def test_seed_reset(self): spa = SPA(self.benchmark, self.models, reps=10) spa.seed(23456) initial_state = spa.bootstrap.random_state assert_equal(spa.bootstrap._seed, 23456) spa.compute() spa.reset() assert_equal(spa._pvalues, None) assert_equal(spa.bootstrap.random_state, initial_state)
def test_seed_reset(self): spa = SPA(self.benchmark, self.models, reps=10) spa.seed(23456) initial_state = spa.bootstrap.random_state assert_equal(spa.bootstrap._seed, 23456) spa.compute() spa.reset() assert_equal(spa._pvalues, None) assert_equal(spa.bootstrap.random_state, initial_state)
def test_variances_and_selection(self): adj_models = self.models + linspace(-2, 0.5, self.k) spa = SPA(self.benchmark, adj_models, block_size=10, reps=10) spa.seed(23456) spa.compute() variances = spa._loss_diff_var loss_diffs = spa._loss_diff demeaned = spa._loss_diff - loss_diffs.mean(0) t = loss_diffs.shape[0] kernel_weights = np.zeros(t) p = 1 / 10.0 for i in range(1, t): kernel_weights[i] = ((1.0 - (i / t)) * ((1 - p) ** i)) + \ ((i / t) * ((1 - p) ** (t - i))) direct_vars = (demeaned**2).sum(0) / t for i in range(1, t): direct_vars += 2 * kernel_weights[i] * \ (demeaned[:t - i, :] * demeaned[i:, :]).sum(0) / t assert_allclose(direct_vars, variances) selection_criteria = -1.0 * np.sqrt( (direct_vars / t) * 2 * np.log(np.log(t))) valid = loss_diffs.mean(0) >= selection_criteria assert_equal(valid, spa._valid_columns) # Bootstrap variances spa = SPA(self.benchmark, self.models, block_size=10, reps=100, nested=True) spa.seed(23456) spa.compute() spa.reset() bs = spa.bootstrap.clone(demeaned) variances = spa._loss_diff_var bootstrap_variances = t * bs.var( lambda x: x.mean(0), reps=100, recenter=True) assert_allclose(bootstrap_variances, variances)
def test_variances_and_selection(self): adj_models = self.models + linspace(-2, 0.5, self.k) spa = SPA(self.benchmark, adj_models, block_size=10, reps=10) spa.seed(23456) spa.compute() variances = spa._loss_diff_var loss_diffs = spa._loss_diff demeaned = spa._loss_diff - loss_diffs.mean(0) t = loss_diffs.shape[0] kernel_weights = np.zeros(t) p = 1 / 10.0 for i in range(1, t): kernel_weights[i] = ((1.0 - (i / t)) * ((1 - p) ** i)) + \ ((i / t) * ((1 - p) ** (t - i))) direct_vars = (demeaned ** 2).sum(0) / t for i in range(1, t): direct_vars += 2 * kernel_weights[i] * \ (demeaned[:t - i, :] * demeaned[i:, :]).sum(0) / t assert_allclose(direct_vars, variances) selection_criteria = -1.0 * np.sqrt((direct_vars / t) * 2 * np.log(np.log(t))) valid = loss_diffs.mean(0) >= selection_criteria assert_equal(valid, spa._valid_columns) # Bootstrap variances spa = SPA(self.benchmark, self.models, block_size=10, reps=100, nested=True) spa.seed(23456) spa.compute() spa.reset() bs = spa.bootstrap.clone(demeaned) variances = spa._loss_diff_var bootstrap_variances = t * bs.var(lambda x: x.mean(0), reps=100, recenter=True) assert_allclose(bootstrap_variances, variances)