diff --git a/data/README.md b/data/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/results_5reruns/README.md b/results_5reruns/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..7b7192357011922b2db286a1c3af3998d6bc2e96
--- /dev/null
+++ b/results_5reruns/README.md
@@ -0,0 +1 @@
+Folder to store results.
diff --git a/results_stability/README.md b/results_stability/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/section_4.1_compute_visualize.py b/section_4.1_compute_visualize.py
index 3f6dcfd2ea462efd588586c91d129fadd59cac27..08af491fc61827b5756be3b9d94abba8edda5cf8 100644
--- a/section_4.1_compute_visualize.py
+++ b/section_4.1_compute_visualize.py
@@ -5,8 +5,8 @@
import torch
-from P36_Francesco_Emma.utilities.main_function import run_everything, run_cross_validation
-from P36_Francesco_Emma.utilities.hyperparameters import dic_hyperparams
+from utils.main_function import run_everything, run_cross_validation
+from utils.hyperparameters import dic_hyperparams
import numpy as np
from matplotlib import pyplot as plt
@@ -46,7 +46,8 @@ array_eps, array_cv_f, array_cv_f_val, _, list_timings_1L = run_cross_validation
## Store in matlab for beautiful tikzfigure plots
-path_for_results = os.getcwd() + '/P36_Francesco_Emma/paper_experiments/results/'
+path_for_results = os.path.abspath(os.path.join(os.path.dirname(__file__), 'results_5reruns/'))
+
os.makedirs(path_for_results, exist_ok=True)
io.savemat(path_for_results + name_dataset + '.mat',
dict(array_concatenate=array_concatenate,
diff --git a/section_4.2_compute.py b/section_4.2_compute.py
index 150836e0c5d5269a0b738ca14a38f5db5b4385f3..c60aa4a4bf64bfb9a24b52210598f90a8a62af64 100644
--- a/section_4.2_compute.py
+++ b/section_4.2_compute.py
@@ -4,12 +4,9 @@
# sets", especially to produce the results for the Figures 3-5.
-import torch
-from P36_Francesco_Emma.utilities.main_function import run_everything, run_cross_validation
-from P36_Francesco_Emma.utilities.hyperparameters import dic_hyperparams
+from utils.main_function import run_everything, run_cross_validation
+from utils.hyperparameters import dic_hyperparams
-import numpy as np
-from matplotlib import pyplot as plt
from scipy import io
import os
@@ -54,7 +51,8 @@ for idx_indices in [0, 1, 2, 3, 4]:
## Store in matlab for beautiful tikzfigure plots
- path_for_results = os.getcwd() + '/P36_Francesco_Emma/paper_experiments/results/'
+ path_for_results = os.path.abspath(os.path.join(os.path.dirname(__file__), 'results_5reruns/'))
+
os.makedirs(path_for_results, exist_ok=True)
io.savemat(path_for_results + name_dataset + '_{}'.format(idx_indices) + '.mat',
dict(array_concatenate=array_concatenate,
diff --git a/section_4.2_visualize.py b/section_4.2_visualize.py
index d286d581f736bc5208317f6116b92894a3f1ada9..1b63fa051db2a68ca3ecc169eb7af71323847403 100644
--- a/section_4.2_visualize.py
+++ b/section_4.2_visualize.py
@@ -9,7 +9,6 @@ from matplotlib import pyplot as plt
from scipy import io
import os
import scipy
-import utilities
## Some settings
@@ -18,8 +17,7 @@ list_datasets = ['fried', 'sarcos', 'protein', 'ct', 'diamonds',
'wecs', 'mlr_knn_rng', 'query_agg_count',
'sgemm', 'road_network', 'methane', 'poker'] #, 'susy', 'higgs']
-basepath = utilities.get_basepath()
-path_for_results = basepath + 'P36_Francesco_Emma/paper_experiments/results_5reruns/'
+path_for_results = os.path.abspath(os.path.join(os.path.dirname(__file__), '..', 'results_5reruns/'))
## Initialize dictionaries to store several quantities
diff --git a/section_4.3_compute.py b/section_4.3_compute.py
index 74e6a3bfcc492b8b2fabb9bb16e895032d22911a..eeed37947103120675a0ebc2a951caaafaeb76ac 100644
--- a/section_4.3_compute.py
+++ b/section_4.3_compute.py
@@ -3,25 +3,21 @@
# Code related to the numerical experiment within Section 4.3. "Stability of the kernel
# optimization", especially to produce the results for Figure 6.
-# Similar to 03_4_stability_investigations.py, but now also running the VKOGA after the kernel optimization.
