Merge branch 'main' into remove_backends

Author: samanklesaria

.github/scripts/unittest-linux/install.sh

@@ -24,41 +24,7 @@ esac
 conda create -n ci -y python="${PYTHON_VERSION}"
 conda activate ci
 
-# 1. Install PyTorch
-# if [ -z "${CUDA_VERSION:-}" ] ; then
-#     if [ "${os}" == MacOSX ] ; then
-#         cudatoolkit=''
-#     else
-#         cudatoolkit="cpuonly"
-#     fi
-#     version="cpu"
-# else
-#     version="$(python -c "print('.'.join(\"${CUDA_VERSION}\".split('.')[:2]))")"
-#     export CUDATOOLKIT_CHANNEL="nvidia"
-#     cudatoolkit="pytorch-cuda=${version}"
-# fi
-
-# printf "Installing PyTorch with %s\n" "${cudatoolkit}"
-# (
-#     if [ "${os}" == MacOSX ] ; then
-#       # TODO: this can be removed as soon as linking issue could be resolved
-#       #  see https://github.com/pytorch/pytorch/issues/62424 from details
-#       MKL_CONSTRAINT='mkl==2021.2.0'
-#       pytorch_build=pytorch
-#     else
-#       MKL_CONSTRAINT=''
-#       pytorch_build="pytorch[build="*${version}*"]"
-#     fi
-#     set -x
-
-#     if [[ -z "$cudatoolkit" ]]; then
-#         conda install ${CONDA_CHANNEL_FLAGS:-} -y -c "pytorch-${UPLOAD_CHANNEL}" $MKL_CONSTRAINT "pytorch-${UPLOAD_CHANNEL}::${pytorch_build}"
-#     else
-#         conda install pytorch ${cudatoolkit} ${CONDA_CHANNEL_FLAGS:-} -y -c "pytorch-${UPLOAD_CHANNEL}" -c nvidia  $MKL_CONSTRAINT
-#     fi
-# )
-
-export GPU_ARCH_TYPE="cpu"  # TODO change this
+export GPU_ARCH_TYPE="cpu"
 
 case $GPU_ARCH_TYPE in
   cpu)
@@ -89,22 +55,4 @@ printf "* Installing test tools\n"
 # On this CI, for whatever reason, we're only able to install ffmpeg 4.
 conda install -y "ffmpeg<5"
 
-NUMBA_DEV_CHANNEL=""
-if [[ "$(python --version)" = *3.9* || "$(python --version)" = *3.10* ]]; then
-    # Numba isn't available for Python 3.9 and 3.10 except on the numba dev channel and building from source fails
-    # See https://github.com/librosa/librosa/issues/1270#issuecomment-759065048
-    NUMBA_DEV_CHANNEL="-c numba/label/dev"
-fi
-(
-    set -x
-    conda install -y -c conda-forge ${NUMBA_DEV_CHANNEL} libvorbis parameterized 'requests>=2.20'
-    pip install SoundFile coverage pytest pytest-cov scipy expecttest unidecode inflect Pillow sentencepiece pytorch-lightning 'protobuf<4.21.0' demucs tinytag pyroomacoustics flashlight-text git+https://github.com/kpu/kenlm
-
-    # TODO: might be better to fix the single call to `pip install` above
-    pip install pillow scipy "numpy>=1.26"
-)
-# Install fairseq
-git clone https://github.com/pytorch/fairseq
-cd fairseq
-git checkout e47a4c8
-pip install .
+pip3 install parameterized requests coverage pytest pytest-cov scipy numpy expecttest

.github/scripts/unittest-linux/run_test.sh

@@ -29,6 +29,10 @@ fi
 )
 
 (
+    export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_CTC_DECODER=true
+    export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_unidecode=true
+    export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_inflect=true
+    export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_pytorch_lightning=true
     cd test
-    pytest torchaudio_unittest -k "not backend and not (sox and not compatibility) and not /io/ and not prototype and not ffmpeg and not fairseq and not hdemucs and not (torchscript and rnnt) and not torchscript_consistency"
+    pytest torchaudio_unittest -k "not backend and not /io/ and not prototype and not ffmpeg and not fairseq and not hdemucs and not (torchscript and rnnt) and not torchscript_consistency"
 )

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

@@ -40,6 +40,7 @@ jobs:
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX_DECODER=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_SOX_ENCODER=true
         export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_FFMPEG=true
+        export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_MOD_demucs=true
         # Avoid reproducibility errors with CUBLAS: https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility
         export CUBLAS_WORKSPACE_CONFIG=:4096:8
 
