random_features_helper.py 文件源码

def run_solver(self, unit, n_units, arm):
        start_time=time.time()
        #kernel_map=dict(zip([1,2,3],['rbf','poly','sigmoid']))
        preprocess_map=dict(zip([1,2,3,4],['none','min_max','scaled','normalized']))
        self.compute_preprocessor(preprocess_map[arm['preprocessor']])
        # Create random features
        features=kernel_approximation.RBFSampler(gamma=arm['gamma'],n_components=n_units, random_state=1)
        train_features=features.fit_transform(self.data['X_train'])
        val_features=features.transform(self.data['X_val'])
        test_features=features.transform(self.data['X_test'])
        approx_time=(time.time()-start_time)/60.0
        print 'approximating kernel took %r' % approx_time

        clf = linear_model.RidgeClassifier(alpha=1.0/(arm['C']*n_units),solver='lsqr',copy_X=False)
        clf.fit(train_features, self.data['y_train'])
        print 'fitting model took %r' % ((time.time()-start_time)/60.0 - approx_time)
        # Validate this hyperparameter configuration on the full validation data
        #y_loss = 1 - clf.score(self.data['X_train'], self.data['y_train'])
        y_loss=1
        test_acc=0
        val_acc= clf.score(val_features, self.data['y_val'])
        test_acc = clf.score(test_features, self.data['y_test'])
        del self.data
        del train_features, val_features, test_features
        gc.collect()

        return y_loss,val_acc,test_acc