-# I ran this file on ic6, ic7, ic8 to compute the results which are collected in the folder results_stability.
import os
import numpy as np
from scipy.stats import ortho_group # Requires version 0.18 of scipy
import pickle
-import utilities
-from P36_Francesco_Emma.utilities.dataset_collection import Dataset
-from P36_Francesco_Emma.utilities.hyperparameters import dic_hyperparams
-from P36_Francesco_Emma.utilities.main_function import run_everything
+from utils.dataset_collection import Dataset
+from utils.hyperparameters import dic_hyperparams
+from utils.main_function import run_everything
## Some settings
list_nctrs = [int(np.round(nr)) for nr in np.logspace(np.log(10) / np.log(10), np.log(1000) / np.log(10), 10)]
-basepath = utilities.get_basepath()
-path_for_results = basepath + 'P36_Francesco_Emma/paper_experiments/results_stability/'
+path_for_results = os.path.abspath(os.path.join(os.path.dirname(__file__), 'results_stability/'))
flag_gaussian = False
diff --git a/section_4.3_visualize.py b/section_4.3_visualize.py
index 2d2ce7d26ed5c6bda9807e1e45b0a79f7e12264c..8eb70a6eeaf2bc889ea2c1b3ef4530eb12a925c3 100644
--- a/section_4.3_visualize.py
+++ b/section_4.3_visualize.py
@@ -3,9 +3,6 @@
# Code related to the numerical experiment within Section 4.3. "Stability of the kernel
# optimization", especially to produce the plots for Figure 6.
-# Evaluation and analysis of the results which are obtained from 04_1_stability_investigations.py.
-# Only the wecs dataset is a bit weird: Here I do not understand what is going on:
-# Only the first singular vectors are aligned, the other ones are not!
import os
from scipy import io
@@ -14,14 +11,12 @@ from scipy.stats import ortho_group # Requires version 0.18 of scipy
from matplotlib import pyplot as plt
import pickle
import scipy
-import utilities
## Some settings
list_nctrs = [int(np.round(nr)) for nr in np.logspace(np.log(10) / np.log(10), np.log(1000) / np.log(10), 10)]
-basepath = utilities.get_basepath()
-path_for_results = basepath + 'P36_Francesco_Emma/paper_experiments/results_stability/'
+path_for_results = os.path.abspath(os.path.join(os.path.dirname(__file__), 'results_stability/'))
## First, collect results from all files in a common dictionary
@@ -167,81 +162,3 @@ for idx_dataset, name_dataset in enumerate(list(dic_results.keys())):
break
-
-# =============================================================================
-# ============ The following code is meant for some investigations ============
-# =============================================================================
-
-
-
-## Single plot: Compute the distances between the largest eigenspaces. We use nfolds = 64 as reference!
-max_dim = 15 # maximum dimension to which we want to check similarity
-name_dataset = 'wecs'
-
-nfold_ref = 64 # we will use this as a reference model!
-list_idx = [0, 1, 2, 3, 4]
-
-for idx_index in list_idx:
-
- plt.figure(110 + idx_index)
- plt.clf()
- list_legend = []
-
- for nfold in dic_results[name_dataset][idx_index]:
- if nfold == nfold_ref:
- continue # this is our reference model
-
- list_similarity = []
- for dim in range(max_dim):
- array_angles = scipy.linalg.subspace_angles(dic_singvecs[name_dataset][idx_index][nfold][:, :(dim + 1)],
- dic_singvecs[name_dataset][idx_index][nfold_ref][:, :(dim + 1)])
-
- # Maybe modify this measure here? 90 degress means orthogonality!!
