#!/usr/bin/env python
# coding: utf-8

# # Feature extraction with tsfresh transformer
# 
# [tsfresh](https://tsfresh.readthedocs.io) is a tool for extacting summary features
# from a collection of time series. It is an unsupervised transformation, and as such
# can easily be used as a pipeline stage in classification, clustering and regression
# in conjunction with a scikit-learn compatible estimator.
# 
# ## Preliminaries
# You have to install tsfresh if you haven't already. To install it, uncomment the cell below:

# In[1]:


# !pip install --upgrade tsfresh


# In[2]:


from sklearn.ensemble import RandomForestClassifier
from sklearn.pipeline import make_pipeline

from aeon.datasets import load_arrow_head, load_basic_motions
from aeon.transformations.collection.feature_based import TSFresh, TSFreshRelevant


# ## Example data set
# 
# We use the ArrowHead data from the [UCR TSC archive](https://timeseriesclassification.com).
# as an example dataset. See
# [dataset notebook](https://github.com/aeon-toolkit/aeon/blob/main/examples/datasets/provided_data.ipynb) for more details. We only use the first few cases for examples to speed up the 
# notebook. 

# In[3]:


X, y = load_arrow_head()
n_cases = 24
X_train = X[:n_cases, :, :]
y_train = y[:n_cases]
X_test = X[n_cases : 2 * n_cases, :, :]
y_test = y[n_cases : 2 * n_cases]
print(X_train.shape, y_train.shape, X_test.shape, y_test.shape)


# ## Using tsfresh to extract features
# 
# There are two versions of TSFresh feature extractors wrapped in aeon. The
# first is the unsupervised
# `TSFresh` which by default extracts all 4662 features. See the
# documentation for parameter configuration.

# In[4]:


t = TSFresh()
Xt = t.fit_transform(X_train)
Xt2 = t.transform(X_test)
print(f"Train shape = {Xt.shape} test shape = {Xt2.shape}")


# The second is `TSFreshRelevant` which uses `y` to select the most
# relevant features.

# In[5]:


t = TSFreshRelevant()
t.fit(X_train, y_train)
Xt = t.transform(X_test)
Xt.shape


# ## Using tsfresh with scikit estimators
# 
# You can use the tsfresh transformer with any scikit-learn compatible estimator.
# 

# In[6]:


classifier = make_pipeline(
    TSFresh(default_fc_parameters="efficient", show_warnings=False),
    RandomForestClassifier(),
)
classifier.fit(X_train, y_train)
classifier.score(X_test, y_test)


# For convenience and consistency of use we also have hard coded TSFresh classifier,
# regressor and clusterer.

# In[7]:


from aeon.classification.feature_based import TSFreshClassifier
from aeon.clustering.feature_based import TSFreshClusterer

cls = TSFreshClassifier(relevant_feature_extractor=False)
clst = TSFreshClusterer(n_clusters=2)

cls.fit(X_train, y_train)
cls.score(X_test, y_test)
clst.fit(X_train)
print(cls.predict(X_test))
print(clst.predict(X_test))


# By default, the `TSFreshClassifier` uses the supervised
# `TSFreshRelevant` and the scitkit `RandomForestClassifier`.
#  You can
# change this through the constructor

# In[8]:


from aeon.classification.sklearn import RotationForestClassifier

cls = TSFreshClassifier(estimator=RotationForestClassifier(n_estimators=5))
cls.fit(X_train, y_train)
cls.score(X_test, y_test)


# By default, the `TSFreshClusterer` uses the unsupervised `TSFresh`
# and the `sklearn` clusterer `KMeans` with default parameters (which fits 8 clusters).
#  You can also configure this through the constructor.

# In[9]:


from sklearn.cluster import KMeans

clst = TSFreshClusterer(estimator=KMeans(n_clusters=3))
clst.fit(X_train)
print(clst.predict(X_test))


# The `TSFreshRegressor` uses the supervised
# `TSFreshRelevant` and the scitkit `RandomForestRegressor`.

# In[10]:


from aeon.regression.feature_based import TSFreshRegressor

reg = TSFreshRegressor(relevant_feature_extractor=False)
from aeon.datasets import load_covid_3month

X, y = load_covid_3month(split="train")
reg.fit(X, y)


# 

# ## TSFresh with multivariate time series data
# 
# ``TSFresh`` transformers and all three estimators can be used with multivariate time 
# series. The transform calculates the features on each channel independently then 
# concatenate the results. The full transform creates `777*n_channels` features.

# In[14]:


X_train, y_train = load_basic_motions(split="train")
X_test, y_test = load_basic_motions(split="test")
print(X_train.shape, y_train.shape, X_test.shape, y_test.shape)


# In[15]:


tsfresh = TSFresh()
X = tsfresh.fit_transform(X_train, y_train)
X.shape


# In[ ]: