Import torchaudio #1597 284bd10

Summary: Import from Github

Reviewed By: carolineechen

Differential Revision: D29518488

fbshipit-source-id: 34b3d3f2f8035bf734d047c7b6e6ec6e15ff65f1

Author: parmeet

.circleci/config.yml

@@ -52,9 +52,7 @@ commands:
           name: Adding CONDA_CHANNEL_FLAGS to BASH_ENV
           command: |
               CONDA_CHANNEL_FLAGS=""
-              if [[ "${PYTHON_VERSION}" = *3.9* ]]; then
-                echo "export CONDA_CHANNEL_FLAGS=-c=conda-forge" >> ${BASH_ENV}
-              fi
+              # formerly used to add conda-forge flags for Python 3.9, reserving the mechanism for future python upgrades
 
 binary_common: &binary_common
   parameters:

.circleci/config.yml.in

@@ -52,9 +52,7 @@ commands:
           name: Adding CONDA_CHANNEL_FLAGS to BASH_ENV
           command: |
               CONDA_CHANNEL_FLAGS=""
-              if [[ "${PYTHON_VERSION}" = *3.9* ]]; then
-                echo "export CONDA_CHANNEL_FLAGS=-c=conda-forge" >> ${BASH_ENV}
-              fi
+              # formerly used to add conda-forge flags for Python 3.9, reserving the mechanism for future python upgrades
 
 binary_common: &binary_common
   parameters:

.circleci/torchscript_bc_test/common.sh

@@ -67,5 +67,5 @@ build_master() {
     printf "* Installing torchaudio\n"
     cd "${_root_dir}" || exit 1
     git submodule update --init --recursive
-    BUILD_TRANSDUCER=1 BUILD_SOX=1 python setup.py clean install
+    BUILD_RNNT=1 BUILD_SOX=1 python setup.py clean install
 }

.circleci/torchscript_bc_test/environment.yml

@@ -1,5 +1,4 @@
 channels:
-  - conda-forge
   - defaults
 dependencies:
   - flake8

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

@@ -23,40 +23,26 @@ eval "$("${conda_dir}/bin/conda" shell.bash hook)"
 conda activate "${env_dir}"
 
 # 1. Install PyTorch
-# [2021/06/22 Temporary workaround] Disabling the original installation
-# The orignal, conda-based instartion is working for GPUs, but not for CPUs
-# For CPUs we use pip-based installation
-# if [ -z "${CUDA_VERSION:-}" ] ; then
-#     if [ "${os}" == MacOSX ] ; then
-#         cudatoolkit=''
-#     else
-#         cudatoolkit="cpuonly"
-#     fi
-# else
-#     version="$(python -c "print('.'.join(\"${CUDA_VERSION}\".split('.')[:2]))")"
-#     cudatoolkit="cudatoolkit=${version}"
-# fi
-# printf "Installing PyTorch with %s\n" "${cudatoolkit}"
-# (
-#     set -x
-#     conda install ${CONDA_CHANNEL_FLAGS:-} -y -c "pytorch-${UPLOAD_CHANNEL}" "pytorch-${UPLOAD_CHANNEL}::pytorch" ${cudatoolkit}
-# )
-
-if [ "${os}" == MacOSX ] || [ -z "${CUDA_VERSION:-}" ] ; then
-    device="cpu"
+if [ -z "${CUDA_VERSION:-}" ] ; then
+    if [ "${os}" == MacOSX ] ; then
+        cudatoolkit=''
+    else
+        cudatoolkit="cpuonly"
+    fi
 else
-    device=cu"$(python -c "print(''.join(\"${CUDA_VERSION}\".split('.')[:2]))")"
+    version="$(python -c "print('.'.join(\"${CUDA_VERSION}\".split('.')[:2]))")"
+    cudatoolkit="cudatoolkit=${version}"
 fi
-printf "Installing PyTorch with %s\n" "${device}"
+printf "Installing PyTorch with %s\n" "${cudatoolkit}"
 (
     set -x
-    pip install --pre torch==1.10.0.dev20210618 -f "https://download.pytorch.org/whl/nightly/${device}/torch_nightly.html"
+    conda install ${CONDA_CHANNEL_FLAGS:-} -y -c "pytorch-${UPLOAD_CHANNEL}" "pytorch-${UPLOAD_CHANNEL}::pytorch" ${cudatoolkit}
 )
 