- similarity = array_angles[0]
-
- list_similarity.append(np.rad2deg(similarity))
-
- plt.plot(list(range(max_dim)), list_similarity)
- list_legend.append('{} vs {}'.format(nfold_ref, nfold))
-
- # energy_captured = np.sum(np.abs(dic_singvals[idx_index][idx_nfold][-10:])) / np.sum(np.abs(dic_singvals[idx_index][idx_nfold]))
- # plt.title(name_dataset + ': energy captured within {} dims: {:.3f}%'.format(max_dim, 100*energy_captured))
- plt.title(name_dataset + '(dim={}): Alignment of subspaces'.format(dic_singvals[name_dataset][idx_index][nfold].shape[0]))
- plt.legend(list_legend)
-
-
-## Single plot: Plot the first few singular vectors
-dic_colors = {0: 'k', 4: 'm', 8: 'g', 16: 'r', 32: 'b', 64: 'y'}
-max_vectors = 3 # maximum dimension to which we want to check similarity
-name_dataset = 'wecs'
-
-list_idx = [0, 1, 2, 3, 4]
-
-for idx_vector in range(max_vectors):
-
- plt.figure(1000 + idx_vector)
- plt.clf()
- list_legend = []
-
- flag_legend = True
- for idx_index in list_idx:
-
- for nfold in dic_results[name_dataset][idx_index]:
-
- if flag_legend:
- list_legend.append(nfold)
-
- # ToDo: Flip if necessary
- plt.plot(dic_singvecs[name_dataset][idx_index][nfold][:, idx_vector], color=dic_colors[nfold])
-
-
- flag_legend = False
-
-
- # energy_captured = np.sum(np.abs(dic_singvals[idx_index][idx_nfold][-10:])) / np.sum(np.abs(dic_singvals[idx_index][idx_nfold]))
- # plt.title(name_dataset + ': energy captured within {} dims: {:.3f}%'.format(max_dim, 100*energy_captured))
- plt.title(name_dataset + '(dim={}): Plot of singular vectors'.format(dic_singvals[name_dataset][idx_index][nfold].shape[0]))
- plt.legend(list_legend)
-
-
-
diff --git a/utils/dataset_collection.py b/utils/dataset_collection.py
index 58a1fab991874b512e135ae841e90d794d0f1361..65b849cbb6abc3988fc5ef1bc92bff7890bbcec3 100644
--- a/utils/dataset_collection.py
+++ b/utils/dataset_collection.py
@@ -1,11 +1,5 @@
import numpy as np
-import pandas as pd
import os
-import h5py
-import math
-import scipy.io as spio
-from P32_just_interpolate.utilities.utils import load_mnist_pair, get_basepath
-import utilities
class Dataset():
"""
@@ -19,27 +13,10 @@ class Dataset():
self.y = None
self.dic_dataset = {
- 'example_2d': (lambda x: .02 * (x[:, [0]] + x[:, [1]]) ** 2 + np.sin(2 * math.pi * (x[:, [0]] - x[:, [1]])), 2),
- 'example_2d_tiz': (lambda x: x[:, [0]] + (x[:, [0]] + .1 * x[:, [1]]) ** 2, 2),
- 'example_2d_radial': (lambda x: np.exp(-4 * np.sum((x - .5 * np.ones_like(x)) ** 2, axis=1, keepdims=True)), 2),
- 'example_2d_active_1': (lambda x: np.abs(x[:, [0]] - 2*x[:, [1]]), 2),
- 'example_2d_franke': (lambda x: 0.75 * np.exp(-(9 * x[:, [0]] - 2) ** 2 / 4 - (9 * x[:, [1]]- 2) ** 2 / 4)
- + 0.75 * np.exp(-(9 * x[:, [0]] + 1) ** 2 / 49 - (9 * x[:, [1]] + 1) / 10)
- + 0.5 * np.exp(-(9 * x[:, [0]] - 7) ** 2 / 4 - (9 * x[:, [1]] - 3) ** 2 / 4)
