import os
import yaml
import numpy as np
from sklearn.datasets import make_classification
from random import gauss
from math import ceil, floor
import pandas as pd
import shutil
import h5py
class MultiviewDatasetGenetator():
def __init__(self, n_samples=100, n_views=2, n_classes=2,
Z_factor=2,
R=0,
n_clusters_per_class=1,
class_sep_factor=10,
n_informative_divid=2,
d=4,
D=10,
standard_deviation=2,
weights=None,
flip_y=0.0,
random_state=42, config_path=None):
if config_path is not None:
with open(config_path) as config_file:
args = yaml.safe_load(config_file)
self.__init__(**args)
else:
self.n_samples = n_samples
self.n_views = n_views
self.n_classes = n_classes
self.Z_factor = Z_factor
self.R = R
self.n_clusters_per_class = n_clusters_per_class
self.class_sep_factor = class_sep_factor
self.n_informative_divid = n_informative_divid
self.d = d
self.D = D
self.standard_deviation = standard_deviation
self.weights = weights
self.flip_y = flip_y
self.random_state = random_state
def generate(self):
if self.n_views < 2:
raise ValueError("n_views >= 2")
if self.n_classes < 2:
raise ValueError("n_classes >= 2")
if self.Z_factor < 1:
raise ValueError(
"Z_factor >= 1 pour le bon fonctionnement de l'algorithme")
if (self.R < 0) or (self.R > 1):
raise ValueError("0 <= R <= 1")
if self.n_clusters_per_class < 1:
raise ValueError("n_clusters_per_class >= 1")
if self.class_sep_factor < 0:
raise ValueError("class_sep_factor >= 0")
if self.n_informative_divid < 1:
raise ValueError("n_informative_divid >= 1")
if self.d < 1:
raise ValueError("d >= 1")
if (self.d + self.D) / 2 - 3 * self.standard_deviation < 1:
raise ValueError(
"Il faut que (d+D)/2 - 3*standard_deviation >= 1 pour avoir des valeurs positives non nulles lors de l'emploi de la loi normale")
# n_views dimension of view v values randomly from N((d+D)/2, standard_deviation^2)
d_v = np.random.normal(loc=(self.d + self.D) / 2,
scale=self.standard_deviation,
size=self.n_views)
d_v = list(d_v)
remove_list, add_list = [], []
for dim_view in d_v:
if dim_view < self.d or dim_view > self.D: # 1 <= d <= dim_view <= D
remove_list.append(dim_view)
add = -1
while add < self.d or add > self.D:
add = gauss((self.d + self.D) / 2, self.standard_deviation)
add_list.append(add)
d_v = [view for view in d_v if view not in remove_list] + add_list
d_v = [int(view) for view in d_v] # dimension of views = integer
# d_v = list of views dimension from the highest to the lowest
d_v.sort(reverse=True)
# Dimension of latent space Z (multiplied by Z_factor)
self.dim_Z = self.Z_factor * self.latent_space_dimension(d_v)
# Number of informative features
self.n_informative = round(self.dim_Z / self.n_informative_divid)
# Generation of latent space Z
self.Z, self.y = make_classification(n_samples=self.n_samples, n_features=self.dim_Z,
n_informative=self.n_informative, n_redundant=0,
n_repeated=0, n_classes=self.n_classes,
n_clusters_per_class=self.n_clusters_per_class,
weights=self.weights,
flip_y=self.flip_y,
class_sep=self.n_clusters_per_class * self.class_sep_factor,
random_state=self.random_state, shuffle=False)
I_q = np.arange(self.Z.shape[1])
meta_I_v = []
self.results = []
for view in range(n_views):
# choice d_v[view] numeros of Z columns uniformly from I_q
I_v = np.random.choice(I_q, size=d_v[view],
replace=False) # tirage dans I_q sans remise de taille d_v[view]
meta_I_v += list(I_v)
# projection of Z along the columns in I_v
X_v = self.projection( I_v)
self.results.append((X_v, I_v))
# remove R*d_v[view] columns numeros of I_v form I_q
elements_to_remove = np.random.choice(I_v,
size=floor(self.R * d_v[view]),
replace=False) # tirage dans I_v sans remise de taille floor(R*d_v[view])
I_q = np.setdiff1d(I_q,
elements_to_remove) # I_q less elements from elements_to_remove
self.unsued_dimensions_list = [column for column in I_q if
column not in meta_I_v]
self.unsued_dimensions_percent = round(
(len(self.unsued_dimensions_list) / self.dim_Z) * 100, 2)
def projection(self, chosen_columns_list):
"""
Returns the projection of latent_space on the columns of chosen_columns_list (in chosen_columns_list order)
Parameters:
-----------
chosen_columns_list : list
Returns:
--------
an array of dimension (number of rows of latent_space, length of chosen_columns_list)
"""
return self.Z[:, chosen_columns_list]
def latent_space_dimension(self, views_dimensions_list):
"""
Returns the minimal dimension of latent space (enough to build the dataset) for generator_multiviews_dataset compared to views_dimensions_list
