def test_mismatch_kernel(simple_seq): n = len(simple_seq) kernel = kernels.MismatchKernel(3, 1, 2) K, _ = kernel.fit(simple_seq) print(K) assert (K[:n // 2, :n // 2] >= 1).all() assert (K[n // 2:, n // 2:] >= 1).all() zeros = np.zeros((n // 2, n // 2)) np.testing.assert_equal(K[:n // 2, n // 2:], zeros) np.testing.assert_equal(K[n // 2:, :n // 2], zeros)
def main_sum():
ls_kernel = [
kernels.MismatchKernel(12, 1, 4, False),
kernels.MismatchKernel(12, 1, 4, False),
kernels.MismatchKernel(9, 1, 4, False)
]
ls_kernel_prime = [
kernels.Gaussian(.1, False),
kernels.Gaussian(.1, False),
kernels.Gaussian(.1, False)
]
# ls_methods = [
# methods.SVM(kernel_0, reg_val=.1),
# methods.SVM(kernel_1, reg_val=.1),
# methods.SVM(kernel_2, reg_val=.01),
# ]
for i in range(3):
print("##################", f"i={i}")
# X = read_write.read_X100(f"data/Xtr{i}_mat100.csv")
X = read_write.read(f"data/Xtr{i}.csv")
length = X.shape[1]
X_cat = np.concatenate((X, X), axis=-1)
# X_test = read_write.read_X100(f"data/Xte{i}_mat100.csv")
# X_test = read_write.read(f"data/Xte{i}.csv")
y = read_write.read_labels(f"data/Ytr{i}.csv")
kernel_class = kernels.SumKernel
method_class = methods.KernelRidgeRegression
params_kernel = [(ls_kernel[i], ls_kernel_prime[i], length, length,
False)]
# [(k, max(floor(k / 10), 1), 4) for k in range(6, 18, 3)] # 10. **
reg_vals = 10.**np.arange(-2, 3, 1 / 4)
validation(X_cat, y, kernel_class, method_class, params_kernel,
reg_vals)
def main_rendu(accuracy_on_train_set=False):
ls_kernel = [
kernels.MismatchKernel(12, 2, 4, False),
kernels.MismatchKernel(12, 2, 4, True),
kernels.MismatchKernel(9, 2, 4, False)
]
ls_reg_val = [100 * 0.03162277660168379, .1, 1000 * 0.03162277660168379]
ls_methods = [
methods.KernelRidgeRegression(ls_kernel[i], reg_val=ls_reg_val[i])
for i in range(3)
]
for i in range(3):
print("##################", f"i={i}")
# X = read_write.read_X100(f"data/Xtr{i}_mat100.csv")
X = read_write.read(f"data/Xtr{i}.csv")
# print(X.shape)
# X_cat = np.concatenate((X, X), axis=-1)
# X_test = read_write.read_X100(f"data/Xte{i}_mat100.csv")
X_test = read_write.read(f"data/Xte{i}.csv")
# X_test_cat = np.concatenate((X_test, X_test), axis=-1)
y = read_write.read_labels(f"data/Ytr{i}.csv")
# X_cat = np.concatenate((X, X), axis=-1)
# X_test_cat = np.concatenate((X_test, X_test), axis=-1)
ls_methods[i].learn(X, y)
# FOR ACCURACY ON TRAINING SET
if accuracy_on_train_set:
y_pred = ls_methods[i].predict(X)
print(methods.accuracy(y, y_pred))
y_test = ls_methods[i].predict(X_test)
read_write.write(y_test,
"predictions/Yte.csv",
offset=i * 1000,
append=(i != 0))
def get_kernel(conf) -> kernels.Kernel:
all_kernels = conf.kernels.values_()
list_kernels = []
list_coefs = []
for k in range(len(all_kernels)):
kernel = all_kernels[k]["name"]
if not conf.mkl:
coef = conf.coefs[k]
assert coef >= 0, f"Coefficient for kernel {k + 1} must be positive."
list_coefs.append(coef)
kernel_conf = conf.kernels[k]
assert kernel in ['onehot', 'spectrum', "mismatch", "substring", "local-alignment"], "Unknown requested kernel."
if kernel == "spectrum":
default_args = {"length": 3}
default_args.update(kernel_conf.args.values_())
kernel = kernels.SpectrumKernel(conf.memoize, default_args['length'])
elif kernel == "mismatch":
default_args = {"k": 3, "m": 2}
default_args.update(kernel_conf.args.values_())
kernel = kernels.MismatchKernel(conf.memoize, default_args['k'], default_args['m'])
elif kernel == "substring":
default_args = {"length": 3, "lambda_decay": 0.05}
default_args.update(kernel_conf.args.values_())
kernel = kernels.SubstringKernel(conf.memoize, default_args['length'])
elif kernel == "local-alignment":
default_args = {"beta": 0.05, "d": 1, "e": 11}
default_args.update(kernel_conf.args.values_())
kernel = kernels.LocalAlignmentKernel(conf.memoize, default_args['beta'], default_args['d'],
default_args['e'])
else:
kernel = kernels.OneHotKernel(conf.memoize)
kernel.set_args(kernel_conf.type, kernel_conf.gamma, kernel_conf.degree, kernel_conf.r)
list_kernels.append(kernel)
if conf.mkl:
kernel = kernels.SimpleMKL(conf.memoize, list_kernels)
else:
kernel = kernels.SumKernel(conf.memoize, list_kernels, list_coefs)
kernel.set_args(normalize=conf.normalize)
return kernel
TEST_METHODS = [
methods.KernelRidgeRegression,
methods.KernelLogisticRegression,
methods.SVM
]
TEST_KERNELS = [
(kernels.Linear(), .1),
(kernels.Gaussian(.1), .1),
(kernels.Polynomial(degree=3), .1)
]
TEST_SEQ_KERNELS = [
# kernels.SpectrumKernel(1),
# kernels.SpectrumKernel(2),
kernels.MismatchKernel(3, 1, A)
]
@pytest.mark.parametrize('method', TEST_METHODS)
@pytest.mark.parametrize('kernel_with_reg', TEST_KERNELS)
def test_methods(method, kernel_with_reg, x_train, y_train, x_test, y_test):
kernel, reg = kernel_with_reg
kernel.normalize = True
meth = method(kernel, reg_val=reg)
meth.learn(x_train, y_train)
y_est = meth.predict(x_train)
np.testing.assert_equal(y_est, y_train)
y_est = meth.predict(x_test)
np.testing.assert_equal(y_est, y_test)