[feat] Relax strict autograd.numpy requirement (#285)

* [feat] Relax strict autograd.numpy requirement

* [Fix] Re-impose --Werror for python 3.14 runs after new autograd release

.github/workflows/pytest.yml

@@ -42,10 +42,5 @@ jobs:
           uv pip install pytest pytest-cov pytest-benchmark hypothesis --system
           uv pip freeze --system
 
-      - name: Run tests with -Werror
+      - name: Run tests
-        if: matrix.python-version != '3.14'
         run: pytest --cov=pyerrors -vv -Werror
-
-      - name: Run tests without -Werror for python 3.14
-        if: matrix.python-version == '3.14'
-        run: pytest --cov=pyerrors -vv

pyerrors/fits.py

@@ -481,7 +481,7 @@ def least_squares(x, y, func, priors=None, silent=False, **kwargs):
 
     try:
         hess = hessian(chisqfunc)(fitp)
-    except TypeError:
+    except (TypeError, ValueError, np.linalg.LinAlgError):
         raise Exception("It is required to use autograd.numpy instead of numpy within fit functions, see the documentation for details.") from None
 
     len_y = len(y_f)
@@ -722,7 +722,7 @@ def total_least_squares(x, y, func, silent=False, **kwargs):
     fitp = out.beta
     try:
         hess = hessian(odr_chisquare)(np.concatenate((fitp, out.xplusd.ravel())))
-    except TypeError:
+    except (TypeError, ValueError, np.linalg.LinAlgError):
         raise Exception("It is required to use autograd.numpy instead of numpy within fit functions, see the documentation for details.") from None
 
     def odr_chisquare_compact_x(d):

pyerrors/roots.py

@@ -32,10 +32,10 @@ def find_root(d, func, guess=1.0, **kwargs):
     root = scipy.optimize.fsolve(func, guess, d_val)
 
     # Error propagation as detailed in arXiv:1809.01289
-    dx = jacobian(func)(root[0], d_val)
     try:
+        dx = jacobian(func)(root[0], d_val)
         da = jacobian(lambda u, v: func(v, u))(d_val, root[0])
-    except TypeError:
+    except (TypeError, ValueError, np.linalg.LinAlgError):
         raise Exception("It is required to use autograd.numpy instead of numpy within root functions, see the documentation for details.") from None
     deriv = - da / dx
     res = derived_observable(lambda x, **kwargs: (x[0] + np.finfo(np.float64).eps) / (np.array(d).reshape(-1)[0].value + np.finfo(np.float64).eps) * root[0],

tests/fits_test.py

@@ -698,13 +698,27 @@ def test_fit_no_autograd():
         y = a[0] * np.exp(-a[1] * x)
         return y
 
-    with pytest.raises(Exception):
+    def func_autograd(a, x):
-        pe.least_squares(x, oy, func)
+        return a[0] * anp.exp(-a[1] * x)
 
-    pe.least_squares(x, oy, func, num_grad=True)
+    # Since autograd 1.9.0 plain numpy ufuncs are dispatched to the autograd
+    # wrappers (ArrayBox.__array_ufunc__), so a function using numpy.exp now
+    # yields exactly the same result as one using autograd.numpy.exp.
+    for r_np, r_ag in zip(pe.least_squares(x, oy, func), pe.least_squares(x, oy, func_autograd)):
+        assert r_np == r_ag
+    for r_np, r_ag in zip(pe.total_least_squares(oy, oy, func), pe.total_least_squares(oy, oy, func_autograd)):
+        assert r_np == r_ag
+
+    # A function that genuinely cannot be traced by autograd must still raise a
+    # clear error pointing the user to autograd.numpy.
+    def func_invalid(a, x):
+        return np.array(a[0] * np.exp(-a[1] * x), dtype=np.float64)
 
     with pytest.raises(Exception):
-        pe.total_least_squares(oy, oy, func)
+        pe.least_squares(x, oy, func_invalid)
+
+    with pytest.raises(Exception):
+        pe.total_least_squares(oy, oy, func_invalid)
 
 
 def test_invalid_fit_function():
@@ -818,7 +832,14 @@ def test_combined_fit_no_autograd():
     def func_b(a,x):
         return a[0]*np.exp(a[2]*x)
 
