[PoC] use pytest configuration file

.circleci/unittest/linux/scripts/install.sh

@@ -71,7 +71,7 @@ fi
 # Note: installing librosa via pip fail because it will try to compile numba.
 (
     set -x
-    conda install -y -c conda-forge ${NUMBA_DEV_CHANNEL} 'librosa==0.10.0' parameterized 'requests>=2.20'
+    conda install -y -c conda-forge ${NUMBA_DEV_CHANNEL} sox libvorbis 'librosa==0.10.0' parameterized 'requests>=2.20'
     pip install kaldi-io SoundFile coverage pytest pytest-cov 'scipy==1.7.3' transformers expecttest unidecode inflect Pillow sentencepiece pytorch-lightning 'protobuf<4.21.0' demucs tinytag pyroomacoustics flashlight-text git+https://github.com/kpu/kenlm
 )
 # Install fairseq

.circleci/unittest/windows/scripts/install.sh

@@ -67,7 +67,7 @@ case "$(python --version)" in
 esac
 # Note: installing librosa via pip fail because it will try to compile numba.
 (
-    conda install -y -c conda-forge ${NUMBA_DEV_CHANNEL} 'librosa==0.10.0' parameterized 'requests>=2.20'
+    conda install -y -c conda-forge ${NUMBA_DEV_CHANNEL} sox libvorbis 'librosa==0.10.0' parameterized 'requests>=2.20'
     # Need to disable shell check since this'll fail out if SENTENCEPIECE_DEPENDENCY is empty
     # shellcheck disable=SC2086
     pip install \

.github/workflows/build_docs.yml

@@ -119,7 +119,7 @@ jobs:
         pip --quiet install git+https://github.com/kpu/kenlm/ flashlight-text
 
         # Install build tools
-        conda install --quiet -y -c conda-forge pandoc doxygen pysoundfile
+        conda install --quiet -y -c conda-forge sox pandoc doxygen pysoundfile
         pip install --quiet -r docs/requirements.txt -r docs/requirements-tutorials.txt
 
         # Build docs

.github/workflows/unittest-linux-cpu.yml

@@ -50,6 +50,8 @@ jobs:
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_AUDIO_OUT_DEVICE=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MACOS=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX_DECODER=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX_ENCODER=true
         echo '::endgroup::'
 
         set -euxo pipefail

.github/workflows/unittest-linux-gpu.yml

@@ -60,7 +60,7 @@ jobs:
         USE_FFMPEG=1 python3 -m pip install -v -e . --no-use-pep517
 
         # Install test tools
-        conda install -y --quiet -c conda-forge -c numba/label/dev 'librosa==0.10.0' parameterized 'requests>=2.20'
+        conda install -y --quiet -c conda-forge -c numba/label/dev sox libvorbis 'librosa==0.10.0' parameterized 'requests>=2.20'
         python3 -m pip install --quiet kaldi-io SoundFile coverage pytest pytest-cov 'scipy==1.7.3' transformers expecttest unidecode inflect Pillow sentencepiece pytorch-lightning 'protobuf<4.21.0' demucs tinytag flashlight-text git+https://github.com/kpu/kenlm/
         python3 -m pip install --quiet git+https://github.com/pytorch/fairseq.git@e47a4c8
 

.github/workflows/unittest-macos-cpu.yml

@@ -49,6 +49,8 @@ jobs:
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_ON_PYTHON_310=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_sentencepiece=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_AUDIO_OUT_DEVICE=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX_DECODER=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX_ENCODER=true
         echo '::endgroup::'
 
         set -euxo pipefail

pytest.ini

@@ -0,0 +1,13 @@
+[pytest]
+addopts =
+    # show tests that (f)ailed, (E)rror, or (X)passed in the summary
+    -rfEX
+    # Make tracebacks shorter
+    --tb=native
+    # enable all warnings
+    -Wd
+    --ignore=test/test_datasets_download.py
+    --ignore-glob=test/test_prototype_*.py
+testpaths =
+    test
+xfail_strict = True

setup.py

@@ -104,31 +104,13 @@ def _parse_url(path):
                 yield url
 
 
-def _parse_sources():
-    third_party_dir = ROOT_DIR / "third_party"
-    libs = ["sox"]
-    archive_dir = third_party_dir / "archives"
-    archive_dir.mkdir(exist_ok=True)
-    for lib in libs:
-        cmake_file = third_party_dir / lib / "CMakeLists.txt"
-        for url in _parse_url(cmake_file):
-            path = archive_dir / os.path.basename(url)
-            yield path, url
-
-
 def _fetch_archives(src):
     for dest, url in src:
         if not dest.exists():
             print(f" --- Fetching {os.path.basename(dest)}")
             torch.hub.download_url_to_file(url, dest, progress=False)
 
