basic transforms

vctk dataset

add label processing

added tests

chunk dataset

simple caching scheme

caching fixed

vctk downsample

yesno dataset

.gitignore

@@ -5,6 +5,7 @@ __pycache__/
 
 # C extensions
 *.so
+_ext/
 
 # Distribution / packaging
 .Python
@@ -68,7 +69,7 @@ docs/_build/
 target/
 
 # Jupyter Notebook
-.ipynb_checkpoints
+.ipynb_checkpoints/
 
 # pyenv
 .python-version
@@ -100,5 +101,5 @@ ENV/
 # mypy
 .mypy_cache/
 
-# Jupyter Notebooks
-.ipynb_checkpoints/
+# Generated Files
+test/assets/sinewave.wav

test/test.py

@@ -2,17 +2,32 @@
 import torch
 import torch.nn as nn
 import torchaudio
+import math
 
-x, sample_rate = torchaudio.load("steam-train-whistle-daniel_simon.mp3")
+steam_train = "assets/steam-train-whistle-daniel_simon.mp3"
+
+x, sample_rate = torchaudio.load(steam_train)
 print(sample_rate)
 print(x.size())
 print(x[10000])
 print(x.min(), x.max())
 print(x.mean(), x.std())
 
-x, sample_rate = torchaudio.load("steam-train-whistle-daniel_simon.mp3",
+x, sample_rate = torchaudio.load(steam_train,
                                  out=torch.LongTensor())
 print(sample_rate)
 print(x.size())
 print(x[10000])
 print(x.min(), x.max())
+
+sine_wave = "assets/sinewave.wav"
+sr = 16000
+freq = 440
+volume = 0.3
+
+y = (torch.cos(2*math.pi*torch.arange(0, 4*sr) * freq/sr)).float()
+y.unsqueeze_(1)
+# y is between -1 and 1, so must scale
+y = (y*volume*2**31).long()
+torchaudio.save(sine_wave, y, sr)
+print(y.min(), y.max())

test/test_transforms.py

+import torch
+import torchaudio
+import torchaudio.transforms as transforms
+import numpy as np
+import unittest
+
+STEAM_TRAIN = "assets/steam-train-whistle-daniel_simon.mp3"
+
+class Tester(unittest.TestCase):
+
+    sr = 16000
+    freq = 440
+    volume = 0.3
+    sig = (torch.cos(2*np.pi*torch.arange(0, 4*sr) * freq/sr)).float()
+    sig.unsqueeze_(1)
+    sig = (sig*volume*2**31).long()
+
+    def test_scale(self):
+
+        audio_orig = self.sig.clone()
+        result = transforms.Scale()(audio_orig)
+        self.assertTrue(result.min() >= -1. and result.max() <= 1.,
+                        "min: {}, max: {}".format(result.min(), result.max()))
+
+        maxminmax = np.abs([audio_orig.min(), audio_orig.max()]).max().astype(np.float)
+        result = transforms.Scale(factor=maxminmax)(audio_orig)
+        self.assertTrue((result.min() == -1. or result.max() == 1.) and
+                        result.min() >= -1. and result.max() <= 1.,
+                        "min: {}, max: {}".format(result.min(), result.max()))
+
+    def test_pad_trim(self):
+
+        audio_orig = self.sig.clone()
+        length_orig = audio_orig.size(0)
+        length_new = int(length_orig * 1.2)
+
+        result = transforms.PadTrim(max_len=length_new)(audio_orig)
+
+        self.assertTrue(result.size(0) == length_new,
+                        "old size: {}, new size: {}".format(audio_orig.size(0), result.size(0)))
+
+        audio_orig = self.sig.clone()
+        length_orig = audio_orig.size(0)
+        length_new = int(length_orig * 0.8)
+
+        result = transforms.PadTrim(max_len=length_new)(audio_orig)
+
+        self.assertTrue(result.size(0) == length_new,
+                        "old size: {}, new size: {}".format(audio_orig.size(0), result.size(0)))
+
+
+    def test_downmix_mono(self):
+        
+        audio_L = self.sig.clone()
+        audio_R = self.sig.clone()
+        R_idx = int(audio_R.size(0) * 0.1)
+        audio_R = torch.cat((audio_R[R_idx:], audio_R[:R_idx]))
+
+        audio_Stereo = torch.cat((audio_L, audio_R), dim=1)
+
+        self.assertTrue(audio_Stereo.size(1) == 2)
+
+        result = transforms.DownmixMono()(audio_Stereo)
+
+        self.assertTrue(result.size(1) == 1)
+
+    def test_compose(self):
+
+        audio_orig = self.sig.clone()
+        length_orig = audio_orig.size(0)
+        length_new = int(length_orig * 1.2)
+        maxminmax = np.abs([audio_orig.min(), audio_orig.max()]).max().astype(np.float)
+
+        tset = (transforms.Scale(factor=maxminmax),
+                transforms.PadTrim(max_len=length_new))
+        result = transforms.Compose(tset)(audio_orig)
+
+        self.assertTrue(np.abs([result.min(), result.max()]).max() == 1.)
+
+        self.assertTrue(result.size(0) == length_new)
+
+
+if __name__ == '__main__':
+    unittest.main()

