decode mp3 with librosa if torchaudio >= 0.12 doesn't work as a tempo…

…rary workaround (#4923)

* decode mp3 with librosa if torchaudio is > 0.12 (ideally version of ffmpeg should be checked too)

* decode mp3 with torchaudio>=0.12 if it works (instead of librosa)

* fix incorrect marks for mp3 tests (require torchaudio, not sndfile)

* add tests for latest torchaudio + separate stage in CI for it (first try)

* install ffmpeg only on ubuntu

* use mock to emulate torchaudio fail, add tests for librosa (not all of them)

* test torchaudio_latest only on ubuntu

* try/except decoding with librosa for file-like objects

* more tests for latest torchaudio, should be comlpete set now

* replace logging with warnings

* fix tests: catch warnings with a pytest context manager

Co-authored-by: Quentin Lhoest <42851186+lhoestq@users.noreply.github.com>

.github/workflows/ci.yml

@@ -72,3 +72,13 @@ jobs:
       - name: Test with pytest
         run: |
           python -m pytest -rfExX -m ${{ matrix.test }} -n 2 --dist loadfile -sv ./tests/
+      - name: Install dependencies to test torchaudio>=0.12 on Ubuntu
+        if: ${{ matrix.os == 'ubuntu-latest' }}
+        run: |
+          pip uninstall -y torchaudio torch
+          pip install "torchaudio>=0.12"
+          sudo apt-get -y install ffmpeg
+      - name: Test torchaudio>=0.12 on Ubuntu
+        if: ${{ matrix.os == 'ubuntu-latest' }}
+        run: |
+          python -m pytest -rfExX -m torchaudio_latest -n 2 --dist loadfile -sv ./tests/features/test_audio.py

src/datasets/features/audio.py

@@ -1,4 +1,5 @@
 import os
+import warnings
 from dataclasses import dataclass, field
 from io import BytesIO
 from typing import TYPE_CHECKING, Any, ClassVar, Dict, Optional, Union
@@ -268,7 +269,7 @@ def _decode_non_mp3_file_like(self, file, format=None):
             if version.parse(sf.__libsndfile_version__) < version.parse("1.0.30"):
                 raise RuntimeError(
                     "Decoding .opus files requires 'libsndfile'>=1.0.30, "
-                    + "it can be installed via conda: `conda install -c conda-forge libsndfile>=1.0.30`"
+                    + 'it can be installed via conda: `conda install -c conda-forge "libsndfile>=1.0.30"`'
                 )
         array, sampling_rate = sf.read(file)
         array = array.T
@@ -282,19 +283,44 @@ def _decode_non_mp3_file_like(self, file, format=None):
     def _decode_mp3(self, path_or_file):
         try:
             import torchaudio
-            import torchaudio.transforms as T
         except ImportError as err:
-            raise ImportError(
-                "Decoding 'mp3' audio files, requires 'torchaudio<0.12.0': pip install 'torchaudio<0.12.0'"
-            ) from err
-        if not version.parse(torchaudio.__version__) < version.parse("0.12.0"):
-            raise RuntimeError(
-                "Decoding 'mp3' audio files, requires 'torchaudio<0.12.0': pip install 'torchaudio<0.12.0'"
-            )
-        try:
-            torchaudio.set_audio_backend("sox_io")
-        except RuntimeError as err:
-            raise ImportError("To support decoding 'mp3' audio files, please install 'sox'.") from err
+            raise ImportError("To support decoding 'mp3' audio files, please install 'torchaudio'.") from err
+        if version.parse(torchaudio.__version__) < version.parse("0.12.0"):
+            try:
+                torchaudio.set_audio_backend("sox_io")
+            except RuntimeError as err:
+                raise ImportError("To support decoding 'mp3' audio files, please install 'sox'.") from err
+            array, sampling_rate = self._decode_mp3_torchaudio(path_or_file)
+        else:
+            try:  # try torchaudio anyway because sometimes it works (depending on the os and os packages installed)
+                array, sampling_rate = self._decode_mp3_torchaudio(path_or_file)
+            except RuntimeError:
+                try:
+                    # flake8: noqa
+                    import librosa
+                except ImportError as err:
+                    raise ImportError(
+                        "Your version of `torchaudio` (>=0.12.0) doesn't support decoding 'mp3' files on your machine. "
+                        "To support 'mp3' decoding with `torchaudio>=0.12.0`, please install `ffmpeg>=4` system package "
+                        'or downgrade `torchaudio` to <0.12: `pip install "torchaudio<0.12"`. '
+                        "To support decoding 'mp3' audio files without `torchaudio`, please install `librosa`: "
+                        "`pip install librosa`. Note that decoding will be extremely slow in that case."
+                    ) from err
+                # try to decode with librosa for torchaudio>=0.12.0 as a workaround
+                warnings.warn("Decoding mp3 with `librosa` instead of `torchaudio`, decoding is slow.")
+                try:
+                    array, sampling_rate = self._decode_mp3_librosa(path_or_file)
+                except RuntimeError as err:
+                    raise RuntimeError(
+                        "Decoding of 'mp3' failed, probably because of streaming mode "
+                        "(`librosa` cannot decode 'mp3' file-like objects, only path-like)."
+                    ) from err
+
+        return array, sampling_rate
+
+    def _decode_mp3_torchaudio(self, path_or_file):
+        import torchaudio
+        import torchaudio.transforms as T
 
