Refactor libtorchaudio example (#1486)

examples/libtorchaudio/CMakeLists.txt

@@ -14,7 +14,4 @@ message("libtorchaudio CMakeLists: ${TORCH_CXX_FLAGS}")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS}")
 
 add_subdirectory(../.. libtorchaudio)
-
-add_executable(main main.cpp)
-target_link_libraries(main "${TORCH_LIBRARIES}" "${TORCHAUDIO_LIBRARY}")
-set_property(TARGET main PROPERTY CXX_STANDARD 14)
+add_subdirectory(augmentation)

examples/libtorchaudio/README.md

-# Example usage: libtorchaudio
+# Libtorchaudio Examples
 
-This example demonstrates how you can use torchaudio's I/O features in C++ application, in addition to PyTorch's operations.
+* [Augmentation](./augmentation)
 
-To try this example, simply run `./run.sh`. This script will
+## Build
 
-1. Create an audio preprocessing pipeline with TorchScript and dump it to a file.
-2. Build the application using `libtorch` and `libtorchaudio`.
-3. Execute the preprocessing pipeline on an example audio.
+The example applications in this directory depend on `libtorch` and `libtorchaudio`.
+If you have a working `PyTorch`, you already have `libtorch`.
+Please refer to [this tutorial](https://pytorch.org/tutorials/advanced/torch_script_custom_classes.html) for the use of `libtorch` and TorchScript.
 
-The detail of the preprocessing pipeline can be found in [`create_jittable_pipeline.py`](./create_jittable_pipeline.py).
+`libtorchaudio` is the library of torchaudio's C++ components without Python component.
+It is currently not distributed, and it will be built alongside with the applications.
+
+The following commands will build `libtorchaudio` and applications.
+
+```bash
+mkdir build
+cd build
+cmake -GNinja \
+      -DCMAKE_PREFIX_PATH="$(python -c 'import torch;print(torch.utils.cmake_prefix_path)')" \
+      -DBUILD_SOX=ON \
+      -DBUILD_KALDI=OFF \
+      ..
+cmake --build .
+```
+
+For the usages of each application, refer to the corresponding application directory.

examples/libtorchaudio/augmentation/CMakeLists.txt

+add_executable(augment main.cpp)
+target_link_libraries(augment "${TORCH_LIBRARIES}" "${TORCHAUDIO_LIBRARY}")
+set_property(TARGET augment PROPERTY CXX_STANDARD 14)

examples/libtorchaudio/augmentation/README.md

+# Augmentation
+
+This example demonstrates how you can use torchaudio's I/O features and augmentations in C++ application.
+
+**NOTE**
+This example uses `"sox_io"` backend, thus does not work on Windows.
+
+## Steps
+### 1. Create augmentation pipeline TorchScript file.
+
+First, we implement our data process pipeline as a regular Python, and save it as a TorchScript object.
+We will load and execute it in our C++ application. The C++ code is found in [`main.cpp`](./main.cpp).
+
+```python
+python create_jittable_pipeline.py \
+    --rir-path "../data/rir.wav" \
+    --output-path "./pipeline.zip"
+```
+
+### 2. Build the application
+
+Please refer to [the top level README.md](../README.md)
+
+### 3. Run the application
+
+Now we run the C++ application `augment`, with the TorchScript object we created in Step.1 and an input audio file.
+
+In [the top level directory](../)
+
+```bash
+input_audio_file="./data/input.wav"
+./build/augmentation/augment ./augmentation/pipeline.zip "${input_audio_file}" "output.wav"
+```
+
+When you give a clean speech file, the output audio sounds like it's a phone conversation.
+

examples/libtorchaudio/create_jittable_pipeline.py → examples/libtorchaudio/augmentation/create_jittable_pipeline.py

@@ -14,9 +14,9 @@ class Pipeline(torch.nn.Module):
 
     This example load waveform from a file then apply effects and save it to a file.
     """
-    def __init__(self):
+    def __init__(self, rir_path: str):
         super().__init__()
-        rir, sample_rate = _load_rir()
+        rir, sample_rate = torchaudio.load(rir_path)
         self.register_buffer('rir', rir)
         self.rir_sample_rate: int = sample_rate
 
@@ -44,8 +44,8 @@ class Pipeline(torch.nn.Module):
         torchaudio.save(output_path, waveform, sample_rate)
 
 
-def _create_jit_pipeline(output_path):
-    module = torch.jit.script(Pipeline())
+def _create_jit_pipeline(rir_path, output_path):
+    module = torch.jit.script(Pipeline(rir_path))
     print("*" * 40)
     print("* Pipeline code")
     print("*" * 40)
@@ -59,16 +59,16 @@ def _get_path(*paths):
     return os.path.join(os.path.dirname(__file__), *paths)
 
 
-def _load_rir():
-    path = _get_path("data", "rir.wav")
-    return torchaudio.load(path)
-
-
 def _parse_args():
     parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--rir-path",
+        default=_get_path("..", "data", "rir.wav"),
+        help="Audio dara for room impulse response."
+    )
     parser.add_argument(
         "--output-path",
-        default=_get_path("data", "pipeline.zip"),
+        default=_get_path("pipeline.zip"),
         help="Output JIT file."
     )
     return parser.parse_args()
@@ -76,7 +76,7 @@ def _parse_args():
 
 def _main():
     args = _parse_args()
-    _create_jit_pipeline(args.output_path)
+    _create_jit_pipeline(args.rir_path, args.output_path)
 
 
 if __name__ == '__main__':

examples/libtorchaudio/main.cpp → examples/libtorchaudio/augmentation/main.cpp

@@ -7,7 +7,7 @@ int main(int argc, char* argv[]) {
   }
 
   torch::jit::script::Module module;
-  std::cout << "Loading module from: " << argv[0] << std::endl;
+  std::cout << "Loading module from: " << argv[1] << std::endl;
   try {
     module = torch::jit::load(argv[1]);
   } catch (const c10::Error &error) {

examples/libtorchaudio/run.sh → examples/libtorchaudio/build.sh

@@ -3,22 +3,14 @@
 set -eux
 
 this_dir="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
-
 build_dir="${this_dir}/build"
-data_dir="${this_dir}/data"
-jit_file="${data_dir}/pipeline.zip"
-input_file="${data_dir}/input.wav"
-output_file="${data_dir}/output.wav"
-
-cd "${this_dir}"
-python create_jittable_pipeline.py
 
 mkdir -p "${build_dir}"
 cd "${build_dir}"
+
 cmake -GNinja \
       -DCMAKE_PREFIX_PATH="$(python -c 'import torch;print(torch.utils.cmake_prefix_path)')" \
       -DBUILD_SOX=ON \
       -DBUILD_KALDI=OFF \
       ..
-cmake --build . --target main
-./main "${jit_file}" "${input_file}" "${output_file}"
+cmake --build .