)
return n_slack_svm
def test_simple_1d_dataset_cutting_plane():
# 10 1d datapoints between 0 and 1
X = np.random.uniform(size=(30, 1))
Y = (X.ravel() > 0.5).astype(np.int)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
pbl = MultiClassClf(n_features=2)
svm = NSlackSSVM(pbl, check_constraints=True, C=10000)
svm.fit(X, Y)
assert_array_equal(Y, np.hstack(svm.predict(X)))
def test_equal_class_weights():
# test that equal class weight is the same as no class weight
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
pbl = MultiClassClf(n_features=3, n_classes=3)
svm = OneSlackSSVM(pbl, C=10)
svm.fit(X_train, Y_train)
predict_no_class_weight = svm.predict(X_test)
pbl_class_weight = MultiClassClf(n_features=3,
n_classes=3,
class_weight=np.ones(3))
svm_class_weight = OneSlackSSVM(pbl_class_weight, C=10)
svm_class_weight.fit(X_train, Y_train)
predict_class_weight = svm_class_weight.predict(X_test)
assert_array_equal(predict_no_class_weight, predict_class_weight)
assert_array_almost_equal(svm.w, svm_class_weight.w)
def test_blobs_2d_subgradient():
# make two gaussian blobs
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
pbl = MultiClassClf(n_features=3, n_classes=3)
svm = SubgradientSSVM(pbl, C=1000)
svm.fit(X_train, Y_train)
assert_array_equal(Y_test, np.hstack(svm.predict(X_test)))
def test_blobs_2d_cutting_plane():
# make two gaussian blobs
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
X_train, X_test, Y_train, Y_test = X[:40], X[40:], Y[:40], Y[40:]
pbl = MultiClassClf(n_features=3, n_classes=3)
svm = NSlackSSVM(pbl, check_constraints=True, C=1000, batch_size=1)
svm.fit(X_train, Y_train)
assert_array_equal(Y_test, np.hstack(svm.predict(X_test)))
def test_class_weights():
X, Y = make_blobs(n_samples=210,
centers=3,
random_state=1,
cluster_std=3,
shuffle=False)
X = np.hstack([X, np.ones((X.shape[0], 1))])
X, Y = X[:170], Y[:170]
pbl = MultiClassClf(n_features=3, n_classes=3)
svm = OneSlackSSVM(pbl, C=10)
svm.fit(X, Y)
weights = 1. / np.bincount(Y)
weights *= len(weights) / np.sum(weights)
pbl_class_weight = MultiClassClf(n_features=3,
n_classes=3,
class_weight=weights)
svm_class_weight = OneSlackSSVM(pbl_class_weight, C=10)
svm_class_weight.fit(X, Y)
assert_greater(f1_score(Y, svm_class_weight.predict(X), average='macro'),
f1_score(Y, svm.predict(X), average='macro'))
# Loss = toeplitz(np.arange(n_classes))
# Loss = np.ones((n_classes, n_classes))
np.fill_diagonal(Loss, 0.0)
# cond_loss = np.dot(Loss, np.expand_dims(prob, 1))
# opt = np.argmin(cond_loss)
# print(cond_loss)
# print("OPTIMAL IS %f", opt)
# we add a constant 1 feature for the bias
X_train_bias = np.hstack([X_train, np.ones((X_train.shape[0], 1))])
X_test_bias = np.hstack([X_test, np.ones((X_test.shape[0], 1))])
model = MultiClassClf(n_features=X_train_bias.shape[1],
n_classes=n_classes,
Loss=Loss)
gmodel = GeneralizedMultiClassClf(n_features=X_train_bias.shape[1],
n_classes=n_classes,
Loss=Loss)
method = 'generalized'
# method = 'vanilla'
Cs = [1.]