         array, sampling_rate = torchaudio.load(path_or_file, format="mp3")
         if self.sampling_rate and self.sampling_rate != sampling_rate:
@@ -306,3 +332,9 @@ def _decode_mp3(self, path_or_file):
         if self.mono:
             array = array.mean(axis=0)
         return array, sampling_rate
+
+    def _decode_mp3_librosa(self, path_or_file):
+        import librosa
+
+        array, sampling_rate = librosa.load(path_or_file, mono=self.mono, sr=self.sampling_rate)
+        return array, sampling_rate

tests/conftest.py

@@ -15,6 +15,10 @@ def pytest_collection_modifyitems(config, items):
         item.add_marker(pytest.mark.unit)
 
 
+def pytest_configure(config):
+    config.addinivalue_line("markers", "torchaudio_latest: mark test to run with torchaudio>=0.12")
+
+
 @pytest.fixture(autouse=True)
 def set_test_cache_config(tmp_path_factory, monkeypatch):
     # test_hf_cache_home = tmp_path_factory.mktemp("cache")  # TODO: why a cache dir per test function does not work?

tests/features/test_audio.py

@@ -1,13 +1,21 @@
 import os
 import tarfile
+from contextlib import nullcontext
+from unittest.mock import patch
 
 import pyarrow as pa
 import pytest
 
 from datasets import Dataset, concatenate_datasets, load_dataset
 from datasets.features import Audio, Features, Sequence, Value
 
-from ..utils import require_libsndfile_with_opus, require_sndfile, require_sox, require_torchaudio
+from ..utils import (
+    require_libsndfile_with_opus,
+    require_sndfile,
+    require_sox,
+    require_torchaudio,
+    require_torchaudio_latest,
+)
 
 
 @pytest.fixture()
@@ -135,6 +143,26 @@ def test_audio_decode_example_mp3(shared_datadir):
     assert decoded_example["sampling_rate"] == 44100
 
 
+@pytest.mark.torchaudio_latest
+@require_torchaudio_latest
+@pytest.mark.parametrize("torchaudio_failed", [False, True])
+def test_audio_decode_example_mp3_torchaudio_latest(shared_datadir, torchaudio_failed):
+    audio_path = str(shared_datadir / "test_audio_44100.mp3")
+    audio = Audio()
+
+    with patch("torchaudio.load") if torchaudio_failed else nullcontext() as load_mock, pytest.warns(
+        UserWarning, match=r"Decoding mp3 with `librosa` instead of `torchaudio`.+?"
+    ) if torchaudio_failed else nullcontext():
+
+        if torchaudio_failed:
+            load_mock.side_effect = RuntimeError()
+
+        decoded_example = audio.decode_example(audio.encode_example(audio_path))
+        assert decoded_example["path"] == audio_path
+        assert decoded_example["array"].shape == (110592,)
+        assert decoded_example["sampling_rate"] == 44100
+
+
 @require_libsndfile_with_opus
 def test_audio_decode_example_opus(shared_datadir):
     audio_path = str(shared_datadir / "test_audio_48000.opus")
@@ -178,6 +206,34 @@ def test_audio_resampling_mp3_different_sampling_rates(shared_datadir):
     assert decoded_example["sampling_rate"] == 48000
 
 
+@pytest.mark.torchaudio_latest
+@require_torchaudio_latest
+@pytest.mark.parametrize("torchaudio_failed", [False, True])
+def test_audio_resampling_mp3_different_sampling_rates_torchaudio_latest(shared_datadir, torchaudio_failed):
+    audio_path = str(shared_datadir / "test_audio_44100.mp3")
+    audio_path2 = str(shared_datadir / "test_audio_16000.mp3")
+    audio = Audio(sampling_rate=48000)
+
+    # if torchaudio>=0.12 failed, mp3 must be decoded anyway (with librosa)
+    with patch("torchaudio.load") if torchaudio_failed else nullcontext() as load_mock, pytest.warns(
+        UserWarning, match=r"Decoding mp3 with `librosa` instead of `torchaudio`.+?"
+    ) if torchaudio_failed else nullcontext():
+        if torchaudio_failed:
+            load_mock.side_effect = RuntimeError()
+
+        decoded_example = audio.decode_example(audio.encode_example(audio_path))
+        assert decoded_example.keys() == {"path", "array", "sampling_rate"}
+        assert decoded_example["path"] == audio_path
+        assert decoded_example["array"].shape == (120373,)
+        assert decoded_example["sampling_rate"] == 48000
+
+        decoded_example = audio.decode_example(audio.encode_example(audio_path2))
+        assert decoded_example.keys() == {"path", "array", "sampling_rate"}
+        assert decoded_example["path"] == audio_path2
+        assert decoded_example["array"].shape == (122688,)
+        assert decoded_example["sampling_rate"] == 48000
+
+
 @require_sndfile
 def test_dataset_with_audio_feature(shared_datadir):
     audio_path = str(shared_datadir / "test_audio_44100.wav")
@@ -266,6 +322,38 @@ def test_dataset_with_audio_feature_tar_mp3(tar_mp3_path):
     assert column[0]["sampling_rate"] == 44100
 
 
+@pytest.mark.torchaudio_latest
+@require_torchaudio_latest
+def test_dataset_with_audio_feature_tar_mp3_torchaudio_latest(tar_mp3_path):
+    # no test for librosa here because it doesn't support file-like objects, only paths
+    audio_filename = "test_audio_44100.mp3"
+    data = {"audio": []}
+    for file_path, file_obj in iter_archive(tar_mp3_path):
+        data["audio"].append({"path": file_path, "bytes": file_obj.read()})
+        break
+    features = Features({"audio": Audio()})
+    dset = Dataset.from_dict(data, features=features)
+    item = dset[0]
+    assert item.keys() == {"audio"}
+    assert item["audio"].keys() == {"path", "array", "sampling_rate"}
+    assert item["audio"]["path"] == audio_filename
+    assert item["audio"]["array"].shape == (110592,)
+    assert item["audio"]["sampling_rate"] == 44100
+    batch = dset[:1]
+    assert batch.keys() == {"audio"}
+    assert len(batch["audio"]) == 1
+    assert batch["audio"][0].keys() == {"path", "array", "sampling_rate"}
+    assert batch["audio"][0]["path"] == audio_filename
+    assert batch["audio"][0]["array"].shape == (110592,)
+    assert batch["audio"][0]["sampling_rate"] == 44100
+    column = dset["audio"]
+    assert len(column) == 1
+    assert column[0].keys() == {"path", "array", "sampling_rate"}
+    assert column[0]["path"] == audio_filename
+    assert column[0]["array"].shape == (110592,)
+    assert column[0]["sampling_rate"] == 44100
+
+
 @require_sndfile
 def test_dataset_with_audio_feature_with_none():
     data = {"audio": [None]}
@@ -328,7 +416,7 @@ def test_resampling_at_loading_dataset_with_audio_feature(shared_datadir):
 
 
 @require_sox
-@require_sndfile
+@require_torchaudio
 def test_resampling_at_loading_dataset_with_audio_feature_mp3(shared_datadir):
     audio_path = str(shared_datadir / "test_audio_44100.mp3")
     data = {"audio": [audio_path]}
@@ -355,6 +443,43 @@ def test_resampling_at_loading_dataset_with_audio_feature_mp3(shared_datadir):
     assert column[0]["sampling_rate"] == 16000
 
 
+@pytest.mark.torchaudio_latest
+@require_torchaudio_latest
+@pytest.mark.parametrize("torchaudio_failed", [False, True])
+def test_resampling_at_loading_dataset_with_audio_feature_mp3_torchaudio_latest(shared_datadir, torchaudio_failed):
+    audio_path = str(shared_datadir / "test_audio_44100.mp3")
+    data = {"audio": [audio_path]}
+    features = Features({"audio": Audio(sampling_rate=16000)})
+    dset = Dataset.from_dict(data, features=features)
+
+    # if torchaudio>=0.12 failed, mp3 must be decoded anyway (with librosa)
+    with patch("torchaudio.load") if torchaudio_failed else nullcontext() as load_mock, pytest.warns(
+        UserWarning, match=r"Decoding mp3 with `librosa` instead of `torchaudio`.+?"
+    ) if torchaudio_failed else nullcontext():
+        if torchaudio_failed:
+            load_mock.side_effect = RuntimeError()
+
+        item = dset[0]
+        assert item.keys() == {"audio"}
+        assert item["audio"].keys() == {"path", "array", "sampling_rate"}
+        assert item["audio"]["path"] == audio_path
+        assert item["audio"]["array"].shape == (40125,)
+        assert item["audio"]["sampling_rate"] == 16000
+        batch = dset[:1]
+        assert batch.keys() == {"audio"}
+        assert len(batch["audio"]) == 1
+        assert batch["audio"][0].keys() == {"path", "array", "sampling_rate"}
+        assert batch["audio"][0]["path"] == audio_path
+        assert batch["audio"][0]["array"].shape == (40125,)
+        assert batch["audio"][0]["sampling_rate"] == 16000
+        column = dset["audio"]
+        assert len(column) == 1
+        assert column[0].keys() == {"path", "array", "sampling_rate"}
+        assert column[0]["path"] == audio_path
+        assert column[0]["array"].shape == (40125,)
+        assert column[0]["sampling_rate"] == 16000
+
+
 @require_sndfile
 def test_resampling_after_loading_dataset_with_audio_feature(shared_datadir):
     audio_path = str(shared_datadir / "test_audio_44100.wav")
@@ -386,7 +511,7 @@ def test_resampling_after_loading_dataset_with_audio_feature(shared_datadir):
 
