def check_jthunder(self, clf: thundersvm.SVC):
self.assertTrue(
np.allclose(
clf.decision_function(self.ds_test.xs).flatten(),
-jthunder.decision_function(clf, self.jtest_xs),
rtol=1e-4,
atol=1e-4,
))
self.assertTrue(
all(
clf.predict(self.ds_test.xs).astype(np.int) ==
jthunder.predict(clf, self.jtest_xs)))
self.assertTrue(
np.allclose(jthunder.norm2(clf), jthunder.norm2_naive(clf)))
def run(self):
model = self.exp_data['model']
base_workspace = self.exp_data['base_workspace']
if model == 'ridge':
space = [Real(1e-3, 1e+3, prior='log-uniform')]
optimize_types = ['alpha']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model = Ridge
x0 = [1.0]
elif model == 'kernel_ridge':
space = [
Categorical(['poly', 'rbf', 'sigmoid']),
Real(1e-3, 1e+3, prior='log-uniform'),
Integer(1, 8),
Real(1e-6, 1e+1, prior='log-uniform'),
Real(-10, 10)
]
optimize_types = ['kernel', 'alpha', 'degree', 'gamma', 'coef0']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model = KernelRidge
x0 = ['poly', 1.0, 3, 1 / 300, 0]
elif model == 'kernel_ridge_separation':
space = [Real(1e-3, 1e+3, prior='log-uniform')]
optimize_types = ['alpha']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model_fixed_params = {
'kernel': 'poly',
'degree': 3,
'gamma': 1 / 300,
'coef0': 0
}
minimizer.model = KernelRidge
x0 = [1.0]
elif model == 'kernel_ridge_random':
space = [Real(1e-3, 1e+3, prior='log-uniform')]
optimize_types = ['alpha']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
kernel = random.choice(
['poly', 'rbf', 'laplacian', 'sigmoid', 'cosine'])
degree = random.randint(1, 10)
coef = random.uniform(-5, 5)
minimizer.model_fixed_params = {
'kernel': kernel,
'degree': degree,
'gamma': None,
'coef0': coef
}
minimizer.model = KernelRidge
x0 = [1.0]
elif model == 'nn':
space = [
Integer(16, 256),
Real(1e-5, 1, prior='log-uniform'),
]
optimize_types = ['n_hidden_units', 'lr']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model_fixed_params = {
'build_fn': self.build_nn,
'epochs': 20,
'batch_size': 256
}
minimizer.model = KerasRegressor
x0 = [64, 0.001]
elif model == 'kernel_svm':
space = [
Categorical(['poly', 'rbf', 'sigmoid']),
Real(1e-3, 1e+3, prior='log-uniform'),
Integer(1, 8),
Real(1e-6, 1e+1, prior='log-uniform'),
Real(-10, 10)
]
optimize_types = ['kernel', 'C', 'degree', 'gamma', 'coef0']
minimizer = MagnitudeAxisMinimizer(base_workspace, optimize_types,
gp_minimize)
minimizer.model_fixed_params = {
'cache_size': 8000,
'max_iter': 10000
}
minimizer.model = SVR
x0 = ['poly', 1.0, 3, 1 / 300, 0]
elif model == 'pca':
if not hasattr(self, 'pca_component'):
pca = PCA(n_components=1)
_, X_nums, *_ = prepare_separation_data(
self.name.split('_')[0] + '.txt')
pca.fit(X_nums)
# pca.fit(np.concatenate([base_workspace['X'],base_workspace['X_test']]))
self.pca_component = pca.components_ # 1xd
error = self.evaluate_w(base_workspace['X_test'],
self.pca_component,
base_workspace['y_test'])
return error
elif model == 'proj_pca':
if not hasattr(self, 'proj_pca_component'):
X, X_nums, y_label, _ = prepare_separation_data(
self.name.split('_')[0] + '.txt')
svc = SVC(kernel='linear',
degree=3,
gamma=1 / 300,
coef0=0,
C=1,
cache_size=4000,
class_weight='balanced',
verbose=True)
svc.fit(X, y_label)
beta = svc.coef_ # 1xd
# X_pred = svc.decision_function(X) nx1
# X_nums = np.concatenate([base_workspace['X'],base_workspace['X_test']])
X_proj = X_nums - (
(svc.decision_function(X_nums) / beta @ beta.T) @ beta
) # nxd
pca = PCA(n_components=1)
pca.fit(X_proj)
self.proj_pca_component = pca.components_ # 1xd
error = self.evaluate_w(base_workspace['X_test'],
self.proj_pca_component,
base_workspace['y_test'])
return error
elif model == 'kernel_proj_pca':
pass
else:
assert False
res = minimizer.minimize(space, n_calls=40, verbose=True, x0=x0)
if self.save_results:
skopt.dump(res, self.name + '.pkl', store_objective=False)
params = {type: v for type, v in zip(minimizer.optimize_types, res.x)}
if hasattr(minimizer, 'model_fixed_params'):
params = {**params, **minimizer.model_fixed_params}
error = self.fit_test_best_model(minimizer.model, base_workspace['X'],
base_workspace['y'],
base_workspace['X_test'],
base_workspace['y_test'], **params)
return error