Import torchaudio #1711 2c11582

Summary: as titled

Reviewed By: carolineechen

Differential Revision: D30449599

fbshipit-source-id: 7b3faaf6d7dbfa2e5ca9c263554b18e7364be77e

Author: nateanl

CMakeLists.txt

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

README.md

@@ -82,6 +82,7 @@ conda install -y -c pytorch-nightly torchaudio
 
 The build process builds libsox and some codecs that torchaudio need to link to. This is achieved by setting the environment variable `BUILD_SOX=1`.
 The build process will fetch and build libmad, lame, flac, vorbis, opus, and libsox before building extension. This process requires `cmake` and `pkg-config`.
+The build process also builds the RNN transducer loss. This functionality can be disabled by setting the environment variable `BUILD_RNNT=0`.
 
 ```bash
 # Linux

build_tools/setup_helpers/extension.py

@@ -36,7 +36,7 @@ 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_RNNT = _get_build("BUILD_RNNT")
+_BUILD_RNNT = _get_build("BUILD_RNNT", True)
 _USE_ROCM = _get_build("USE_ROCM")
 _USE_CUDA = _get_build("USE_CUDA", torch.cuda.is_available())
 

docs/source/functional.rst

@@ -211,6 +211,11 @@ vad
 
 .. autofunction:: spectrogram
 
+:hidden:`inverse_spectrogram`
+-----------------------------
+
+.. autofunction:: inverse_spectrogram
+
 :hidden:`griffinlim`
 --------------------
 
@@ -251,6 +256,14 @@ vad
 
 .. autofunction:: spectral_centroid
 
+:hidden:`Loss`
+~~~~~~~~~~~~~~
+
+rnnt_loss
+---------
+
+.. autofunction:: rnnt_loss
+
 References
 ~~~~~~~~~~
 

docs/source/index.rst

@@ -39,8 +39,6 @@ The :mod:`torchaudio` package consists of I/O, popular datasets and common audio
    compliance.kaldi
    kaldi_io
    utils
-   rnnt_loss
-   tacotron2
 
 
 .. toctree::

docs/source/models.rst

@@ -25,6 +25,24 @@ DeepSpeech
   .. automethod:: forward
 
 
+Tacotron2
+~~~~~~~~~
+
+.. autoclass:: Tacotron2
+
+  .. automethod:: forward
+
+  .. automethod:: infer
+
+Factory Functions
+-----------------
+
+tacotron2
+---------
+
+.. autofunction:: tacotron2
+
+
 Wav2Letter
 ~~~~~~~~~~
 

docs/source/rnnt_loss.rst

@@ -16,6 +16,13 @@ Transforms are common audio transforms. They can be chained together using :clas
 
   .. automethod:: forward
 
+:hidden:`InverseSpectrogram`
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: InverseSpectrogram
+
+  .. automethod:: forward
+
 :hidden:`GriffinLim`
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
@@ -164,6 +171,12 @@ Transforms are common audio transforms. They can be chained together using :clas
 
   .. automethod:: forward
 
+:hidden:`RNNTLoss`
+~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: RNNTLoss
+
+  .. automethod:: forward
 
 References
 ~~~~~~~~~~

docs/source/transforms.rst

@@ -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_RNNT OFF CACHE BOOL "Build RNN transducer into libtorchaudio")
+SET(BUILD_RNNT ON CACHE BOOL "Build RNN transducer into libtorchaudio")
 SET(BUILD_TORCHAUDIO_PYTHON_EXTENSION OFF CACHE BOOL "Build Python binding")
 
 find_package(Torch REQUIRED)

examples/libtorchaudio/CMakeLists.txt

@@ -22,6 +22,7 @@ cmake -GNinja \
       -DCMAKE_PREFIX_PATH="$(python -c 'import torch;print(torch.utils.cmake_prefix_path)')" \
       -DBUILD_SOX=ON \
       -DBUILD_KALDI=OFF \
+      -DBUILD_RNNT=ON \
       ..
 cmake --build .
 ```