"""
Implementation of the Deep Temporal Clustering model
Dataset loading functions
@author Florent Forest (FlorentF9)
"""
import numpy as np
from tslearn.datasets import UCR_UEA_datasets
from tslearn.preprocessing import TimeSeriesScalerMeanVariance
from sklearn.preprocessing import LabelEncoder
ucr = UCR_UEA_datasets()
# UCR/UEA univariate and multivariate datasets.
all_ucr_datasets = ucr.list_datasets()
def load_ucr(dataset='CBF'):
X_train, y_train, X_test, y_test = ucr.load_dataset(dataset)
X = np.concatenate((X_train, X_test))
y = np.concatenate((y_train, y_test))
if dataset == 'HandMovementDirection': # this one has special labels
y = [yy[0] for yy in y]
y = LabelEncoder().fit_transform(
y) # sometimes labels are strings or start from 1
assert (y.min() == 0) # assert labels are integers and start from 0
# preprocess data (standardization)
X_scaled = TimeSeriesScalerMeanVariance().fit_transform(X)
return X_scaled, y
"""
Implementation of the Deep Temporal Clustering model
Dataset loading functions
"""
import numpy as np
from tslearn.datasets import UCR_UEA_datasets
from tslearn.preprocessing import TimeSeriesScalerMeanVariance
from sklearn.preprocessing import LabelEncoder
ucr = UCR_UEA_datasets()
# try to use UCR/UEA univariate and multivariate datasets.
# requires forked version of tslearn from https://github.com/yichangwang/tslearn
try:
all_ucr_datasets = ucr.list_datasets() + ucr._multivariate_dataset
except AttributeError:
all_ucr_datasets = ucr.list_datasets()
def load_ucr(dataset='CBF'):
X_train, y_train, X_test, y_test = ucr.load_dataset(dataset)
X = np.concatenate((X_train, X_test))
y = np.concatenate((y_train, y_test))
if dataset == 'HandMovementDirection': # this one has special labels
y = [yy[0] for yy in y]
y = LabelEncoder().fit_transform(
y) # sometimes labels are strings or start from 1
assert (y.min() == 0) # assert labels are integers and start from 0
# preprocess data (standardization)
X_scaled = TimeSeriesScalerMeanVariance().fit_transform(X)
class UCRDataset(torch.utils.data.Dataset):
"""
A torch wrapper around tslearn UCR_Datasets datasets
"""
def __init__(self,
name,
partition="train",
ratio=.75,
randomstate=None,
silent=True,
augment_data_noise=0):
r = np.random.RandomState(seed=randomstate)
self.name = name
self.dataset = UCR_UEA_datasets()
self.augment_data_noise = augment_data_noise
if name not in self.dataset.list_datasets():
raise ValueError("Dataset not found: Please choose from " +
", ".join(self.dataset.list_datasets()))
X_trainvalid, y_trainvalid, X_test, y_test = self.dataset.load_dataset(
name)
self.nclasses = len(np.unique(np.append(y_test, y_trainvalid, axis=0)))
self.ndims = 1 # UCR datasets have one featuredimension
train_mask = r.rand(len(X_trainvalid)) < ratio
valid_mask = np.logical_not(train_mask)
if partition == "train":
self.X = X_trainvalid[train_mask]
self.y = y_trainvalid[train_mask]
elif partition == "valid":
self.X = X_trainvalid[valid_mask]
self.y = y_trainvalid[valid_mask]
elif partition == "trainvalid":
self.X = X_trainvalid
self.y = y_trainvalid
elif partition == "test":
self.X = X_test
self.y = y_test
else:
raise ValueError(
"Invalid partition! please provide either 'train','valid', 'trainvalid', or 'test'"
)
# some binary datasets e.g. EGC200 or Lightning 2 have classes: -1, 1 -> clipping to 1:2
if self.y.min() < 0:
if not silent:
print("Found class ids < 0 in dataset. clipping to zero!")
self.y = np.clip(self.y, 0, None)
# some datasets (e.g. Coffee) have classes with zero index while all other start with 1...
if self.y.min() > 0:
if not silent:
print(
"Found class id starting from 1. reducing all class ids by one to start from zero"
)
self.y -= 1
#self.classes = np.unique(self.y)
self.sequencelength = X_trainvalid.shape[1]
if not silent:
msg = "Loaded dataset {}-{} T={}, classes={}: {}/{} samples"
print(
msg.format(name, partition, self.sequencelength, self.nclasses,
len(self.X),
len(X_trainvalid) + len(X_test)))
def __len__(self):
return self.X.shape[0]
def __getitem__(self, idx):
X = self.X[idx]
X += np.random.rand(*X.shape) * self.augment_data_noise
X = torch.from_numpy(X).type(torch.FloatTensor)
y = torch.from_numpy(np.array([self.y[idx]])).type(torch.LongTensor)
# add 1d hight and width dimensions and copy y for each time
return X, y.expand(X.shape[0]), idx
def __str__(self):
str = """
UCR Dataset = {dataset}
X.shape = {Xshape}
y.shape = {yshape}
nclasses = {nclasses}
ndims = {ndims}
""".format(dataset=self.name,
Xshape=self.X.shape,
yshape=self.y.shape,
nclasses=self.nclasses,
ndims=self.ndims)
return str
