Add a hook for torchvision.ops (#80)

* tests: add a test for torchvision.ops.nms

Reproduces the error from pyinstaller/pyinstaller#5386.

* Add a hook for torchvision.ops

Add a hidden import of torchvision._C extension module that we
fail to pick up automatically. This extension module is required
in functions from torchvision.ops.* modules, where its absence
causes the _assert_has_ops() call to fail.

Fixes the failing test_torchvision_nms.
Fixes pyinstaller/pyinstaller#5386.

news/80.new.rst

@@ -0,0 +1 @@
+Add hook for ``torchvision.ops`` to ensure that the required extension module (``torchvision._C``) is collected.

src/_pyinstaller_hooks_contrib/hooks/stdhooks/hook-torchvision.ops.py

@@ -0,0 +1,15 @@
+# ------------------------------------------------------------------
+# Copyright (c) 2020 PyInstaller Development Team.
+#
+# This file is distributed under the terms of the GNU General Public
+# License (version 2.0 or later).
+#
+# The full license is available in LICENSE.GPL.txt, distributed with
+# this software.
+#
+# SPDX-License-Identifier: GPL-2.0-or-later
+# ------------------------------------------------------------------
+
+# Functions from torchvision.ops.* modules require torchvision._C
+# extension module, which PyInstaller fails to pick up automatically...
+hiddenimports = ['torchvision._C']

src/_pyinstaller_hooks_contrib/tests/test_libraries.py

@@ -429,3 +429,36 @@ def test_pydantic(pyi_builder):
     pyi_builder.test_source("""
         import pydantic
         """)
+
+
+def torch_onedir_only(test):
+    def wrapped(pyi_builder):
+        if pyi_builder._mode != 'onedir':
+            pytest.skip('PyTorch tests support only onedir mode '
+                        'due to potential distribution size.')
+        test(pyi_builder)
+    return wrapped
+
+
+@importorskip('torchvision')
+@torch_onedir_only
+def test_torchvision_nms(pyi_builder):
+    pyi_builder.test_source("""
+        import torch
+        import torchvision
+        # boxes: Nx4 tensor (x1, y1, x2, y2)
+        boxes = torch.tensor([
+            [0.0, 0.0, 1.0, 1.0],
+            [0.45, 0.0, 1.0, 1.0],
+        ])
+        # scores: Nx1 tensor
+        scores = torch.tensor([
+            1.0,
+            1.1
+        ])
+        keep = torchvision.ops.nms(boxes, scores, 0.5)
+        # The boxes have IoU of 0.55, and the second one has a slightly
+        # higher score, so we expect it to be kept while the first one
+        # is discarded.
+        assert keep == 1
+    """)