 # 2. Install torchaudio
 printf "* Installing torchaudio\n"
 git submodule update --init --recursive
-BUILD_TRANSDUCER=1 BUILD_SOX=1 python setup.py install
+BUILD_RNNT=1 BUILD_SOX=1 python setup.py install
 
 # 3. Install Test tools
 printf "* Installing test tools\n"

.circleci/unittest/windows/scripts/environment.yml

@@ -1,5 +1,4 @@
 channels:
-  - conda-forge
   - defaults
 dependencies:
   - flake8

CMakeLists.txt

@@ -56,7 +56,7 @@ endif()
 # Options
 option(BUILD_SOX "Build libsox statically" OFF)
 option(BUILD_KALDI "Build kaldi statically" ON)
-option(BUILD_TRANSDUCER "Enable transducer" OFF)
+option(BUILD_RNNT "Enable RNN transducer" OFF)
 option(BUILD_LIBTORCHAUDIO "Build C++ Library" ON)
 option(BUILD_TORCHAUDIO_PYTHON_EXTENSION "Build Python extension" OFF)
 option(USE_CUDA "Enable CUDA support" OFF)

build_tools/setup_helpers/extension.py

@@ -36,9 +36,9 @@ def _get_build(var, default=False):
 
 _BUILD_SOX = False if platform.system() == 'Windows' else _get_build("BUILD_SOX")
 _BUILD_KALDI = False if platform.system() == 'Windows' else _get_build("BUILD_KALDI", True)
-_BUILD_TRANSDUCER = _get_build("BUILD_TRANSDUCER")
+_BUILD_RNNT = _get_build("BUILD_RNNT")
 _USE_ROCM = _get_build("USE_ROCM")
-_USE_CUDA = torch.cuda.is_available()
+_USE_CUDA = _get_build("USE_CUDA", torch.cuda.is_available())
 
 
 def get_ext_modules():
@@ -73,7 +73,7 @@ def build_extension(self, ext):
             f"-DPython_INCLUDE_DIR={distutils.sysconfig.get_python_inc()}",
             f"-DBUILD_SOX:BOOL={'ON' if _BUILD_SOX else 'OFF'}",
             f"-DBUILD_KALDI:BOOL={'ON' if _BUILD_KALDI else 'OFF'}",
-            f"-DBUILD_TRANSDUCER:BOOL={'ON' if _BUILD_TRANSDUCER else 'OFF'}",
+            f"-DBUILD_RNNT:BOOL={'ON' if _BUILD_RNNT else 'OFF'}",
             "-DBUILD_TORCHAUDIO_PYTHON_EXTENSION:BOOL=ON",
             "-DBUILD_LIBTORCHAUDIO:BOOL=OFF",
             f"-DUSE_ROCM:BOOL={'ON' if _USE_ROCM else 'OFF'}",

examples/libtorchaudio/CMakeLists.txt

@@ -6,7 +6,7 @@ SET(BUILD_LIBTORCHAUDIO ON CACHE BOOL "Build libtorchaudio")
 SET(BUILD_SOX ON CACHE BOOL "Build libsox into libtorchaudio")
 
 SET(BUILD_KALDI OFF CACHE BOOL "Build Kaldi into libtorchaudio")
-SET(BUILD_TRANSDUCER OFF CACHE BOOL "Build transducer into libtorchaudio")
+SET(BUILD_RNNT OFF CACHE BOOL "Build RNN transducer into libtorchaudio")
 SET(BUILD_TORCHAUDIO_PYTHON_EXTENSION OFF CACHE BOOL "Build Python binding")
 
 find_package(Torch REQUIRED)

examples/pipeline_wav2letter/main.py

@@ -11,12 +11,12 @@
 from torch.optim.lr_scheduler import ExponentialLR, ReduceLROnPlateau
 from torch.utils.data import DataLoader
 from torchaudio.datasets.utils import bg_iterator
+from torchaudio.functional import edit_distance
 from torchaudio.models.wav2letter import Wav2Letter
 
 from ctc_decoders import GreedyDecoder
 from datasets import collate_factory, split_process_librispeech
 from languagemodels import LanguageModel
-from metrics import levenshtein_distance
 from transforms import Normalize, UnsqueezeFirst
 from utils import MetricLogger, count_parameters, save_checkpoint
 
@@ -217,7 +217,7 @@ def compute_error_rates(outputs, targets, decoder, language_model, metric):
         target_print = target[i].ljust(print_length)[:print_length]
         logging.info("Target: %s    Output: %s", target_print, output_print)
 
-    cers = [levenshtein_distance(t, o) for t, o in zip(target, output)]
+    cers = [edit_distance(t, o) for t, o in zip(target, output)]
     cers = sum(cers)
     n = sum(len(t) for t in target)
     metric["batch char error"] = cers
@@ -232,7 +232,7 @@ def compute_error_rates(outputs, targets, decoder, language_model, metric):
     output = [o.split(language_model.char_space) for o in output]
     target = [t.split(language_model.char_space) for t in target]
 
-    wers = [levenshtein_distance(t, o) for t, o in zip(target, output)]
+    wers = [edit_distance(t, o) for t, o in zip(target, output)]
     wers = sum(wers)
     n = sum(len(t) for t in target)
     metric["batch word error"] = wers

examples/pipeline_wav2letter/metrics.py

@@ -15,5 +15,5 @@ if [[ "$OSTYPE" == "msys" ]]; then
     python_tag="$(echo "cp$PYTHON_VERSION" | tr -d '.')"
     "$script_dir/vc_env_helper.bat" python setup.py bdist_wheel --plat-name win_amd64 --python-tag $python_tag
 else
-    BUILD_TRANSDUCER=1 BUILD_SOX=1 python setup.py bdist_wheel
+    BUILD_RNNT=1 BUILD_SOX=1 python setup.py bdist_wheel
 fi

packaging/build_wheel.sh

@@ -231,10 +231,6 @@ setup_conda_pytorch_constraint() {
   else
     export CONDA_CHANNEL_FLAGS="${CONDA_CHANNEL_FLAGS} -c pytorch -c pytorch-test -c pytorch-nightly"
   fi
-  # Some dependencies for Python 3.9 are only on conda-forge
-  if [[ "${PYTHON_VERSION}" = "3.9" ]]; then
-    export CONDA_CHANNEL_FLAGS="${CONDA_CHANNEL_FLAGS} -c conda-forge"
-  fi
   if [[ "$CU_VERSION" == cpu ]]; then
     export CONDA_PYTORCH_BUILD_CONSTRAINT="- pytorch==$PYTORCH_VERSION${PYTORCH_VERSION_SUFFIX}"
     export CONDA_PYTORCH_CONSTRAINT="- pytorch==$PYTORCH_VERSION"

packaging/pkg_helpers.bash

@@ -1,4 +1,4 @@
 #!/usr/bin/env bash
 set -ex
 
-BUILD_TRANSDUCER=1 BUILD_SOX=1 python setup.py install --single-version-externally-managed --record=record.txt
+BUILD_RNNT=1 BUILD_SOX=1 python setup.py install --single-version-externally-managed --record=record.txt

packaging/torchaudio/build.sh

@@ -382,6 +382,46 @@ def test_phase_vocoder_shape(self, rate, test_pseudo_complex):
         output_shape = (torch.view_as_complex(spec_stretch) if test_pseudo_complex else spec_stretch).shape
         assert output_shape == expected_shape
 