# Cs = [6.5, 7., 7.5]
for C in Cs:
fw_bc_svm = FrankWolfeSSVM(model,
C=C,
max_iter=300,
def test_crammer_singer_model_class_weight():
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
pbl = MultiClassClf(n_features=3, n_classes=3, class_weight=[1, 2, 1])
rng = np.random.RandomState(0)
w = rng.uniform(size=pbl.size_joint_feature)
# test inference energy
x = X[0]
y, energy = pbl.inference(x, w, return_energy=True)
assert_almost_equal(energy, np.dot(w, pbl.joint_feature(x, y)))
# test inference_result:
energies = [np.dot(w, pbl.joint_feature(x, y_hat)) for y_hat in range(3)]
assert_equal(np.argmax(energies), y)
# test loss_augmented inference energy
y, energy = pbl.loss_augmented_inference(x, Y[0], w, return_energy=True)
assert_almost_equal(energy,
np.dot(w, pbl.joint_feature(x, y)) + pbl.loss(Y[0], y))
# test batch versions
Y_batch = pbl.batch_inference(X, w)
Y_ = [pbl.inference(x, w) for x in X]
assert_array_equal(Y_batch, Y_)
Y_batch = pbl.batch_loss_augmented_inference(X, Y, w)
Y_ = [pbl.loss_augmented_inference(x, y, w) for x, y in zip(X, Y)]
assert_array_equal(Y_batch, Y_)
loss_batch = pbl.batch_loss(Y, Y_)
loss = [pbl.loss(y, y_) for y, y_ in zip(Y, Y_)]
assert_array_equal(loss_batch, loss)
def test_crammer_singer_model_class_weight():
X, Y = make_blobs(n_samples=80, centers=3, random_state=42)
# we have to add a constant 1 feature by hand :-/
X = np.hstack([X, np.ones((X.shape[0], 1))])
pbl = MultiClassClf(n_features=3, n_classes=3, class_weight=[1, 2, 1])
rng = np.random.RandomState(0)
w = rng.uniform(size=pbl.size_psi)
# test inference energy
x = X[0]
y, energy = pbl.inference(x, w, return_energy=True)
assert_almost_equal(energy, np.dot(w, pbl.psi(x, y)))
# test inference_result:
energies = [np.dot(w, pbl.psi(x, y_hat)) for y_hat in xrange(3)]
assert_equal(np.argmax(energies), y)
# test loss_augmented inference energy
y, energy = pbl.loss_augmented_inference(x, Y[0], w, return_energy=True)
assert_almost_equal(energy, np.dot(w, pbl.psi(x, y)) + pbl.loss(Y[0], y))
# test batch versions
Y_batch = pbl.batch_inference(X, w)
Y_ = [pbl.inference(x, w) for x in X]
assert_array_equal(Y_batch, Y_)
Y_batch = pbl.batch_loss_augmented_inference(X, Y, w)
Y_ = [pbl.loss_augmented_inference(x, y, w) for x, y in zip(X, Y)]
assert_array_equal(Y_batch, Y_)
loss_batch = pbl.batch_loss(Y, Y_)
loss = [pbl.loss(y, y_) for y, y_ in zip(Y, Y_)]
assert_array_equal(loss_batch, loss)
import numpy as np
from sklearn.datasets import load_digits
from pystruct.models import MultiClassClf
from pystruct.learners import FrankWolfeSSVM
digits = load_digits()
X_train, y_train = digits.data, digits.target
X_train = X_train / 16.
y_train = y_train.astype(np.int)
model = MultiClassClf()
bcfw = FrankWolfeSSVM(model=model,
C=.1,
max_iter=1000,
tol=0.1,
verbose=3,
check_dual_every=10)
batch_fw = FrankWolfeSSVM(model=model,
C=.1,
max_iter=1000,
batch_mode=True,
tol=0.1,
verbose=3,
check_dual_every=10)
bcfw.fit(X_train, y_train)
batch_fw.fit(X_train, y_train)
itr = [i * 10 for i in range(0, 12)]
plt.plot(list(itr), list(d_gapBCFW), 'go-', label='line 1', linewidth=2)
plt.title('BCFW duality gap at each 10 timesteps')
#Add the heart rate features
x = np.append(x, window_with_timestamp_and_label[-2])
X = np.append(X, np.reshape(x, (1,-1)), axis=0)
y = np.append(y, label)
print("Finished feature extraction over {} windows".format(len(X)))
print("Unique labels found: {}".format(set(y)))
sys.stdout.flush()
y = y.astype(int)
model = MultiClassClf(n_features=X.shape[1], n_classes=2)
print("\n")
print("---------------------- Grid Search -------------------------")
#grid_search
param_grid = [
{
'C':[0.075,0.05,0.1],
'batch_size':[100,1000,150],
'tol':[1e-2,-10],
'inactive_window':[50,100,10],
'inactive_threshold':[1e-5,1e-4,1e-6]
}
]