@@ -73,34 +74,8 @@ jobs:
 
         echo "::endgroup::"
         echo "::group::Install other Dependencies"
-        # conda install \
-        #   --quiet --yes \
-        #   -c conda-forge \
-        #   -c numba/label/dev \
-        #   libvorbis parameterized 'requests>=2.20'
-        # pip3 install --progress-bar off \
-        #   SoundFile \
-        #   coverage \
-        #   pytest \
-        #   pytest-cov \
-        #   scipy \
-        #   transformers \
-        #   expecttest \
-        #   unidecode \
-        #   inflect \
-        #   Pillow \
-        #   sentencepiece \
-        #   pytorch-lightning \
-        #   'protobuf<4.21.0' \
-        #   demucs \
-        #   tinytag \
-        #   flashlight-text \
-        #   git+https://github.com/kpu/kenlm/ \
-        #   git+https://github.com/pytorch/fairseq.git@e47a4c8
         
-        pip3 install parameterized requests
-        pip3 install SoundFile coverage pytest pytest-cov scipy expecttest unidecode inflect Pillow sentencepiece pytorch-lightning 'protobuf<4.21.0' demucs tinytag 
-        pip3 install pillow scipy "numpy>=1.26"
+        pip3 install parameterized requests coverage pytest pytest-cov scipy numpy expecttest 
 
         echo "::endgroup::"
         echo "::group::Run tests"

docs/requirements-tutorials.txt

@@ -1,11 +1 @@
 IPython
-deep-phonemizer
-boto3
-cython
-pandas
-librosa==0.10.0
-sentencepiece
-pandoc
-mir_eval
-pesq
-pystoi

examples/tutorials/audio_feature_extractions_tutorial.py

@@ -25,7 +25,6 @@
 print(torch.__version__)
 print(torchaudio.__version__)
 
-import librosa
 import matplotlib.pyplot as plt
 
 ######################################################################

test/conftest.py

@@ -1,4 +1,23 @@
 import sys
 from pathlib import Path
 
+# Note: [TorchCodec test dependency mocking hack]
+# We are adding the `test/` directory to the system path. This causes the
+# `tests/torchcodec` folder to be importable, and in particular, this makes it
+# possible to mock torchcodec utilities. E.g. executing:
+#
+# ```
+# from torchcodec.decoders import AudioDecoder
+# ```
+# directly or indirectly when running the tests will effectively be loading the
+# mocked `AudioDecoder` implemented in `test/torchcodec/decoders.py`, which
+# relies on scipy instead of relying on torchcodec.
+#
+# So whenever `torchaudio.load()` is called from within the tests, it's the
+# mocked scipy `AudioDecoder` that gets used.  Ultimately, this allows us *not*
+# to add torchcodec as a test dependency of torchaudio: we can just rely on
+# scipy.
+#
+# This is VERY hacky and ideally we should implement a more robust way to mock
+# torchcodec.
 sys.path.append(str(Path(__file__).parent.resolve()))

test/torchaudio_unittest/test_load_save_torchcodec.py

@@ -12,8 +12,13 @@
 from torchaudio import load_with_torchcodec, save_with_torchcodec
 from torchaudio_unittest.common_utils import get_asset_path
 
-# Now, load/save_torchcodec are the same as torchaudio.load/save, so
-# there is no need to test this.
+# These tests were ran when `torchaudio.load()` and `torchaudio.save()` were
+# still relying on their previous backends (ffmpeg, sox, soundfile). We needed
+# to validate that the newly introduced `load_with_torchcodec()` and
+# save_with_torchcodec() were matching their results.
+# From 2.9, `load()` and `save()` now internally rely on `load_with_torchcodec()` and
+# `save_with_torchcodec()` directly, so these tests are now redundant and we
+# skip them unconditionally.
 pytest.skip(allow_module_level=True)
 
 def get_ffmpeg_version():

test/torchaudio_unittest/utils/sox_utils_test.py

@@ -1,6 +1,9 @@
 import torchaudio_unittest.common_utils.wav_utils as wav_utils
 from types import SimpleNamespace
 
+# See corresponding [TorchCodec test dependency mocking hack] note in
+# conftest.py
+
 class AudioDecoder:
     def __init__(self, uri):
         self.uri = uri

test/torchcodec/decoders.py

@@ -1,6 +1,9 @@
 import torchaudio_unittest.common_utils.wav_utils as wav_utils
 from types import SimpleNamespace
 
+# See corresponding [TorchCodec test dependency mocking hack] note in
+# conftest.py
+
 class AudioEncoder:
     def __init__(self, data, sample_rate):
         self.data = data