From eaefe49f3d00ae7e085f9328c8e0037ead06d6a2 Mon Sep 17 00:00:00 2001
From: Christoph Boeddeker <cbj@mail.uni-paderborn.de>
Date: Tue, 18 Jun 2024 12:56:54 +0200
Subject: [PATCH] fix doctests for numpy 2

---
 padertorch/data/segment.py          | 7 +++++--
 padertorch/ops/losses/regression.py | 4 ++--
 padertorch/utils.py                 | 6 ++++--
 3 files changed, 11 insertions(+), 6 deletions(-)

diff --git a/padertorch/data/segment.py b/padertorch/data/segment.py
index 9b778b00..8fccb159 100644
--- a/padertorch/data/segment.py
+++ b/padertorch/data/segment.py
@@ -346,7 +346,7 @@ def _get_rand_int(rng, *args, **kwargs):
 
 def get_anchor(
         num_samples: int, length: int, shift: int = None,
-        mode: str = 'left', rng=np.random
+        mode: str = 'left', rng: 'np.random.RandomState' = np.random
 ) -> int:
     """
     Calculates anchor for the boundaries for segmentation of a signal
@@ -372,6 +372,9 @@ def get_anchor(
             random_max_segments: Randomly chooses the anchor such that
                 the maximum number of segments are created
         rng: random number generator (`numpy.random`)
+            e.g. `np.random.RandomState(0)` or `np.random.default_rng(0)`.
+            Not, that `np.random.default_rng` differs from `np.random` and has
+            sometimes other return types (e.g. python `int` vs `np.int64`)
 
     Returns:
        integer value describing the anchor
@@ -387,7 +390,7 @@ def get_anchor(
     1
     >>> get_anchor(24, 10, 3, mode='random')
     10
-    >>> get_anchor(24, 10, 3, mode='random', rng=np.random.default_rng(seed=4))
+    >>> get_anchor(24, 10, 3, mode='random', rng=np.random.RandomState(seed=4))
     10
     >>> get_anchor(24, 10, 3, mode='random_max_segments')
     3
diff --git a/padertorch/ops/losses/regression.py b/padertorch/ops/losses/regression.py
index 79eb75b7..83c0cb45 100644
--- a/padertorch/ops/losses/regression.py
+++ b/padertorch/ops/losses/regression.py
@@ -220,7 +220,7 @@ def si_sdr_loss(estimate, target, reduction='mean', offset_invariant=False,
     ...     print('Numpy metric:', si_sdr(estimate.numpy(), target.numpy()))
 
     Perfect estimation
-    >>> si_sdr(reference.numpy(), reference.numpy())
+    >>> print(si_sdr(reference.numpy(), reference.numpy()))
     inf
     >>> sdr_loss(reference, reference)
     tensor(-inf, dtype=torch.float64)
@@ -228,7 +228,7 @@ def si_sdr_loss(estimate, target, reduction='mean', offset_invariant=False,
     tensor(True)
     >>> si_sdr_loss(reference.to(torch.float32), reference.to(torch.float32)) < -130  # Torch CPU is not hardware independent
     tensor(True)
-    >>> si_sdr(reference.numpy(), (reference * 2).numpy())
+    >>> print(si_sdr(reference.numpy(), (reference * 2).numpy()))
     inf
     >>> si_sdr_loss(reference, reference * 2) < -300  # Torch CPU is not hardware independent
     tensor(True)
diff --git a/padertorch/utils.py b/padertorch/utils.py
index 40bb165c..ac594ff4 100644
--- a/padertorch/utils.py
+++ b/padertorch/utils.py
@@ -32,14 +32,16 @@ def to_list(x, length=None):
     Complicated corner cases are e.g. `range()` and `dict.values()`, which are
     handled here.
 
+    >>> from paderbox.utils.pretty import pprint
+
     >>> to_list(1)
     [1]
     >>> to_list([1])
     [1]
     >>> to_list((i for i in range(3)))
     [0, 1, 2]
-    >>> to_list(np.arange(3))
-    [0, 1, 2]
+    >>> pprint(to_list(np.arange(3)), nep51=True)  # use pprint to support numpy 1 and 2
+    [np.int64(0), np.int64(1), np.int64(2)]
     >>> to_list({'a': 1})
     [{'a': 1}]
     >>> to_list({'a': 1}.keys())