Add libtorchaudio cpp example (#1349)

Author: mthrok

examples/libtorchaudio/.gitignore

@@ -0,0 +1,3 @@
+build
+data/output.wav
+data/pipeline.zip

examples/libtorchaudio/CMakeLists.txt

@@ -0,0 +1,20 @@
+cmake_minimum_required(VERSION 3.5)
+
+project(libtorchaudio-cpp-example)
+
+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 Python binding")
+SET(BUILD_TORCHAUDIO_PYTHON_EXTENSION OFF CACHE BOOL "Build Python binding")
+
+find_package(Torch REQUIRED)
+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)

examples/libtorchaudio/README.md

@@ -0,0 +1,11 @@
+# Example usage: libtorchaudio
+
+This example demonstrates how you can use torchaudio's I/O features in C++ application, in addition to PyTorch's operations.
+
+To try this example, simply run `./run.sh`. This script will
+
+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 detail of the preprocessing pipeline can be found in [`create_jittable_pipeline.py`](./create_jittable_pipeline.py).

examples/libtorchaudio/create_jittable_pipeline.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+"""
+Create a data preprocess pipeline that can be run with libtorchaudio
+"""
+import os
+import argparse
+
+import torch
+import torchaudio
+
+
+class Pipeline(torch.nn.Module):
+    """Example audio process pipeline.
+
+    This example load waveform from a file then apply effects and save it to a file.
+    """
+    def __init__(self):
+        super().__init__()
+        rir, sample_rate = _load_rir()
+        self.register_buffer('rir', rir)
+        self.rir_sample_rate: int = sample_rate
+
+    def forward(self, input_path: str, output_path: str):
+        torchaudio.sox_effects.init_sox_effects()
+
+        # 1. load audio
+        waveform, sample_rate = torchaudio.load(input_path)
+
+        # 2. Add background noise
+        alpha = 0.01
+        waveform = alpha * torch.randn_like(waveform) + (1 - alpha) * waveform
+
+        # 3. Reample the RIR filter to much the audio sample rate
+        rir, _ = torchaudio.sox_effects.apply_effects_tensor(
+            self.rir, self.rir_sample_rate, effects=[["rate", str(sample_rate)]])
+        rir = rir / torch.norm(rir, p=2)
+        rir = torch.flip(rir, [1])
+
+        # 4. Apply RIR filter
+        waveform = torch.nn.functional.pad(waveform, (rir.shape[1] - 1, 0))
+        waveform = torch.nn.functional.conv1d(waveform[None, ...], rir[None, ...])[0]
+
+        # Save
+        torchaudio.save(output_path, waveform, sample_rate)
+
+
+def _create_jit_pipeline(output_path):
+    module = torch.jit.script(Pipeline())
+    print("*" * 40)
+    print("* Pipeline code")
+    print("*" * 40)
+    print()
+    print(module.code)
+    print("*" * 40)
+    module.save(output_path)
+
+
+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(
+        "--output-path",
+        default=_get_path("data", "pipeline.zip"),
+        help="Output JIT file."
+    )
+    return parser.parse_args()
+
+
+def _main():
+    args = _parse_args()
+    _create_jit_pipeline(args.output_path)
+
+
+if __name__ == '__main__':
+    _main()

examples/libtorchaudio/data/README.md

@@ -0,0 +1,5 @@
+The files in this directory are originated from [VOiCES](https://iqtlabs.github.io/voices/) dataset, which is licensed under Creative Commos BY 4.0. They are modified to fit into the tutorial.
+
+* `input.wav`: `VOiCES_devkit/source-16k/train/sp0307/Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.wav`
+
+* `rir.wav`: `VOiCES_devkit/distant-16k/room-response/rm1/impulse/Lab41-SRI-VOiCES-rm1-impulse-mc01-stu-clo.wav`

examples/libtorchaudio/data/input.wav

@@ -0,0 +1,21 @@
+#include <torch/script.h>
+
+int main(int argc, char* argv[]) {
+  if (argc !=4) {
+    std::cerr << "Usage: " << argv[0] << " <JIT_OBJECT> <INPUT_FILE> <OUTPUT_FILE>" << std::endl;
+    return -1;
+  }
+
+  torch::jit::script::Module module;
+  std::cout << "Loading module from: " << argv[0] << std::endl;
+  try {
+    module = torch::jit::load(argv[1]);
+  } catch (const c10::Error &error) {
+    std::cerr << "Failed to load the module:" << error.what() << std::endl;
+    return -1;
+  }
+
+  std::cout << "Performing the process ..." << std::endl;
+  module.forward({c10::IValue(argv[2]), c10::IValue(argv[3])});
+  std::cout << "Done." << std::endl; 
+}

examples/libtorchaudio/data/rir.wav

@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+
+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}"

examples/libtorchaudio/main.cpp

@@ -49,6 +49,7 @@ if(BUILD_LIBTORCHAUDIO)
 
   target_link_libraries(
     libtorchaudio
+    ${TORCH_LIBRARIES}
     ${TORCHAUDIO_THIRD_PARTIES}
     )