[MRG] Fix numpy versions problems (#411)

* add new empty line at the end of python file in tslearn/clustering/__init.py

* Empty commit

* Upgrade numpy

* Upgrade numpy with --ignore-installed

* Upgrade numpy with --ignore-installed before scikit-learn install

* Specify numpy version 1.22 for Numba compatibility

* Correct numpy version 1.22 instead of 0.22

* Use numpy version 1.21.6 for Python 3.7 compatibility

* Try to unspecify scikit-learn version

* Empty commit

* Remove commented lines

* Try to install numpy before requirements

* Specify scikit-learn==1.0

* Add --upgrade option for requirements.

* Add eager option for requirements install.

* Install numpy==1.22 for Python 3.8 adn 3.9 only (for linux).

* Remove tests on Python 3.7 for now (on linux)

* Fix error in writing numpy==1.22

* Install scikit-learn==1.0 before reading the requirements

* Back to the main branch of tslearn

* Add numpy.import_array() after cimport numpy

* Replace cython files (.pyx) by usual python files (.py)

* Remove cython from requirements and setup file

* Use numba in python files which where previously written in cython

* Change prange into range for loop with reversed index

* Remove reversed order in loop for numba nopython compatibility

* Correct dtype error

* Remove useless variables and vectorize some for loops

* Define forgotten variable

* Specify float dtype of a matrix

* Add docstrings

* Temporarily comment jit decorator to focus on other sources of code errors.

* Correct indice error in for loop.

* Add necessary for loop.

* Add jit decorators.

* Try to change np.float64 to float in type checks for Python 3.7 compatibility.

* Comment assert dtype == np.float64 to solve Python 3.7 error

* Update macOS-10.15 to macOS-12 in azure-pipelines.yml

* Add input variable types in jit decorator

* Add type signature in git decorators of cycc.py

* Add type signature in jit decorator of functions of cysax.py

* Correct input type in jit decorator

* Correct input type in jit decorator of cydist_sax function

* Try intp input type in jit decorator of function cydist_sax

* Try to use typeof for type inference in jit decorator of function cydist_sax

* Try to use typeof array for type inference in jit decorator of function cydist_sax

* Use typeof array for type inference in jit decorator of function cydist_1d_sax

* Use typeof array for type inference in jit decorator of function inv_transform_sax

* Solve np.linalg.norm problem in cycc.py for use of nopython mode in jit decorated functions

* Use nopython mode in jit decorated functions of cysax.py and soft_dtw_fast.py

* Add parallel option in jit decorators in files cycc.py and cysax.py

* Add parallel option in jit decorator of function _jacobian_product_sq_euc

* Delete useless white space to restart the continuous integration tests

* Add fastmath option in jit decorators

* Simplify arrays in Numba typeof for type induction in jit decorators

* Add parallel option in jit decorators of functions _soft_dtw and _soft_dtw_grad (also removed prange from these functions)

* Use range in decreasing order in function _soft_dtw_grad

* Modify loop structure in funtion cdist_normalized_cc to use Numba prange

* Correct docstring error in normalized_cc function

* Add void as output type in jit decorator of functions _soft_dtw, _soft_dtw_grad and _jacobian_product_sq_euc

* Remove comments in file cycc.py

* Remove comments in file cysax.py

* Remove comments in file soft_dtw_fast.py

* Remove unused imports

* Remove signatures (input and output types) in jit decorators

* Empty commit

* Vectorize a for loop in function inv_transform_sax

* Restore previous code formulation in function gamma_soft_dtw

Author: YannCabanes

azure-pipelines.yml

@@ -26,24 +26,24 @@ jobs:
     inputs:
       versionSpec: '$(python.version)'
     displayName: 'Use Python $(python.version)'
-  
+
   - script: |
       set -xe
       python --version
       python -m pip install --upgrade pip
       python -m pip install -r requirements.txt
     displayName: 'Install dependencies'
-  
+
   - script: |
       set -xe
       python -m pip install -e .
     displayName: 'Install tslearn'
-  
+
   - script: |
       set -xe
       python -m pip install pytest pytest-azurepipelines
       python -m pip install scikit-learn==1.0
-      python -m pytest -v tslearn/ --doctest-modules;
+      python -m pytest -v tslearn/ --doctest-modules
     displayName: 'Test'
 
 

pyproject.toml

@@ -1,2 +1,2 @@
 [build-system]
-requires = ["setuptools", "wheel", "numpy", "Cython"]
+requires = ["setuptools", "wheel", "numpy"]

requirements.txt

@@ -1,4 +1,3 @@
-Cython
 numpy
 numba
 scipy

requirements_build.txt

@@ -1,4 +1,3 @@
-Cython
 numpy
 numba
 scipy

requirements_nocast.txt

@@ -1,4 +1,3 @@
-Cython
 numpy
 numba
 scipy

setup.py

@@ -2,7 +2,6 @@
 from codecs import open
 import numpy
 import os
-from Cython.Build import cythonize
 