- - 0.2 * np.exp(-(9 * x[:, [0]] - 4) ** 2 - (9 * x[:, [1]] - 7) ** 2), 2),
- 'example_3d_active_1': (lambda x: x[:, [0]] + 2 * x[:, [1]], 3),
- 'example_3d_active_2': (lambda x: (x[:, [0]] + x[:, [1]]) ** 2 + np.sin(2 * math.pi * (x[:, [0]] - x[:, [1]])), 3),
- 'example_10d_active': (lambda x: x[:, [0]] * (x[:, [1]] - x[:, [2]]) ** 3 + 2 * np.sin(math.pi * (x[:, [1]] - x[:, [2]])) - np.exp(-2 * x[:, [2]]), 10),
- 'example_10d_vanishing': (lambda x: x @ ((np.arange(10) + 1.0) ** (-2)).reshape(-1, 1), 10),
- 'example_10d_radial': (lambda x: np.exp(-4 * np.sum((x[:, :5] - .5 * np.ones_like(x[:, :5])) ** 2, axis=1, keepdims=True)), 10),
- 'example_5d_radial': (lambda x: np.exp(-4 * np.sum((x[:, :] - .5 * np.ones_like(x[:, :])) ** 2, axis=1, keepdims=True)), 5),
- 'example_5d_active': (lambda x: (x[:, [1]] - x[:, [2]]), 5),
'example_5d_faster_conv': (lambda x: np.exp(-4 * np.sum(x[:, :] - .5 * np.ones_like(x[:, :]), axis=1, keepdims=True) ** 2), 5),
'example_6d_kink': (lambda x: (np.exp(-4 * np.sum((x[:, :] - .5 * np.ones_like(x[:, :])) ** 2, axis=1, keepdims=True)) + 2 * np.abs(x[:, [0]] - .5)), 6),
'example_7d_semiactive': (lambda x: np.exp(-4 * np.sum((x[:, :] - .5 * np.ones_like(x[:, :])) ** 2, axis=1, keepdims=True))
+ np.exp(-9 * np.sum((x[:, :2] - .3 * np.ones_like(x[:, :2])) ** 2, axis=1, keepdims=True)), 7),
- # 'example_winkle': (lambda x: ((np.sin(x[:, [0]]) * np.exp(x[:, [1]]) + x[:, [2]]) ** ((x[:, [3]] + 1) ** 2))
- # / ((x[:, [3]] + 1) ** 2), 5),
}
@@ -52,141 +29,24 @@ class Dataset():
X = np.random.rand(self.N_points, dim)
y = fcn(X)
-
- elif name_dataset == 'example_2d_clusters':
- X, y = self.example_2d_clusters()
- elif name_dataset[:24] == 'example_highdim_clusters':
- dim = int(name_dataset.split('_')[-1])
- X, y = self.example_highdim_clusters(dim)
- elif name_dataset[:10] == 'toyexample':
- X, y = self.toyexample(name_dataset)
+ elif name_dataset in ['ct', 'diamonds', 'fried', 'kegg_undir_uci',
+ 'methane', 'mlr_knn_rng', 'online_video', 'poker',
+ 'protein', 'query_agg_count', 'road_network', 'sarcos', 'sgemm',
+ 'stock', 'wecs']:
+ X, y = self.example_holzmuller(name_dataset)
else:
- if name_dataset[:13] == 'example_mnist':
- X, y = self.example_mnist(name_dataset)
- elif name_dataset == 'example_TUD':
- X, y = self.example_TUD()
- elif name_dataset in ['ct', 'diamonds', 'fried', 'kegg_undir_uci',
- 'methane', 'mlr_knn_rng', 'online_video', 'poker',
- 'protein', 'query_agg_count', 'road_network', 'sarcos', 'sgemm',
- 'stock', 'wecs']:
- X, y = self.example_holzmuller(name_dataset)
- elif name_dataset in ['flights', 'higgs', 'susy', 'taxi', 'timit']:
- X, y = self.falkon_datasets(name_dataset)
- elif name_dataset in ['uci_airfoil_self_noise', 'uci_CCPP']:
- X, y = self.example_uci(name_dataset)
+ pass
assert X is not None, 'Bug in get_data!'