torchaudio/__init__.py

@@ -7,6 +7,8 @@ from cffi import FFI
 ffi = FFI()
 from ._ext import th_sox
 
+from torchaudio import transforms
+from torchaudio import datasets
 
 def check_input(src):
     if not torch.is_tensor(src):
@@ -23,7 +25,7 @@ def load(filename, out=None):
     typename = type(out).__name__.replace('Tensor', '')
     func = getattr(th_sox, 'libthsox_{}_read_audio_file'.format(typename))
     sample_rate_p = ffi.new('int*')
-    func(str(filename).encode("ascii"), out, sample_rate_p)
+    func(str(filename).encode("utf-8"), out, sample_rate_p)
     sample_rate = sample_rate_p[0]
     return out, sample_rate
 
@@ -37,4 +39,4 @@ def save(filepath, src, sample_rate):
     typename = type(src).__name__.replace('Tensor', '')
     func = getattr(th_sox, 'libthsox_{}_write_audio_file'.format(typename))
 
-    func(bytes(filepath, "ascii"), src, extension[1:], sample_rate)
+    func(bytes(filepath, "utf-8"), src, bytes(extension[1:], "utf-8"), sample_rate)

torchaudio/datasets/__init__.py

+from .yesno import YESNO
+from .vctk import VCTK
+
+__all__ = ('YESNO', 'VCTK')