+    @parameterized.expand(
+        [
+            # words
+            ["", "", 0],  # equal
+            ["abc", "abc", 0],
+            ["ᑌᑎIᑕO", "ᑌᑎIᑕO", 0],
+
+            ["abc", "", 3],  # deletion
+            ["aa", "aaa", 1],
+            ["aaa", "aa", 1],
+            ["ᑌᑎI", "ᑌᑎIᑕO", 2],
+
+            ["aaa", "aba", 1],  # substitution
+            ["aba", "aaa", 1],
+            ["aba", "   ", 3],
+
+            ["abc", "bcd", 2],  # mix deletion and substitution
+            ["0ᑌᑎI", "ᑌᑎIᑕO", 3],
+
+            # sentences
+            [["hello", "", "Tᕮ᙭T"], ["hello", "", "Tᕮ᙭T"], 0],  # equal
+            [[], [], 0],
+
+            [["hello", "world"], ["hello", "world", "!"], 1],  # deletion
+            [["hello", "world"], ["world"], 1],
+            [["hello", "world"], [], 2],
+
+            [["Tᕮ᙭T", ], ["world"], 1],  # substitution
+            [["Tᕮ᙭T", "XD"], ["world", "hello"], 2],
+            [["", "XD"], ["world", ""], 2],
+            ["aba", "   ", 3],
+
+            [["hello", "world"], ["world", "hello", "!"], 2],  # mix deletion and substitution
+            [["Tᕮ᙭T", "world", "LOL", "XD"], ["world", "hello", "ʕ•́ᴥ•̀ʔっ"], 3],
+        ]
+    )
+    def test_simple_case_edit_distance(self, seq1, seq2, distance):
+        assert F.edit_distance(seq1, seq2) == distance
+        assert F.edit_distance(seq2, seq1) == distance
+
 
 class FunctionalCPUOnly(TestBaseMixin):
     def test_create_fb_matrix_no_warning_high_n_freq(self):

test/torchaudio_unittest/functional/functional_impl.py

@@ -1,10 +1,10 @@
 import torch
 from .autograd_impl import Autograd
 from torchaudio_unittest import common_utils
-from .utils import skipIfNoTransducer
+from .utils import skipIfNoRNNT
 
 
-@skipIfNoTransducer
+@skipIfNoRNNT
 class TestAutograd(Autograd, common_utils.PytorchTestCase):
     dtype = torch.float32
     device = torch.device('cpu')

test/torchaudio_unittest/rnnt/autograd_cpu_test.py

@@ -1,10 +1,10 @@
 import torch
 from .autograd_impl import Autograd
 from torchaudio_unittest import common_utils
-from .utils import skipIfNoTransducer
+from .utils import skipIfNoRNNT
 
 
-@skipIfNoTransducer
+@skipIfNoRNNT
 @common_utils.skipIfNoCuda
 class TestAutograd(Autograd, common_utils.PytorchTestCase):
     dtype = torch.float32

test/torchaudio_unittest/rnnt/autograd_cuda_test.py

@@ -1,9 +1,9 @@
 import torch
 from torchaudio_unittest import common_utils
-from .utils import skipIfNoTransducer
+from .utils import skipIfNoRNNT
 from .rnnt_loss_impl import RNNTLossTest
 
 
-@skipIfNoTransducer
+@skipIfNoRNNT
 class TestRNNTLoss(RNNTLossTest, common_utils.PytorchTestCase):
     device = torch.device('cpu')

test/torchaudio_unittest/rnnt/rnnt_loss_cpu_test.py

@@ -1,10 +1,10 @@
 import torch
 from .rnnt_loss_impl import RNNTLossTest
 from torchaudio_unittest import common_utils
-from .utils import skipIfNoTransducer
+from .utils import skipIfNoRNNT
 
 
-@skipIfNoTransducer
+@skipIfNoRNNT
 @common_utils.skipIfNoCuda
 class TestRNNTLoss(RNNTLossTest, common_utils.PytorchTestCase):
     device = torch.device('cuda')

test/torchaudio_unittest/rnnt/rnnt_loss_cuda_test.py

@@ -1,10 +1,10 @@
 import torch
 
 from torchaudio_unittest.common_utils import PytorchTestCase
-from .utils import skipIfNoTransducer
+from .utils import skipIfNoRNNT
 from .torchscript_consistency_impl import RNNTLossTorchscript
 