assert y is not None, 'Bug in get_data!'
return X, y
- def example_2d_clusters(self):
-
- X1 = np.random.randn(100, 2)
- X2 = .1 * np.random.randn(500, 2) + np.array([.5, .5])
- X3 = .2 * np.random.randn(500, 2) + np.array([-.5, -.5])
-
- f_func = lambda x: 1 / (1 + np.abs(x[:, 0] - .5))
-
- X = np.concatenate([X1, X2, X3], axis=0)
- y = f_func(X)
-
- return X, y
-
- def example_highdim_clusters(self, dim):
- # In the meeting with Antoine and Giacomo we cam up with a hypothesis when greedy works better.
-
- # X1 = np.random.rand(1000, dim)
- # X2 = .1 * np.random.rand(5000, dim) + .5 * np.zeros((1, dim))
- # X3 = .2 * np.random.rand(5000, dim) - .5 * np.zeros((1, dim))
- # X = np.concatenate([X1, X2, X3], axis=0)
- #
- # y = np.ones((X.shape[0], 1))
- # y[X[:, 0] > .5] = -1
-
- N_points = self.N_points
-
- # Example 1: Peak at center, then decay. randn distribution.
- # X = np.random.randn(N_points, dim)
- # center = .5 * np.ones((1, dim))
- # y = 1 / (1 + 5 * np.linalg.norm(X - center, axis=1, keepdims=True))
- # y = y + 1e-3 * np.random.randn(y.shape[0], 1)
-
-
- # Example2: Peak at center, then decay. sparse randn distribution and two clusters.
- X1 = np.random.randn(int(.1*N_points), dim)
- X2 = .1 * np.random.rand(int(.45*N_points), dim) + .5 * np.ones((1, dim))
- X3 = .2 * np.random.rand(int(.45*N_points), dim) - .5 * np.ones((1, dim))
- X = np.concatenate([X1, X2, X3], axis=0)
-
- center = .5 * np.ones((1, dim))
- y = 1 / (1 + 5 * np.linalg.norm(X - center, axis=1, keepdims=True))
- y = y + 1e-3 * np.random.randn(y.shape[0], 1)
-
- return X, y
-
-
- def example_mnist(self, name_dataset):
- """MNIST dataset"""
-
- X_train, X_test, y_train, y_test = \
- load_mnist_pair(get_basepath(),
- [int(name_dataset[-2]), int(name_dataset[-1])])
-
- X = np.concatenate([X_train, X_test], axis=0)
- y = np.concatenate([y_train, y_test], axis=0).reshape(-1, 1)
-
- return X, y
-
- def example_uci(self, name_dataset):
- """Some UCI datasets"""
-
- # path = '/usr/local/home/wenzeltn/deepkernel-pytorch2/data/UCI/'
- path = '/home/wenzeltn/local_home/deepkernel-pytorch2/data/UCI/'
-
- if name_dataset == 'uci_airfoil_self_noise':
- data = np.loadtxt(path + 'airfoil_self_noise.dat', unpack=True).transpose()
-
- X = data[:, :5]
- y = data[:, [5]]
-
- X = (X - X.mean(axis=0, keepdims=True)) / (X.std(axis=0, keepdims=True) + 1e-30)
- y = (y - y.mean(axis=0, keepdims=True)) / (y.std(axis=0, keepdims=True) + 1e-30)
-
- if name_dataset == 'uci_CCPP':
- import pandas as pd
- data = pd.read_excel(path + 'CCPP/' + 'Folds5x2_pp.xlsx').to_numpy()
-
- X = data[:, :4]
- y = data[:, [4]]
-
- X = (X - X.mean(axis=0, keepdims=True)) / (X.std(axis=0, keepdims=True) + 1e-30)
- y = (y - y.mean(axis=0, keepdims=True)) / (y.std(axis=0, keepdims=True) + 1e-30)
-
- return X, y
-
- def example_TUD(self):
- """Felix Döppel data"""
-
- path = '/home/wenzeltn/local_home/deepkernel-pytorch2/P01_DK_optimization/data/'
- # path = '/usr/local/home/wenzeltn/deepkernel-pytorch2/P01_DK_optimization/data/'
- mat_train = spio.loadmat(path + 'DataTUDKinetics.mat', squeeze_me=True)
- # mat_test = spio.loadmat(path + 'TestTUDKinetics.mat', squeeze_me=True)
-
- X = mat_train['input']
- y = mat_train['output'].reshape(-1, 1)
-
- return X, y
def example_holzmuller(self, ds_name):
"""Data set from David Holzmüller paper of batch active learning."""