torchaudio/datasets/vctk.py

+from __future__ import print_function
+import torch.utils.data as data
+import os
+import os.path
+import shutil
+import errno
+import torch
+import torchaudio
+
+AUDIO_EXTENSIONS = [
+    '.wav', '.mp3', '.flac', '.sph', '.ogg', '.opus',
+    '.WAV', '.MP3', '.FLAC', '.SPH', '.OGG', '.OPUS',
+]
+
+def is_audio_file(filename):
+    return any(filename.endswith(extension) for extension in AUDIO_EXTENSIONS)
+
+def make_manifest(dir):
+    audios = []
+    dir = os.path.expanduser(dir)
+    for target in sorted(os.listdir(dir)):
+        d = os.path.join(dir, target)
+        if not os.path.isdir(d):
+            continue
+
+        for root, _, fnames in sorted(os.walk(d)):
+            for fname in fnames:
+                if is_audio_file(fname):
+                    path = os.path.join(root, fname)
+                    item = path
+                    audios.append(item)
+    return audios
+
+def read_audio(fp, downsample=True):
+    sig, sr = torchaudio.load(fp)
+    if downsample:
+        # 48khz -> 16 khz
+        if sig.size(0) % 3 == 0:
+            sig = sig.view(3, -1, sig.size(1)).mean(0)
+        else:
+            sig = sig[:-(sig.size(0) % 3)].view(3, -1, sig.size(1)).mean(0)
+    return sig, sr
+
+def load_txts(dir):
+    """Create a dictionary with all the text of the audio transcriptions."""
+    utterences = dict()
+    txts = []
+    dir = os.path.expanduser(dir)
+    for target in sorted(os.listdir(dir)):
+        d = os.path.join(dir, target)
+        if not os.path.isdir(d):
+            continue
+
+        for root, _, fnames in sorted(os.walk(d)):
+            for fname in fnames:
+                if fname.endswith(".txt"):
+                    with open(os.path.join(root, fname), "r") as f:
+                        fname_no_ext = os.path.basename(fname).rsplit(".", 1)[0]
+                        utterences[fname_no_ext] = f.readline()
+                        #utterences += f.readlines()
+    #utterences = dict([tuple(u.strip().split(" ", 1)) for u in utterences])
+    return utterences
+
+class VCTK(data.Dataset):
+    """`VCTK <http://homepages.inf.ed.ac.uk/jyamagis/page3/page58/page58.html>`_ Dataset.
+    `alternate url <http://datashare.is.ed.ac.uk/handle/10283/2651>`
+
+    Args:
+        root (string): Root directory of dataset where ``processed/training.pt``
+            and  ``processed/test.pt`` exist.
+        download (bool, optional): If true, downloads the dataset from the internet and
+            puts it in root directory. If dataset is already downloaded, it is not
+            downloaded again.
+        transform (callable, optional): A function/transform that  takes in an PIL image
+            and returns a transformed version. E.g, ``transforms.Scale``
+        target_transform (callable, optional): A function/transform that takes in the
+            target and transforms it.
+        dev_mode(bool, optional): if true, clean up is not performed on downloaded
+            files.  Useful to keep raw audio and transcriptions.
+    """
+    raw_folder = 'vctk/raw'
+    processed_folder = 'vctk/processed'
+    url = 'http://homepages.inf.ed.ac.uk/jyamagis/release/VCTK-Corpus.tar.gz'
+    dset_path = 'VCTK-Corpus'
+
+    def __init__(self, root, downsample=True, transform=None, target_transform=None, download=False, dev_mode=False):
+        self.root = os.path.expanduser(root)
+        self.downsample = downsample
+        self.transform = transform
+        self.target_transform = target_transform
+        self.dev_mode = dev_mode
+        self.data = []
+        self.labels = []
+        self.chunk_size = 1000
+        self.num_samples = 0
+        self.max_len = 0
+        self.cached_pt = 0
+
+        if download:
+            self.download()
+
+        if not self._check_exists():
+            raise RuntimeError('Dataset not found.' +
+                               ' You can use download=True to download it')
+        self._read_info()
+        self.data, self.labels = torch.load(os.path.join(self.root, self.processed_folder, "vctk_{:04d}.pt".format(self.cached_pt)))
+
+    def __getitem__(self, index):
+        """
+        Args:
+            index (int): Index
+
+        Returns:
+            tuple: (image, target) where target is index of the target class.
+        """
+        if self.cached_pt != index // self.chunk_size:
+            self.cached_pt = int(index // self.chunk_size)
+            self.data, self.labels = torch.load(os.path.join(self.root, self.processed_folder, "vctk_{:04d}.pt".format(self.cached_pt)))
+        index = index % self.chunk_size
+        audio, target = self.data[index], self.labels[index]
+
+        if self.transform is not None:
+            audio = self.transform(audio)
+
+        if self.target_transform is not None:
+            target = self.target_transform(target)
+
+        return audio, target
+
+    def __len__(self):
+        return self.num_samples
+
+    def _check_exists(self):
+        return os.path.exists(os.path.join(self.root, self.processed_folder, "vctk_info.txt"))
+
+    def _write_info(self, num_items):
+        info_path = os.path.join(self.root, self.processed_folder, "vctk_info.txt")
+        with open(info_path, "w") as f:
+            f.write("num_samples,{}\n".format(num_items))
+            f.write("max_len,{}\n".format(self.max_len))
+
+    def _read_info(self):
+        info_path = os.path.join(self.root, self.processed_folder, "vctk_info.txt")
+        with open(info_path, "r") as f:
+            self.num_samples = int(f.readline().split(",")[1])
+            self.max_len = int(f.readline().split(",")[1])
+
+    def download(self):
+        """Download the VCTK data if it doesn't exist in processed_folder already."""
+        from six.moves import urllib
+        import tarfile
+
+        if self._check_exists():
+            return
+
+        raw_abs_dir = os.path.join(self.root, self.raw_folder)
+        processed_abs_dir = os.path.join(self.root, self.processed_folder)
+        dset_abs_path = os.path.join(self.root, self.raw_folder, self.dset_path)
+
+        # download files
+        try:
+            os.makedirs(os.path.join(self.root, self.raw_folder))
+            os.makedirs(os.path.join(self.root, self.processed_folder))
+        except OSError as e:
+            if e.errno == errno.EEXIST:
+                pass
+            else:
+                raise
+
+        url = self.url
+        print('Downloading ' + url)
+        filename = url.rpartition('/')[2]
+        file_path = os.path.join(self.root, self.raw_folder, filename)
+        if not os.path.isfile(file_path):
+            data = urllib.request.urlopen(url)
+            with open(file_path, 'wb') as f:
+                f.write(data.read())
+        if not os.path.exists(dset_abs_path):
+            with tarfile.open(file_path) as zip_f:
+                zip_f.extractall(raw_abs_dir)
+        else:
+            print("Using existing raw folder")
+        if not self.dev_mode:
+            os.unlink(file_path)
+
+        # process and save as torch files
+        print('Processing...')
+        shutil.copyfile(
+            os.path.join(dset_abs_path, "COPYING"),
+            os.path.join(processed_abs_dir, "VCTK_COPYING")
+        )
+        audios = make_manifest(dset_abs_path)
+        utterences = load_txts(dset_abs_path)
+        self.max_len = 0
+        print("Found {} audio files and {} utterences".format(len(audios), len(utterences)))
+        for n in range(len(audios) // self.chunk_size + 1):
+            tensors = []
+            labels = []
+            lengths = []
+            st_idx = n * self.chunk_size
+            end_idx = st_idx + self.chunk_size
+            for i, f in enumerate(audios[st_idx:end_idx]):
+                txt_dir = os.path.dirname(f).replace("wav48", "txt")
+                if os.path.exists(txt_dir):
+                    f_rel_no_ext = os.path.basename(f).rsplit(".", 1)[0]
+                    sig = read_audio(f, downsample=self.downsample)[0]
+                    tensors.append(sig)
+                    lengths.append(sig.size(0))
+                    labels.append(utterences[f_rel_no_ext])
+                    self.max_len = sig.size(0) if sig.size(0) > self.max_len else self.max_len
+            # sort sigs/labels: longest -> shortest
+            tensors, labels = zip(*[(b, c) for (a,b,c) in sorted(zip(lengths, tensors, labels), key=lambda x: x[0], reverse=True)])
+            data = (tensors, labels)
+            torch.save(
+                data,
+                os.path.join(
+                    self.root,
+                    self.processed_folder,
+                    "vctk_{:04d}.pt".format(n)
+                )
+            )
+        self._write_info((n*self.chunk_size)+i+1)
+        if not self.dev_mode:
+            shutil.rmtree(raw_abs_dir, ignore_errors=True)
+
+        print('Done!')