 # thanks Pipy for handling markdown now
 ROOT = os.path.abspath(os.path.dirname(__file__))
@@ -33,8 +32,6 @@
             "License :: BSD-2-Clause"
     ],
     include_dirs=[numpy.get_include()],
-    ext_modules=cythonize("tslearn/metrics/*.pyx",
-                          include_path=[numpy.get_include()]),
     packages=find_packages(),
     package_data={"tslearn": [".cached_datasets/Trace.npz"]},
     install_requires=['numpy', 'scipy', 'scikit-learn', 'Cython', 'numba',

tslearn/metrics/cycc.py

@@ -0,0 +1,135 @@
+STUFF_cycc = "cycc"
+
+import numpy as np
+from numba import njit, objmode, prange
+
+__author__ = "Romain Tavenard romain.tavenard[at]univ-rennes2.fr"
+
+
+@njit(parallel=True, fastmath=True)
+def normalized_cc(s1, s2, norm1=-1.0, norm2=-1.0):
+    """Normalize cc.
+
+    Parameters
+    ----------
+    s1 : array-like, shape=[sz, d], dtype=float64
+    s2 : array-like, shape=[sz, d], dtype=float64
+    norm1 : float64, default=-1.0
+    norm2 : float64, default=-1.0
+
+    Returns
+    -------
+    norm_cc : array-like, shape=[2 * sz - 1], dtype=float64
+    """
+    assert s1.shape[1] == s2.shape[1]
+    sz = s1.shape[0]
+    n_bits = 1 + int(np.log2(2 * sz - 1))
+    fft_sz = 2**n_bits
+
+    if norm1 < 0.0:
+        norm1 = np.linalg.norm(s1)
+    if norm2 < 0.0:
+        norm2 = np.linalg.norm(s2)
+
+    denom = norm1 * norm2
+    if denom < 1e-9:  # To avoid NaNs
+        denom = np.inf
+
+    with objmode(cc="float64[:, :]"):
+        cc = np.real(
+            np.fft.ifft(
+                np.fft.fft(s1, fft_sz, axis=0)
+                * np.conj(np.fft.fft(s2, fft_sz, axis=0)),
+                axis=0,
+            )
+        )
+    cc = np.vstack((cc[-(sz - 1) :], cc[:sz]))
+    norm_cc = np.real(cc).sum(axis=-1) / denom
+    return norm_cc
+
+
+@njit(parallel=True, fastmath=True)
+def cdist_normalized_cc(dataset1, dataset2, norms1, norms2, self_similarity):
+    """Compute the distance matrix between two time series dataset.
+
+    Parameters
+    ----------
+    dataset1 : array-like, shape=[n_ts1, sz, d], dtype=float64
+    dataset2 : array-like, shape=[n_ts2, sz, d], dtype=float64
+    norms1 : array-like, shape=[n_ts1], dtype=float64
+    norms2 : array-like, shape=[n_ts2], dtype=float64
+    self_similarity : bool
+
+    Returns
+    -------
+    dists : array-like, shape=[n_ts1, n_ts2], dtype=float64
+    """
+    n_ts1, sz, d = dataset1.shape
+    n_ts2 = dataset2.shape[0]
+    assert d == dataset2.shape[2]
+    dists = np.zeros((n_ts1, n_ts2))
+
+    if (norms1 < 0.0).any():
+        for i_ts1 in prange(n_ts1):
+            norms1[i_ts1] = np.linalg.norm(dataset1[i_ts1, ...])
+    if (norms2 < 0.0).any():
+        for i_ts2 in prange(n_ts2):
+            norms2[i_ts2] = np.linalg.norm(dataset2[i_ts2, ...])
+    if self_similarity:
+        for i in prange(1, n_ts1):
+            for j in range(i):
+                dists[i, j] = normalized_cc(
+                    dataset1[i], dataset2[j], norm1=norms1[i], norm2=norms2[j]
+                ).max()
+        dists += dists.T
+    else:
+        for i in prange(n_ts1):
+            for j in range(n_ts2):
+                dists[i, j] = normalized_cc(
+                    dataset1[i], dataset2[j], norm1=norms1[i], norm2=norms2[j]
+                ).max()
+    return dists
+
+
+@njit(parallel=True, fastmath=True)
+def y_shifted_sbd_vec(ref_ts, dataset, norm_ref, norms_dataset):
+    """Shift a time series dataset w.r.t. a time series of reference.
+
+    Parameters
+    ----------
+    ref_ts : array-like, shape=[sz, d], dtype=float64
+        Time series of reference.
+    dataset : array-like, shape=[n_ts, sz, d], dtype=float64
+        Time series dataset.
+    norm_ref : float64
+    norms_dataset : array-like, shape=[n_ts], dtype=float64
+        Norms of the time series dataset.
+
+    Returns
+    -------
+    dataset_shifted : array-like, shape=[n_ts, sz, d], dtype=float64
+    """
+    n_ts = dataset.shape[0]
+    sz = dataset.shape[1]
+    d = dataset.shape[2]
+    assert sz == ref_ts.shape[0] and d == ref_ts.shape[1]
+    dataset_shifted = np.zeros((n_ts, sz, d))
+
+    if norm_ref < 0:
+        norm_ref = np.linalg.norm(ref_ts)
+    if (norms_dataset < 0.0).any():
+        for i_ts in prange(n_ts):
+            norms_dataset[i_ts] = np.linalg.norm(dataset[i_ts, ...])
+
+    for i in prange(n_ts):
+        cc = normalized_cc(ref_ts, dataset[i], norm1=norm_ref, norm2=norms_dataset[i])
+        idx = np.argmax(cc)
+        shift = idx - sz
+        if shift > 0:
+            dataset_shifted[i, shift:] = dataset[i, :-shift, :]
+        elif shift < 0:
+            dataset_shifted[i, :shift] = dataset[i, -shift:, :]
+        else:
+            dataset_shifted[i] = dataset[i]
+
+    return dataset_shifted