- if os.path.exists('/usr/local/home/wenzeltn'): # anm03
- path = '/usr/local/home/wenzeltn/deepkernel-pytorch2/data/data_holzmuller/'
- elif os.path.exists('/usr/local/storage/wenzeltn'): # ianscluster
- path = '/usr/local/storage/wenzeltn/deepkernel-pytorch2/data/data_holzmuller/'
- elif os.path.exists('/home/wenzeltn/'): # laptop
- path = '/home/wenzeltn/local_home/deepkernel-pytorch2/data/data_holzmuller/'
+ path = os.path.abspath(os.path.join(os.path.dirname(__file__), '..', 'data/'))
X = np.load(path + ds_name + '/X.npy')
y = np.load(path + ds_name + '/y.npy')
@@ -199,93 +59,3 @@ class Dataset():
return X, y
- def toyexample(self, ds_name):
- """Toy datasets from Tizian created on 04.01.23."""
-
- basepath = utilities.get_basepath()
- path = basepath + 'data/TOY_selfcreated/'
-
- X = np.load(path + ds_name + 'X.npy')
- y = np.load(path + ds_name + 'y.npy')
-
- # Standardization of input according to David H (reduce effect of outliers)
- X = (X - X.mean(axis=0, keepdims=True)) / (X.std(axis=0, keepdims=True) + 1e-30)
- X = 5 * np.tanh(0.2 * X)
-
- y = (y - y.mean(axis=0, keepdims=True)) / (y.std(axis=0, keepdims=True) + 1e-30)
-
- return X, y
-
-
- def falkon_datasets(self, ds_name):
- """Datasets from FALKON paper."""
-
- path = None
- if os.path.exists('/usr/local/home/wenzeltn'): # anm03
- path = '/usr/local/home/wenzeltn/deepkernel-pytorch2/data/FALKON_data/'
- elif os.path.exists('/usr/local/storage/wenzeltn'): # ianscluster
- path = '/usr/local/storage/wenzeltn/deepkernel-pytorch2/data/FALKON_data/'
- elif os.path.exists('/home/wenzeltn/'):
- pass
-
- filename = None
- if ds_name == 'susy' or ds_name == 'higgs':
- if ds_name == 'susy':
- filename = 'Susy.mat'
- else:
- filename = 'Higgs.mat'
-
- with h5py.File(path + filename, "r") as h5py_file:
- arr = np.asarray(h5py_file['X'], dtype=np.float32).T
- X = arr[:, 1:]
- y = arr[:, 0].reshape(-1, 1)
-
- # Preprocess input
- mtr = np.mean(X, axis=0, dtype=np.float64, keepdims=True).astype(X.dtype)
- vtr = np.var(X, axis=0, dtype=np.float64, ddof=1, keepdims=True).astype(X.dtype)
-
- X -= mtr
- X /= vtr
-
- # Preprocess outputs
- y = y * 2 - 1
-
-
- elif ds_name == 'timit': # num_train_samples = 1124823
- pass
- # ToDo: Does not work, dataset probably damaged
-
- # filename = 'TIMIT.mat'
- #
- # with h5py.File(path + filename, 'r') as h5py_file:
- # dtype = np.float32
- # Xtr = np.array(h5py_file['Xtr'], dtype=dtype)
- # Xts = np.array(h5py_file['Xts'], dtype=dtype)
- # Ytr = np.array(h5py_file['Ytr'], dtype=dtype).reshape((-1,))
- # Yts = np.array(h5py_file['Yts'], dtype=dtype).reshape((-1,))
- # X = np.concatenate((Xtr, Xts), axis=0)
- # y = np.concatenate((Ytr, Yts), axis=0)
- #
- #
- # #
- # # f = spio.loadmat(path + filename)
- # # dtype = np.float32
- # # Xtr = np.array(f['Xtr'], dtype=dtype)
- # # Xts = np.array(f['Xts'], dtype=dtype)
- # # Ytr = np.array(f['Ytr'], dtype=dtype).reshape((-1,))
- # # Yts = np.array(f['Yts'], dtype=dtype).reshape((-1,))
- # # X = np.concatenate((Xtr, Xts), axis=0)
- # # y = np.concatenate((Ytr, Yts), axis=0)
-
- # elif ds_name == 'flights':
- # filename = 'flights.csv'
- # # Not implemented, because I have a .csv file, Giacomo uses an .hdf5 file
- #
- # elif ds_name == 'taxi':
- # pass
- # # Not implemented because super large
-
- # ToDo: Not implemented: Taxi since very large, flights only in .csv, timit is broken
-
- return X, y
-
diff --git a/utils/hyperparameters.py b/utils/hyperparameters.py
index 02469eea921fc1cbb321ea65b2946e26ffc01f2e..9cd1ca153e3c915ba8ed3e2710c944eb7d3a83c6 100644
--- a/utils/hyperparameters.py
+++ b/utils/hyperparameters.py
@@ -22,54 +22,6 @@ class example_func_approx():
k_matern = 0
-class example_2d(example_func_approx):
- pass
-
-class example_10d_vanishing(example_func_approx):
- flag_gaussian = True
-
-# class example_toy():
-# maxIter_vkoga = 250
-# N_points = 10000
-# noise_level = 1e-3
-#
-# reg_para_optim = 1e-5 # for kernel optimization
-# reg_para_vkoga = 0 # for running VKOGA
-# learning_rate = 5e-3
-# n_epochs = 25
-# batch_size = 64
-# n_folds = None
-#
-# flag_initialize_diagonal = True
-# flag_symmetric = False
-# flag_gaussian = False
-#
-# n_cross_val = 10
-# shape_para = 1
-# k_matern = 0
-
-
-class example_TUD():
- maxIter_vkoga = 500
- N_points = None
- noise_level = 0
-
- reg_para_optim = 1e-5 # for kernel optimization
- reg_para_vkoga = 0 # for running VKOGA
- learning_rate = 5e-3
- n_epochs = 50
- batch_size = 64
- n_folds = None
-
- flag_initialize_diagonal = True
- flag_symmetric = False
- flag_gaussian = True
-
- n_cross_val = 10
- shape_para = 1
- k_matern = 0
-
-
class example_holzmuller():
maxIter_vkoga = 1000
N_points = None
@@ -90,48 +42,15 @@ class example_holzmuller():
shape_para = 1 / 5
k_matern = 0
+
class example_holzmuller_few_epochs(example_holzmuller):
n_epochs = 10
-
-class example_mnist(example_holzmuller):
- n_epochs = 5
- maxIter_vkoga = 200
-
-class example_airfoil(example_holzmuller):
- n_epochs = 25
- learning_rate = 2e-2
- maxIter_vkoga = 500
- k_matern = 0
-
-class example_CCPP(example_holzmuller):
- n_epochs = 25
- learning_rate = 2e-2
- maxIter_vkoga = 500
- k_matern = 0
-
-
-
dic_hyperparams = {
- 'example_2d': example_func_approx(),
- 'example_2d_tiz': example_func_approx(),
- 'example_2d_radial': example_func_approx(),
- 'example_2d_active_1': example_func_approx(),
- 'example_2d_franke': example_func_approx(),
- 'example_3d_active_1': example_func_approx(),
- 'example_3d_active_2': example_func_approx(),
- 'example_10d_active': example_func_approx(),
- 'example_10d_vanishing': example_func_approx(),
- 'example_10d_radial': example_func_approx(),
- 'example_5d_radial': example_func_approx(),
- 'example_5d_active': example_func_approx(),
'example_5d_faster_conv': example_func_approx(),
'example_6d_kink': example_func_approx(),
'example_7d_semiactive': example_func_approx(),
- 'example_winkle': example_func_approx(),
- 'example_TUD': example_TUD(),
-
'ct': example_holzmuller_few_epochs(),
'diamonds': example_holzmuller_few_epochs(),
@@ -147,17 +66,6 @@ dic_hyperparams = {
