Changed parameter eigvals to subset_by_value in scipy.linalg.eigh() f…

…unction. Now support scipy==1.14.0.

.github/workflows/build-code.yml

@@ -67,9 +67,8 @@ jobs:
     - name: Test with pytest
       run: |
         # conda install -yq -c ${CONDA_PREFIX}/conda-bld/ msmbuilder2022
-        pip install pytest==8.0.2
-        conda install scipy=1.13.1
         conda install -yq numdifftools hmmlearn
+        pip install pytest==8.0.2
         mkdir ../../pkgs
         cp -r msmbuilder/tests ../../pkgs
         cd ../../pkgs

msmbuilder/decomposition/_speigh.pyx

@@ -124,7 +124,7 @@ def speigh(double[:, ::1] A, double[:, ::1] B, double rho, double eps=1e-6,
     cdef double[::1] b = np.empty(N)
     # Initialize solver from dominant generalized eigenvector (unregularized
     # solution)
-    x = scipy.linalg.eigh(A, B, eigvals=(N-1, N-1))[1][:,0]
+    x = scipy.linalg.eigh(A, B, subset_by_index=[N-1, N-1])[1][:,0]
 
     cdef double[::1] B_eigvals
     cdef double[:, ::1] B_eigvecs
@@ -182,7 +182,7 @@ def speigh(double[:, ::1] A, double[:, ::1] B, double rho, double eps=1e-6,
         u = Ak[0,0] / Bk[0,0]
     else:
         gevals, gevecs = scipy.linalg.eigh(
-            Ak, Bk, eigvals=(Ak.shape[0]-1, Ak.shape[0]-1))
+            Ak, Bk, subset_by_index=[Ak.shape[0]-1, Ak.shape[0]-1])
         # Usually slower to use sparse linear algebra here
         # gevals, gevecs = scipy.sparse.linalg.eigsh(
         #     A=Ak, M=Bk, k=1, v0=x[mask], which='LA')

msmbuilder/decomposition/tica.py

@@ -186,7 +186,7 @@ def _solve(self):
             raise RuntimeError('correlation matrix is not symmetric')
 
         vals, vecs = scipy.linalg.eigh(lhs, b=rhs,
-            eigvals=(self.n_features-self.n_components, self.n_features-1))
+            subset_by_index=[self.n_features-self.n_components, self.n_features-1])
 
         # sort in order of decreasing value
         ind = np.argsort(vals)[::-1]