 
-def _fetch_third_party_libraries():
-    _init_submodule()
-    if os.name != "nt":
-        _fetch_archives(_parse_sources())
-
-
 def _main():
     sha = _run_cmd(["git", "rev-parse", "HEAD"])
     branch = _run_cmd(["git", "rev-parse", "--abbrev-ref", "HEAD"])
@@ -142,7 +124,7 @@ def _main():
     print("-- Building version", version)
 
     _make_version_file(version, sha)
-    _fetch_third_party_libraries()
+    _init_submodule()
 
     with open("README.md") as f:
         long_description = f.read()

test/torchaudio_unittest/backend/dispatcher/sox/info_test.py

@@ -15,6 +15,7 @@
     skipIfNoExec,
     skipIfNoModule,
     skipIfNoSox,
+    skipIfNoSoxDecoder,
     sox_utils,
     TempDirMixin,
 )
@@ -178,6 +179,7 @@ def test_amb(self, dtype, sample_rate, num_channels):
         assert info.bits_per_sample == bits_per_sample
         assert info.encoding == get_encoding("amb", dtype)
 
+    @skipIfNoSoxDecoder("amr_nb")
     def test_amr_nb(self):
         """`self._info` can check amr-nb file correctly"""
         duration = 1
@@ -256,7 +258,7 @@ def test_htk(self):
         assert info.encoding == "PCM_S"
 
 
-@skipIfNoSox
+@skipIfNoSoxDecoder("opus")
 class TestInfoOpus(PytorchTestCase):
     _info = partial(get_info_func(), backend="sox")
 

test/torchaudio_unittest/backend/dispatcher/sox/load_test.py

@@ -14,6 +14,7 @@
     save_wav,
     skipIfNoExec,
     skipIfNoSox,
+    skipIfNoSoxDecoder,
     sox_utils,
     TempDirMixin,
 )
@@ -238,6 +239,7 @@ def test_vorbis_large(self, sample_rate, num_channels, quality_level):
         ),
         name_func=name_func,
     )
+    @skipIfNoSoxDecoder("opus")
     def test_opus(self, bitrate, num_channels, compression_level):
         """`sox_io_backend.load` can load opus file correctly."""
         ops_path = get_asset_path("io", f"{bitrate}_{compression_level}_{num_channels}ch.opus")
@@ -282,6 +284,7 @@ def test_amb(self, dtype, sample_rate, num_channels, normalize):
             "amb", sample_rate, num_channels, bit_depth=bit_depth, duration=1, encoding=encoding, normalize=normalize
         )
 
+    @skipIfNoSoxDecoder("amr_nb")
     def test_amr_nb(self):
         """`sox_io_backend.load` can load amr_nb format correctly."""
         self.assert_format("amr-nb", sample_rate=8000, num_channels=1, bit_depth=32, duration=1)

test/torchaudio_unittest/backend/dispatcher/sox/save_test.py

@@ -13,6 +13,7 @@
     save_wav,
     skipIfNoExec,
     skipIfNoSox,
+    skipIfNoSoxEncoder,
     sox_utils,
     TempDirMixin,
     TorchaudioTestCase,
@@ -253,6 +254,7 @@ def test_save_amb(self, enc_params):
         encoding, bits_per_sample = enc_params
         self.assert_save_consistency("amb", encoding=encoding, bits_per_sample=bits_per_sample, test_mode="path")
 
+    @skipIfNoSoxEncoder("amr_nb")
     def test_save_amr_nb(self):
         self.assert_save_consistency("amr-nb", num_channels=1, test_mode="path")
 
@@ -269,12 +271,18 @@ def test_save_gsm(self):
             ("flac",),
             ("vorbis",),
             ("sph", "PCM_S", 16),
-            ("amr-nb",),
             ("amb", "PCM_S", 16),
         ],
         name_func=name_func,
     )
     def test_save_large(self, format, encoding=None, bits_per_sample=None):
+        self._test_save_large(format, encoding, bits_per_sample)
+
+    @skipIfNoSoxEncoder("amr_nb")
+    def test_save_large_amr_nb(self):
+        self._test_save_large("amr_nb")
+
+    def _test_save_large(self, format, encoding=None, bits_per_sample=None):
         """`self._save` can save large files."""
         sample_rate = 8000
         one_hour = 60 * 60 * sample_rate

test/torchaudio_unittest/backend/sox_io/info_test.py

@@ -10,6 +10,7 @@
     save_wav,
     skipIfNoExec,
     skipIfNoSox,
+    skipIfNoSoxDecoder,
     sox_utils,
     TempDirMixin,
 )
@@ -196,6 +197,7 @@ def test_amb(self, dtype, sample_rate, num_channels):
         assert info.bits_per_sample == bits_per_sample
         assert info.encoding == get_encoding("amb", dtype)
 