 
-@skipIfNoTransducer
+@skipIfNoRNNT
 class TestRNNTLoss(RNNTLossTorchscript, PytorchTestCase):
     device = torch.device('cpu')

test/torchaudio_unittest/rnnt/torchscript_consistency_cpu_test.py

@@ -1,11 +1,11 @@
 import torch
 
 from torchaudio_unittest.common_utils import PytorchTestCase, skipIfNoCuda
-from .utils import skipIfNoTransducer
+from .utils import skipIfNoRNNT
 from .torchscript_consistency_impl import RNNTLossTorchscript
 
 
-@skipIfNoTransducer
+@skipIfNoRNNT
 @skipIfNoCuda
 class TestRNNTLoss(RNNTLossTorchscript, PytorchTestCase):
     device = torch.device('cuda')

test/torchaudio_unittest/rnnt/torchscript_consistency_cuda_test.py

@@ -31,6 +31,7 @@ def compute_with_pytorch_transducer(data, reuse_logits_for_grads=False):
         blank=data["blank"],
         fused_log_softmax=data.get("fused_log_softmax", True),
         reuse_logits_for_grads=reuse_logits_for_grads,
+        reduction="none",
     )(
         logits=data["logits"],
         logit_lengths=data["logit_lengths"],
@@ -442,7 +443,7 @@ def grad_hook(grad):
     return data
 
 
-def skipIfNoTransducer(test_item):
+def skipIfNoRNNT(test_item):
     try:
         torch.ops.torchaudio.rnnt_loss
         return test_item

test/torchaudio_unittest/rnnt/utils.py

@@ -19,7 +19,7 @@ ExternalProject_Add(mad
   DOWNLOAD_DIR ${ARCHIVE_DIR}
   URL https://downloads.sourceforge.net/project/mad/libmad/0.15.1b/libmad-0.15.1b.tar.gz
   URL_HASH SHA256=bbfac3ed6bfbc2823d3775ebb931087371e142bb0e9bb1bee51a76a6e0078690
-  PATCH_COMMAND patch < ${CMAKE_CURRENT_SOURCE_DIR}/patch/libmad.patch && cp ${CMAKE_CURRENT_SOURCE_DIR}/patch/config.guess ${CMAKE_CURRENT_BINARY_DIR}/src/mad/config.guess
+  PATCH_COMMAND patch < ${CMAKE_CURRENT_SOURCE_DIR}/patch/libmad.patch && cp ${CMAKE_CURRENT_SOURCE_DIR}/patch/config.guess ${CMAKE_CURRENT_BINARY_DIR}/src/mad/config.guess && cp ${CMAKE_CURRENT_SOURCE_DIR}/patch/config.sub ${CMAKE_CURRENT_BINARY_DIR}/src/mad/config.sub
   CONFIGURE_COMMAND ${CMAKE_COMMAND} -E env ${envs} ${CMAKE_CURRENT_BINARY_DIR}/src/mad/configure ${COMMON_ARGS}
   DOWNLOAD_NO_PROGRESS ON
   LOG_DOWNLOAD ON
@@ -53,7 +53,7 @@ ExternalProject_Add(lame
   URL https://downloads.sourceforge.net/project/lame/lame/3.99/lame-3.99.5.tar.gz
   URL_HASH SHA256=24346b4158e4af3bd9f2e194bb23eb473c75fb7377011523353196b19b9a23ff
   CONFIGURE_COMMAND ${CMAKE_COMMAND} -E env ${envs} ${CMAKE_CURRENT_BINARY_DIR}/src/lame/configure ${COMMON_ARGS} --enable-nasm
-  PATCH_COMMAND cp ${CMAKE_CURRENT_SOURCE_DIR}/patch/config.guess ${CMAKE_CURRENT_BINARY_DIR}/src/lame/config.guess
+  PATCH_COMMAND cp ${CMAKE_CURRENT_SOURCE_DIR}/patch/config.guess ${CMAKE_CURRENT_BINARY_DIR}/src/lame/config.guess && cp ${CMAKE_CURRENT_SOURCE_DIR}/patch/config.sub ${CMAKE_CURRENT_BINARY_DIR}/src/lame/config.sub
   DOWNLOAD_NO_PROGRESS ON
   LOG_DOWNLOAD ON
   LOG_UPDATE ON