def ksvm_cross_validation(k,
X,
Y,
kfun,
kparam,
lambda_,
lr=1e-3,
steps=10000):
m = X.shape[0]
# Randomly assign a fold index to each sample.
folds = np.arange(m) % k
np.random.shuffle(folds)
correct_predictions = 0
# For each fold...
for fold in range(k):
Xtrain = X[folds != fold, :]
Ytrain = Y[folds != fold]
# Train a model
alpha, b = pvml.ksvm_train(Xtrain,
Ytrain,
kfun,
kparam,
lambda_,
lr=lr,
steps=steps)
# Evaluate the model on the left-out fold
Xval = X[folds == fold, :]
Yval = Y[folds == fold]
pred, _ = pvml.ksvm_inference(Xval, Xtrain, alpha, b, kfun, kparam)
print((pred == Yval).mean())
correct_predictions += (pred == Yval).sum()
return correct_predictions / m
def inference(self, X):
ret = pvml.ksvm_inference(X, self.Xtrain, self.alpha, self.b,
self.kfun, self.kparam)
labels, logits = ret
return labels, logits + 0.5
def test_wrong_bias(self):
X = np.linspace(-1, 1, 10).reshape(5, 2)
alpha = np.linspace(-1, 1, 5)
b = np.linspace(-1, 1, 5)
with self.assertRaises(ValueError):
pvml.ksvm_inference(X, X, alpha, b, "polynomial", 2)
def test_wrong_coefficients1(self):
X = np.linspace(-1, 1, 10)
alpha = np.linspace(-1, 1, 10)
b = 0
with self.assertRaises(ValueError):
pvml.ksvm_inference(X, X, alpha, b, "polynomial", 2)
def test_train2(self):
X, Y = test_data.separable_hypercubes_data_set(12, 2)
alpha, b = pvml.ksvm_train(X, Y, "polynomial", 2, 0, lr=10, steps=1000)
Yhat, P = pvml.ksvm_inference(X, X, alpha, b, "polynomial", 2)
self.assertListEqual(Y.tolist(), Yhat.tolist())
self.assertListEqual(Yhat.tolist(), (P > 0).tolist())
def test_train1(self):
X, Y = test_data.separable_circle_data_set(50, 2)
alpha, b = pvml.ksvm_train(X, Y, "rbf", 0.1, 0, lr=1e-1, steps=1000)
Yhat, P = pvml.ksvm_inference(X, X, alpha, b, "rbf", 0.1)
self.assertListEqual(Y.tolist(), Yhat.tolist())
self.assertListEqual(Yhat.tolist(), (P > 0).tolist())