 
 @require_sox
-@require_sndfile
+@require_torchaudio
 def test_resampling_after_loading_dataset_with_audio_feature_mp3(shared_datadir):
     audio_path = str(shared_datadir / "test_audio_44100.mp3")
     data = {"audio": [audio_path]}
@@ -416,6 +541,46 @@ def test_resampling_after_loading_dataset_with_audio_feature_mp3(shared_datadir)
     assert column[0]["sampling_rate"] == 16000
 
 
+@pytest.mark.torchaudio_latest
+@require_torchaudio_latest
+@pytest.mark.parametrize("torchaudio_failed", [False, True])
+def test_resampling_after_loading_dataset_with_audio_feature_mp3_torchaudio_latest(shared_datadir, torchaudio_failed):
+    audio_path = str(shared_datadir / "test_audio_44100.mp3")
+    data = {"audio": [audio_path]}
+    features = Features({"audio": Audio()})
+    dset = Dataset.from_dict(data, features=features)
+
+    # if torchaudio>=0.12 failed, mp3 must be decoded anyway (with librosa)
+    with patch("torchaudio.load") if torchaudio_failed else nullcontext() as load_mock, pytest.warns(
+        UserWarning, match=r"Decoding mp3 with `librosa` instead of `torchaudio`.+?"
+    ) if torchaudio_failed else nullcontext():
+        if torchaudio_failed:
+            load_mock.side_effect = RuntimeError()
+
+        item = dset[0]
+        assert item["audio"]["sampling_rate"] == 44100
+        dset = dset.cast_column("audio", Audio(sampling_rate=16000))
+        item = dset[0]
+        assert item.keys() == {"audio"}
+        assert item["audio"].keys() == {"path", "array", "sampling_rate"}
+        assert item["audio"]["path"] == audio_path
+        assert item["audio"]["array"].shape == (40125,)
+        assert item["audio"]["sampling_rate"] == 16000
+        batch = dset[:1]
+        assert batch.keys() == {"audio"}
+        assert len(batch["audio"]) == 1
+        assert batch["audio"][0].keys() == {"path", "array", "sampling_rate"}
+        assert batch["audio"][0]["path"] == audio_path
+        assert batch["audio"][0]["array"].shape == (40125,)
+        assert batch["audio"][0]["sampling_rate"] == 16000
+        column = dset["audio"]
+        assert len(column) == 1
+        assert column[0].keys() == {"path", "array", "sampling_rate"}
+        assert column[0]["path"] == audio_path
+        assert column[0]["array"].shape == (40125,)
+        assert column[0]["sampling_rate"] == 16000
+
+
 @pytest.mark.parametrize(
     "build_data",
     [

tests/utils.py

@@ -64,7 +64,16 @@ def parse_flag_from_env(key, default=False):
     find_library("sox") is None,
     reason="test requires sox OS dependency; only available on non-Windows: 'sudo apt-get install sox'",
 )
-require_torchaudio = pytest.mark.skipif(find_spec("torchaudio") is None, reason="test requires torchaudio")
+require_torchaudio = pytest.mark.skipif(
+    find_spec("torchaudio") is None
+    or version.parse(import_module("torchaudio").__version__) >= version.parse("0.12.0"),
+    reason="test requires torchaudio<0.12",
+)
+require_torchaudio_latest = pytest.mark.skipif(
+    find_spec("torchaudio") is None
+    or version.parse(import_module("torchaudio").__version__) < version.parse("0.12.0"),
+    reason="test requires torchaudio>=0.12",
+)
 
 
 def require_beam(test_case):