+    @skipIfNoSoxDecoder("amr_nb")
     def test_amr_nb(self):
         """`sox_io_backend.info` can check amr-nb file correctly"""
         duration = 1

test/torchaudio_unittest/backend/sox_io/load_test.py

@@ -18,6 +18,7 @@
     skipIfNoExec,
     skipIfNoModule,
     skipIfNoSox,
+    skipIfNoSoxDecoder,
     sox_utils,
     TempDirMixin,
 )
@@ -244,6 +245,7 @@ def test_vorbis_large(self, sample_rate, num_channels, quality_level):
         ),
         name_func=name_func,
     )
+    @skipIfNoSoxDecoder("opus")
     def test_opus(self, bitrate, num_channels, compression_level):
         """`sox_io_backend.load` can load opus file correctly."""
         ops_path = get_asset_path("io", f"{bitrate}_{compression_level}_{num_channels}ch.opus")
@@ -288,6 +290,7 @@ def test_amb(self, dtype, sample_rate, num_channels, normalize):
             "amb", sample_rate, num_channels, bit_depth=bit_depth, duration=1, encoding=encoding, normalize=normalize
         )
 
+    @skipIfNoSoxDecoder("amr_nb")
     def test_amr_nb(self):
         """`sox_io_backend.load` can load amr_nb format correctly."""
         self.assert_format("amr-nb", sample_rate=8000, num_channels=1, bit_depth=32, duration=1)

test/torchaudio_unittest/backend/sox_io/save_test.py

@@ -12,6 +12,7 @@
     save_wav,
     skipIfNoExec,
     skipIfNoSox,
+    skipIfNoSoxEncoder,
     sox_utils,
     TempDirMixin,
     TorchaudioTestCase,
@@ -266,6 +267,7 @@ def test_save_amb(self, enc_params):
             7,
         ],
     )
+    @skipIfNoSoxEncoder("amr-nb")
     def test_save_amr_nb(self, bit_rate):
         self.assert_save_consistency("amr-nb", compression=bit_rate, num_channels=1)
 
@@ -282,12 +284,18 @@ def test_save_gsm(self):
             ("flac",),
             ("vorbis",),
             ("sph", "PCM_S", 16),
-            ("amr-nb",),
             ("amb", "PCM_S", 16),
         ],
         name_func=name_func,
     )
     def test_save_large(self, format, encoding=None, bits_per_sample=None):
+        self._test_save_large(format, encoding, bits_per_sample)
+
+    @skipIfNoSoxEncoder
+    def test_save_large_amr_nb(self):
+        self.test_save_large("amr_nb")
+
+    def _test_save_large(self, format, encoding=None, bits_per_sample=None):
         """`sox_io_backend.save` can save large files."""
         sample_rate = 8000
         one_hour = 60 * 60 * sample_rate

test/torchaudio_unittest/common_utils/__init__.py

@@ -19,6 +19,8 @@
     skipIfNoQengine,
     skipIfNoRIR,
     skipIfNoSox,
+    skipIfNoSoxDecoder,
+    skipIfNoSoxEncoder,
     skipIfPy310,
     skipIfRocm,
     TempDirMixin,

test/torchaudio_unittest/common_utils/case_utils.py

@@ -234,6 +234,24 @@ def skipIfNoModule(module, display_name=None):
     reason="Sox features are not available.",
     key="NO_SOX",
 )
+
+
+def skipIfNoSoxDecoder(ext):
+    return _skipIf(
+        not torchaudio._extension._SOX_INITIALIZED or ext not in torchaudio.utils.sox_utils.list_read_formats(),
+        f'sox does not handle "{ext}" for read.',
+        key=f"NO_SOX_DECODER",
+    )
+
+
+def skipIfNoSoxEncoder(ext):
+    return _skipIf(
+        not torchaudio._extension._SOX_INITIALIZED or ext not in torchaudio.utils.sox_utils.list_write_formats(),
+        f'sox does not handle "{ext}" for read.',
+        key=f"NO_SOX_ENCODER",
+    )
+
+
 skipIfNoKaldi = _skipIf(
     not torchaudio._extension._IS_KALDI_AVAILABLE,
     reason="Kaldi features are not available.",

test/torchaudio_unittest/sox_effect/dataset_test.py

@@ -94,7 +94,7 @@ def test_apply_effects_file(self):
         loader = torch.utils.data.DataLoader(
             dataset,
             batch_size=32,
-            num_workers=16,
+            num_workers=4,
             worker_init_fn=init_random_seed,
         )
         for batch in loader:
@@ -115,7 +115,7 @@ def test_apply_effects_tensor(self):
         loader = torch.utils.data.DataLoader(
             dataset,
             batch_size=32,
-            num_workers=16,
+            num_workers=4,
             worker_init_fn=init_random_seed,
         )
         for batch in loader: