Merge pull request #3485 from ColCarroll/rm-numdifftools

Remove numdifftools

Author: lucianopaz

RELEASE-NOTES.md

@@ -51,7 +51,8 @@
 
 - `nuts_kwargs` and `step_kwargs` have been deprecated in favor of using the standard `kwargs` to pass optional step method arguments.
 - `SGFS` and `CSG` have been removed (Fix for [#3353](https://github.com/pymc-devs/pymc3/issues/3353)). They have been moved to [pymc3-experimental](https://github.com/pymc-devs/pymc3-experimental).
--  References to `live_plot` and corresponding notebooks have been removed.
+- References to `live_plot` and corresponding notebooks have been removed.
+- Function `approx_hessian` was removed, due to `numdifftools` becoming incompatible with current `scipy`. The function was already optional, only available to a user who installed `numdifftools` separately, and not hit on any common codepaths. [#3485](https://github.com/pymc-devs/pymc3/pull/3485).
 
 ## PyMC3 3.6 (Dec 21 2018)
 

pymc3/tests/test_distributions.py

@@ -3,8 +3,8 @@
 
 from .helpers import SeededTest, select_by_precision
 from ..vartypes import continuous_types
-from ..model import Model, Point, Potential, Deterministic
-from ..blocking import DictToVarBijection, DictToArrayBijection, ArrayOrdering
+from ..model import Model, Point, Deterministic
+from ..blocking import DictToVarBijection
 from ..distributions import (
     DensityDist, Categorical, Multinomial, VonMises, Dirichlet,
     MvStudentT, MvNormal, MatrixNormal, ZeroInflatedPoisson,
@@ -471,37 +471,9 @@ def check_int_to_1(self, model, value, domain, paramdomains):
             area = integrate_nd(pdfx, domain, value.dshape, value.dtype)
             assert_almost_equal(area, 1, err_msg=str(pt))
 
-    def check_dlogp(self, model, value, domain, paramdomains):
-        try:
-            from numdifftools import Gradient
-        except ImportError:
-            return
-        if not model.cont_vars:
-            return
-
-        domains = paramdomains.copy()
-        domains['value'] = domain
-        bij = DictToArrayBijection(
-            ArrayOrdering(model.cont_vars), model.test_point)
-        dlogp = bij.mapf(model.fastdlogp(model.cont_vars))
-        logp = bij.mapf(model.fastlogp)
-
-        def wrapped_logp(x):
-            try:
-                return logp(x)
-            except:
-                return np.nan
-
-        ndlogp = Gradient(wrapped_logp)
-        for pt in product(domains, n_samples=100):
-            pt = Point(pt, model=model)
-            pt = bij.map(pt)
-            decimals = select_by_precision(float64=6, float32=4)
-            assert_almost_equal(dlogp(pt), ndlogp(pt), decimal=decimals, err_msg=str(pt))
-
     def checkd(self, distfam, valuedomain, vardomains, checks=None, extra_args=None):
         if checks is None:
-            checks = (self.check_int_to_1, self.check_dlogp)
+            checks = (self.check_int_to_1, )
 
         if extra_args is None:
             extra_args = {}
@@ -940,7 +912,8 @@ def test_wishart(self, n):
         # This check compares the autodiff gradient to the numdiff gradient.
         # However, due to the strict constraints of the wishart,
         # it is impossible to numerically determine the gradient as a small
-        # pertubation breaks the symmetry. Thus disabling.
+        # pertubation breaks the symmetry. Thus disabling. Also, numdifftools was
+        # removed in June 2019, so an alternative would be needed.
         #
         # self.checkd(Wishart, PdMatrix(n), {'n': Domain([2, 3, 4, 2000]), 'V': PdMatrix(n)},
         #             checks=[self.check_dlogp])
@@ -1120,12 +1093,6 @@ def logp(x):
             return -log(2 * .5) - abs(x - .5) / .5
         self.checkd(DensityDist, R, {}, extra_args={'logp': logp})
 
-    def test_addpotential(self):
-        with Model() as model:
-            value = Normal('value', 1, 1)
-            Potential('value_squared', -value ** 2)
-            self.check_dlogp(model, value, R, {})
-
     def test_get_tau_sigma(self):
         sigma = np.array([2])
         assert_almost_equal(continuous.get_tau_sigma(sigma=sigma), [1. / sigma**2, sigma])

pymc3/tuning/__init__.py

@@ -1,2 +1,2 @@
 from .starting import find_MAP
-from .scaling import approx_hessian, find_hessian, trace_cov, guess_scaling
+from .scaling import find_hessian, trace_cov, guess_scaling

pymc3/tuning/scaling.py

@@ -4,41 +4,7 @@
 from ..theanof import hessian_diag, inputvars
 from ..blocking import DictToArrayBijection, ArrayOrdering
 
-__all__ = ['approx_hessian', 'find_hessian', 'trace_cov', 'guess_scaling']
-
-
-def approx_hessian(point, vars=None, model=None):
-    """
-    Returns an approximation of the Hessian at the current chain location.
-
-    Parameters
-    ----------
-    model : Model (optional if in `with` context)
-    point : dict
-    vars : list
-        Variables for which Hessian is to be calculated.
-    """
-    from numdifftools import Jacobian
-
-    model = modelcontext(model)
-    if vars is None:
-        vars = model.cont_vars
-    vars = inputvars(vars)
-
-    point = Point(point, model=model)
-
-    bij = DictToArrayBijection(ArrayOrdering(vars), point)
-    dlogp = bij.mapf(model.fastdlogp(vars))
-
-    def grad_logp(point):
-        return np.nan_to_num(dlogp(point))
-
-    '''
-    Find the jacobian of the gradient function at the current position
-    this should be the Hessian; invert it to find the approximate
-    covariance matrix.
-    '''
-    return -Jacobian(grad_logp)(bij.map(point))
+__all__ = ['find_hessian', 'trace_cov', 'guess_scaling']
 
 
 def fixed_hessian(point, vars=None, model=None):

requirements-dev.txt

@@ -7,7 +7,6 @@ Keras>=2.0.8
 nbsphinx>=0.2.13
 nose>=1.3.7
 nose-parameterized==0.6.0
-numdifftools>=0.9.20
 numpy>=1.13.0
 numpydoc==0.7.0
 pycodestyle>=2.3.1