Parameters:
-----------
views_dimensions_list : list
R : float
Returns:
--------
an int
"""
max_view_dimension = max(views_dimensions_list)
dimension = ceil(self.R * sum(views_dimensions_list))
if dimension < max_view_dimension:
dimension = max_view_dimension
reduced_dimension = dimension
remove_sum = 0
for num_view in range(1, len(views_dimensions_list)):
view_prec = views_dimensions_list[num_view - 1]
view_current = views_dimensions_list[num_view]
remove = floor(self.R * view_prec)
remove_sum += remove
if reduced_dimension - remove < view_current:
dimension += view_current - (reduced_dimension - remove)
reduced_dimension = dimension - remove_sum
return dimension
def to_csv(self, saving_path="."):
"""
Create length of multiviews_list + 2 csv files to the indicated path
Files name :
latent_space.csv for latent_space
integer_labels.csv for integer_labels
view0.csv for multiviews_list[0]
Parameters:
-----------
path : str
latent_space : array
integer_labels : 1D array
multiviews_list : list of tuples
Returns:
--------
None
"""
df_latent_space = pd.DataFrame(self.Z)
df_latent_space.to_csv(os.path.join(saving_path, 'latent_space.csv')
, index=False)
df_labels = pd.DataFrame(self.y)
df_labels.to_csv(os.path.join(saving_path, 'integer_labels.csv'),
index=False)
for view_index, view_tuple in enumerate(self.results):
df_view = pd.DataFrame(view_tuple[0], columns=view_tuple[1])
df_view.to_csv(os.path.join(saving_path,
'view'+str(view_index)+'.csv'),
index=False)
def to_hdf5(self, saving_path=".", name="generated_dset"):
dataset_file = h5py.File(os.path.join(saving_path, name+".hdf5"), 'w')
labels_dataset = dataset_file.create_dataset("Labels",
shape=self.y.shape,
data=self.y)
labels_names = ["Label_1", "Label_0"]
labels_dataset.attrs["names"] = [
label_name.encode() if not isinstance(label_name, bytes)
else label_name for label_name in labels_names]
for view_index, (data, feature_indices) in enumerate(self.results):
df_dataset = dataset_file.create_dataset("View" + str(view_index),
shape=data.shape,
data=data)
df_dataset.attrs["sparse"] = False
df_dataset.attrs["name"] = "GeneratedView"+str(view_index)
meta_data_grp = dataset_file.create_group("Metadata")
meta_data_grp.attrs["nbView"] = len(self.results)
meta_data_grp.attrs["nbClass"] = np.unique(self.y)
meta_data_grp.attrs["datasetLength"] = \
self.results[0][0].shape[0]
meta_data_grp.create_dataset("example_ids", data=np.array(
["gen_example_" + str(ex_indx) for ex_indx in
range(self.results[0][0].shape[0])]).astype(
np.dtype("S100")), dtype=np.dtype("S100"))
dataset_file.close()
if __name__=="__main__":
n_samples = 100 # Number of samples in tha dataset
n_views = 4 # Number of views in the dataset
n_classes = 2 # Number of classes in the dataset
Z_factor = 2 # Z dim = latent_space_dim * z_factor
R = 0 # Precentage of non-redundant features in the view
n_clusters_per_class = 1 # Number of clusters for each class
class_sep_factor = 10000 # Separation between the different classes
n_informative_divid = 2 # Divides the number of informative features in the latent space
standard_deviation = 2
d = 4
D = 10
flip_y = 0.00
random_state = 42
weights = None # The proportions of examples in each class
path = "/home/baptiste/Documents/Datasets/Generated/metrics_dset/"
name = "metrics"
if not os.path.exists(path):
os.mkdir(path)
multiview_generator = MultiviewDatasetGenetator(n_samples=n_samples,
n_views=n_views,
n_classes=n_classes,
Z_factor=Z_factor,
R=R,
n_clusters_per_class=n_clusters_per_class,
class_sep_factor=class_sep_factor,
n_informative_divid=n_informative_divid,
d=d,
D=D,
standard_deviation=standard_deviation,
flip_y=flip_y,
weights=weights,
random_state=random_state)
multiview_generator.generate()
multiview_generator.to_hdf5(saving_path=path, name=name)
# for filename in os.listdir(path):
# file_path = os.path.join(path, filename)
# try:
# if os.path.isfile(file_path) or os.path.islink(file_path):
# os.unlink(file_path)
# elif os.path.isdir(file_path):
# shutil.rmtree(file_path)
# except Exception as e:
# print('Failed to delete %s. Reason: %s' % (file_path, e))
# changing_labels_indices = np.random.RandomState(random_state).choice(np.arange(y.shape[0]), n_outliers)
# print(changing_labels_indices)
# y[changing_labels_indices] = np.invert(y[changing_labels_indices].astype(bool)).astype(int)
# results_to_csv(path, Z, y, results)