+    def func_a_autograd(a,x):
+        return a[0]*anp.exp(a[1]*x)
+
+    def func_b_autograd(a,x):
+        return a[0]*anp.exp(a[2]*x)
+
     funcs = {'a':func_a, 'b':func_b}
+    funcs_autograd = {'a':func_a_autograd, 'b':func_b_autograd}
     xs = {'a':xvals_a, 'b':xvals_b}
     ys = {'a':[pe.Obs([np.random.normal(item, item*1.5, 1000)],['ensemble1']) for item in func_exp1(xvals_a)],
         'b':[pe.Obs([np.random.normal(item, item*1.4, 1000)],['ensemble1']) for item in func_exp2(xvals_b)]}
@@ -826,8 +847,17 @@ def test_combined_fit_no_autograd():
     for key in funcs.keys():
         [item.gamma_method() for item in ys[key]]
 
+    # Since autograd 1.9.0 plain numpy ufuncs are dispatched to the autograd
+    # wrappers, so the fit using numpy now matches the one using autograd.numpy.
+    for r_np, r_ag in zip(pe.least_squares(xs, ys, funcs), pe.least_squares(xs, ys, funcs_autograd)):
+        assert r_np == r_ag
+
+    # A function that genuinely cannot be traced by autograd must still raise.
+    def func_a_invalid(a, x):
+        return np.array(a[0] * np.exp(a[1] * x), dtype=np.float64)
+
     with pytest.raises(Exception):
-        pe.least_squares(xs, ys, funcs)
+        pe.least_squares(xs, ys, {'a': func_a_invalid, 'b': func_b})
 
     pe.least_squares(xs, ys, funcs, num_grad=True)
 
@@ -930,7 +960,14 @@ def test_combined_fit_no_autograd():
     def func_b(a,x):
         return a[0]*np.exp(a[2]*x)
 
+    def func_a_autograd(a,x):
+        return a[0]*anp.exp(a[1]*x)
+
+    def func_b_autograd(a,x):
+        return a[0]*anp.exp(a[2]*x)
+
     funcs = {'a':func_a, 'b':func_b}
+    funcs_autograd = {'a':func_a_autograd, 'b':func_b_autograd}
     xs = {'a':xvals_a, 'b':xvals_b}
     ys = {'a':[pe.Obs([np.random.normal(item, item*1.5, 1000)],['ensemble1']) for item in func_exp1(xvals_a)],
         'b':[pe.Obs([np.random.normal(item, item*1.4, 1000)],['ensemble1']) for item in func_exp2(xvals_b)]}
@@ -938,8 +975,17 @@ def test_combined_fit_no_autograd():
     for key in funcs.keys():
         [item.gamma_method() for item in ys[key]]
 
+    # Since autograd 1.9.0 plain numpy ufuncs are dispatched to the autograd
+    # wrappers, so the fit using numpy now matches the one using autograd.numpy.
+    for r_np, r_ag in zip(pe.least_squares(xs, ys, funcs), pe.least_squares(xs, ys, funcs_autograd)):
+        assert r_np == r_ag
+
+    # A function that genuinely cannot be traced by autograd must still raise.
+    def func_a_invalid(a, x):
+        return np.array(a[0] * np.exp(a[1] * x), dtype=np.float64)
+
     with pytest.raises(Exception):
-        pe.least_squares(xs, ys, funcs)
+        pe.least_squares(xs, ys, {'a': func_a_invalid, 'b': func_b})
 
     pe.least_squares(xs, ys, funcs, num_grad=True)
 

tests/roots_test.py

@@ -1,4 +1,5 @@
 import numpy as np
+import autograd.numpy as anp
 import pyerrors as pe
 import pytest
 
@@ -37,11 +38,23 @@ def test_root_no_autograd():
     def root_function(x, d):
         return x - np.log(np.exp(d))
 
+    def root_function_autograd(x, d):
+        return x - anp.log(anp.exp(d))
+
     value = np.random.normal(0, 100)
     my_obs = pe.pseudo_Obs(value, 0.1, 't')
 
+    # Since autograd 1.9.0 plain numpy ufuncs are dispatched to the autograd
+    # wrappers, so a root function using numpy now yields the same result as
+    # one using autograd.numpy.
+    assert pe.roots.find_root(my_obs, root_function) == pe.roots.find_root(my_obs, root_function_autograd)
+
+    # A function that genuinely cannot be traced by autograd must still raise.
+    def root_invalid(x, d):
+        return x - np.float64(d)
+
     with pytest.raises(Exception):
-        my_root = pe.roots.find_root(my_obs, root_function)
+        pe.roots.find_root(my_obs, root_invalid)
 
 
 def test_root_multi_parameter():