'sarcos': example_holzmuller_few_epochs(),
'sgemm': example_holzmuller_few_epochs(),
'stock': example_holzmuller_few_epochs(),
- 'wecs': example_holzmuller_few_epochs(),
-
- 'susy': example_holzmuller_few_epochs(),
- 'higgs': example_holzmuller_few_epochs(),
-
-
- 'example_mnist_01': example_mnist(),
- 'example_mnist_34': example_mnist(),
-
- 'uci_airfoil_self_noise': example_airfoil(),
- 'uci_CCPP': example_CCPP(),
-
+ 'wecs': example_holzmuller_few_epochs()
}
diff --git a/utils/kernels.py b/utils/kernels.py
index 0d223479acc1118563f9635d7d15a996605645f6..fd34de5687ecab26ab73fdb46c20166e7f519147 100644
--- a/utils/kernels.py
+++ b/utils/kernels.py
@@ -5,6 +5,7 @@ from scipy.spatial import distance_matrix
import numpy as np
import matplotlib.pyplot as plt
+
# Abstract kernel
class Kernel(ABC):
@abstractmethod
diff --git a/utils/main_function.py b/utils/main_function.py
index ea871ee2a46929fee066d0d6c43bba30f3d75333..79f5010227d41cff7f7bad286c1291548a113dd5 100644
--- a/utils/main_function.py
+++ b/utils/main_function.py
@@ -1,29 +1,25 @@
-from P36_Francesco_Emma.models.optimized_kernel import OptimizedKernel
-from P36_Francesco_Emma.utilities.dataset_collection import Dataset
-from P36_Francesco_Emma.utilities.hyperparameters import dic_hyperparams
+from utils.optimized_kernel import OptimizedKernel
+from utils.dataset_collection import Dataset
+from utils.hyperparameters import dic_hyperparams
-from vkoga import tkernels, kernels
+from utils import tkernels, kernels
import torch
from matplotlib import pyplot as plt
import numpy as np
import time
-from sklearn.model_selection import train_test_split
-from vkoga.kernels import Matern
-from vkoga.vkoga import VKOGA
+from utils.vkoga import VKOGA
import os
from datetime import datetime
-import utilities
## Some settings
list_nctrs = [int(np.round(nr)) for nr in np.logspace(np.log(10) / np.log(10), np.log(1000) / np.log(10), 10)]
-basepath = utilities.get_basepath()
-path_for_indices = basepath + 'data/'
+path_for_indices = os.path.abspath(os.path.join(os.path.dirname( __file__ ), '..', 'data/'))
# Main functiono to run kernel optimization with subsequent vkoga
diff --git a/utils/optimized_kernel.py b/utils/optimized_kernel.py
index 423faf4ed875e5e78c1e508ec15c36d58d26f56d..fb3aae3ba8d7d965c9912dcded4149313f82a317 100644
--- a/utils/optimized_kernel.py
+++ b/utils/optimized_kernel.py
@@ -1,6 +1,6 @@
from torch import nn
import torch
-from P36_Francesco_Emma.utilities.cv_rippa_ext import compute_cv_loss_via_rippa_ext_2
+from utils.cv_rippa_ext import compute_cv_loss_via_rippa_ext_2
import numpy as np
diff --git a/utils/vkoga.py b/utils/vkoga.py
index 1731605560da4189f8cb88ab5961f5bf51742a09..5e7fd4bd6c5b6b3a59f24c5d723f56c2fd87dfaa 100644
--- a/utils/vkoga.py
+++ b/utils/vkoga.py
@@ -1,6 +1,6 @@
#!/usr/bin/env python3
-from vkoga.kernels import Gaussian
+from utils.kernels import Gaussian
import numpy as np
from sklearn.base import BaseEstimator
from sklearn.utils.validation import check_X_y, check_array, check_is_fitted