[src&install] Drop librosa and refactor requirements files (#386)

Fix #378 (switch to Julius)

Co-authored-by: Pariente Manuel <[email protected]>

.github/workflows/test_asteroid_package.yml

@@ -33,7 +33,6 @@ jobs:
       run: |
         python -m pip install --upgrade --user pip --quiet
         python -m pip install numpy Cython --upgrade-strategy only-if-needed --quiet
-        python -m pip install pytest coverage codecov --upgrade-strategy only-if-needed --quiet
         if [ $TORCH_INSTALL == "1.6.0" ]; then
           INSTALL="torch==1.6.0 torchaudio==0.6.0"
         elif [ $TORCH_INSTALL == "1.7.0" ]; then
@@ -42,7 +41,7 @@ jobs:
           INSTALL="--pre torch torchaudio -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html"
         fi
         python -m pip install $INSTALL
-        python -m pip install -r requirements.txt --quiet
+        python -m pip install -r requirements/dev.txt --upgrade-strategy only-if-needed --quiet
         python --version
         pip --version
         python -m pip list

.github/workflows/test_docs.yml

@@ -38,8 +38,7 @@ jobs:
       run: |
         python -m pip install --upgrade --user pip --quiet
         python -m pip install numpy Cython --upgrade-strategy only-if-needed --quiet
-        python -m pip install -r requirements.txt --quiet
-        python -m pip install -r docs/requirements.txt
+        python -m pip install -r requirements/docs.txt
         python --version
         pip --version
         python -m pip list

.github/workflows/test_model_consistency.yml

@@ -39,7 +39,6 @@ jobs:
       run: |
         python -m pip install --upgrade --user pip --quiet
         python -m pip install numpy Cython --upgrade-strategy only-if-needed --quiet
-        python -m pip install pytest coverage codecov --upgrade-strategy only-if-needed --quiet
         if [ $TORCH_INSTALL == "1.6.0" ]; then
           INSTALL="torch==1.6.0 torchaudio==0.6.0"
         elif [ $TORCH_INSTALL == "1.7.0" ]; then
@@ -48,7 +47,7 @@ jobs:
           INSTALL="--pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html"
         fi
         python -m pip install $INSTALL
-        python -m pip install -r requirements.txt --quiet
+        python -m pip install -r requirements/dev.txt --quiet
         python --version
         pip --version
         python -m pip list

.github/workflows/test_torch_hub.yml

@@ -19,11 +19,25 @@ jobs:
       with:
         python-version: 3.8
 
+    # Timeout: https://stackoverflow.com/a/59076067/4521646
+    timeout-minutes: 10
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+
+    - name: Install libnsdfile
+      run: |
+        sudo apt update
+        sudo apt install libsndfile1-dev libsndfile1
+
     - name: Install python dependencies
       run: |
         python -m pip install --upgrade pip --quiet
-        python -m pip install torch --quiet
-        python -m pip install numpy scipy asteroid-filterbanks requests filelock --quiet
+        python -m pip install numpy Cython --upgrade-strategy only-if-needed --quiet
+        python -m pip install -r requirements/torchhub.txt --quiet
         python --version
         pip --version
         python -m pip list

CONTRIBUTING.md

@@ -15,7 +15,7 @@ To do that, install asteroid in develop mode either with pip
 
 To avoid formatting roundtrips in PRs, Asteroid relies on [`black`](https://github.com/psf/black)
 and [`pre-commit-hooks`](https://github.com/pre-commit/pre-commit-hooks) to handle formatting
-for us. You'll need to install `requirements.txt` and install git hooks with
+for us. You'll need to install `requirements/dev.txt` and install git hooks with
 `pre-commit install`.
 
 Here is a summary:
@@ -24,7 +24,7 @@ Here is a summary:
 ### Install
 git clone your_fork_url
 cd asteroid
-pip install -r requirements.txt
+pip install -r requirements/dev.txt
 pip install -e .
 pre-commit install  # To run black before commit
 

asteroid/separate.py

@@ -1,7 +1,8 @@
 import os
+import warnings
 import torch
 import numpy as np
-import warnings
+import soundfile as sf
 
 try:
     from typing import Protocol
@@ -120,50 +121,90 @@ def file_separate(
     **kwargs,
 ) -> None:
     """Filename interface to `separate`."""
-    import soundfile as sf
 
     if not hasattr(model, "sample_rate"):
         raise TypeError(
             f"This function requires your model ({type(model).__name__}) to have a "
             "'sample_rate' attribute. See `BaseModel.sample_rate` for details."
         )
 
+    # Estimates will be saved as filename_est1.wav etc...
+    base, _ = os.path.splitext(filename)
+    if output_dir is not None:
+        base = os.path.join(output_dir, os.path.basename(base))
+    save_name_template = base + "_est{}.wav"
+
+    # Bail out early if an estimate file already exists and we shall not overwrite.
+    est1_filename = save_name_template.format(1)
+    if os.path.isfile(est1_filename) and not force_overwrite:
+        warnings.warn(
+            f"File {est1_filename} already exists, pass `force_overwrite=True` to overwrite it",
+            UserWarning,
+        )
+        return
+
     # SoundFile wav shape: [time, n_chan]
-    wav, fs = sf.read(filename, dtype="float32", always_2d=True)
+    wav, fs = _load_audio(filename)
     if wav.shape[-1] > 1:
         warnings.warn(
             f"Received multichannel signal with {wav.shape[-1]} signals, "
             f"using the first channel only."
         )
     # FIXME: support only single-channel files for now.
-    if fs != model.sample_rate:
-        if resample:
-            from librosa import resample
-
-            wav = resample(wav[:, 0], orig_sr=fs, target_sr=model.sample_rate)[:, None]
-        else:
-            raise RuntimeError(
-                f"Received a signal with a sampling rate of {fs}Hz for a model "
-                f"of {model.sample_rate}Hz. You can pass `resample=True` to resample automatically."
-            )
+    if resample:
+        wav = _resample(wav[:, 0], orig_sr=fs, target_sr=int(model.sample_rate))[:, None]
+    elif fs != model.sample_rate:
+        raise RuntimeError(
+            f"Received a signal with a sampling rate of {fs}Hz for a model "
+            f"of {model.sample_rate}Hz. You can pass `resample=True` to resample automatically."
+        )
     # Pass wav as [batch, n_chan, time]; here: [1, 1, time]
     wav = wav[:, 0][None, None]
-    (to_save,) = numpy_separate(model, wav, **kwargs)
+    (est_srcs,) = numpy_separate(model, wav, **kwargs)
+    # Resample to original sr
+    est_srcs = [
+        _resample(est_src, orig_sr=int(model.sample_rate), target_sr=fs) for est_src in est_srcs
+    ]
 
     # Save wav files to filename_est1.wav etc...
-    for src_idx, est_src in enumerate(to_save):
-        base = ".".join(filename.split(".")[:-1])
-        save_name = base + "_est{}.".format(src_idx + 1) + filename.split(".")[-1]
-        if output_dir is not None:
-            save_name = os.path.join(output_dir, save_name.split("/")[-1])
-        if os.path.isfile(save_name) and not force_overwrite:
-            warnings.warn(
-                f"File {save_name} already exists, pass `force_overwrite=True` to overwrite it",
-                UserWarning,
-            )
-            return
-        if fs != model.sample_rate:
-            from librosa import resample
-
-            est_src = resample(est_src, orig_sr=model.sample_rate, target_sr=fs)
-        sf.write(save_name, est_src, fs)
+    for src_idx, est_src in enumerate(est_srcs, 1):
+        sf.write(save_name_template.format(src_idx), est_src, fs)
+
+
+def _resample(wav, orig_sr, target_sr, _resamplers={}):
+    from julius import ResampleFrac
+
+    if orig_sr == target_sr:
+        return wav
+
+    # Cache ResampleFrac instance to speed up resampling if we're repeatedly
+    # resampling between the same two sample rates.
+    try:
+        resampler = _resamplers[(orig_sr, target_sr)]
+    except KeyError:
+        resampler = _resamplers[(orig_sr, target_sr)] = ResampleFrac(orig_sr, target_sr)
+
+    return resampler(torch.from_numpy(wav)).numpy()
+
+
+def _load_audio(filename):
+    try:
+        return sf.read(filename, dtype="float32", always_2d=True)
+    except Exception as sf_err:
+        # If soundfile fails to load the file, try with librosa next, which uses
+        # the 'audioread' library to support a wide range of audio formats.
+        # We try with soundfile first because librosa takes a long time to import.
+        try:
+            import librosa
+        except ModuleNotFoundError:
+            raise RuntimeError(
+                f"Could not load file {filename!r} with soundfile. "
+                "Install 'librosa' to be able to load more file types."
+            ) from sf_err
+
+        wav, sr = librosa.load(filename, dtype="float32", sr=None)
+        # Always return wav of shape [time, n_chan]
+        if wav.ndim == 1:
+            return wav[:, None], sr
+        else:
+            return wav.T, sr

docs/source/_templates/theme_variables.jinja

@@ -9,7 +9,7 @@
   'tutorials': '',
   'previous_pytorch_versions': 'https://pytorch-lightning.rtfd.io/en/latest/',
   'home': 'https://mpariente.github.io/asteroid/',
-  'get_started': 'https://colab.research.google.com/github/mpariente/asteroid/blob/master/notebooks/01_AsteroidGettingStarted.ipynb',
+  'get_started': 'https://colab.research.google.com/github/mpariente/asteroid/blob/master/notebooks/00_GettingStarted.ipynb',
   'features': 'https://mpariente.github.io/asteroid/',
   'blog': 'https://mpariente.github.io/asteroid/',
   'resources': 'https://mpariente.github.io/asteroid/',

docs/source/conf.py

@@ -322,7 +322,7 @@ def setup(app):
 # https://stackoverflow.com/questions/15889621/sphinx-how-to-exclude-imports-in-automodule
 
 MOCK_REQUIRE_PACKAGES = []
-with open(os.path.join(PATH_ROOT, "requirements.txt"), "r") as fp:
+with open(os.path.join(PATH_ROOT, "requirements", "docs.txt"), "r") as fp:
     for ln in fp.readlines():
         found = [ln.index(ch) for ch in list(",=<>#") if ch in ln]
         pkg = ln[: min(found)] if found else ln

docs/source/index.rst

@@ -23,7 +23,7 @@ large range of datasets and architectures, and a set of recipes to reproduce som
    :maxdepth: 1
    :caption: Notebooks and Tutorials
 
-   Getting started with Asteroid <http://colab.research.google.com/github/asteroid-team/asteroid/blob/master/notebooks/00_AsteroidGettingStarted.ipynb>
+   Getting started with Asteroid <http://colab.research.google.com/github/asteroid-team/asteroid/blob/master/notebooks/00_GettingStarted.ipynb>
    Introduction and Overview <http://colab.research.google.com/github/asteroid-team/asteroid/blob/master/notebooks/01_AsteroidOverview.ipynb>
    Understanding the Filterbank API <http://colab.research.google.com/github/asteroid-team/asteroid/blob/master/notebooks/02_Filterbank.ipynb>
    Our PITLossWrapper explained <http://colab.research.google.com/github/asteroid-team/asteroid/blob/master/notebooks/03_PITLossWrapper.ipynb>

egs/wham/DPTNet/utils/prepare_python_env.sh

@@ -23,4 +23,4 @@ fi
 $pip_path install soundfile
 $pip_path install -e $asteroid_root
 #$pip_path install ${asteroid_root}/\[""evaluate""\]
-echo -e "\nAsteroid has been installed in editable mode. Feel free to apply your changes !"
+echo -e "\nAsteroid has been installed in editable mode. Feel free to apply your changes !"