torchaudio/datasets/yesno.py

+from __future__ import print_function
+import torch.utils.data as data
+import os
+import os.path
+import shutil
+import errno
+import torch
+import torchaudio
+
+class YESNO(data.Dataset):
+    """`YesNo Hebrew <http://www.openslr.org/1/>`_ Dataset.
+
+    Args:
+        root (string): Root directory of dataset where ``processed/training.pt``
+            and  ``processed/test.pt`` exist.
+        download (bool, optional): If true, downloads the dataset from the internet and
+            puts it in root directory. If dataset is already downloaded, it is not
+            downloaded again.
+        transform (callable, optional): A function/transform that  takes in an PIL image
+            and returns a transformed version. E.g, ``transforms.Scale``
+        target_transform (callable, optional): A function/transform that takes in the
+            target and transforms it.
+        dev_mode(bool, optional): if true, clean up is not performed on downloaded
+            files.  Useful to keep raw audio and transcriptions.
+    """
+    raw_folder = 'yesno/raw'
+    processed_folder = 'yesno/processed'
+    url = 'http://www.openslr.org/resources/1/waves_yesno.tar.gz'
+    dset_path = 'waves_yesno'
+    processed_file = 'yesno.pt'
+
+    def __init__(self, root, transform=None, target_transform=None, download=False, dev_mode=False):
+        self.root = os.path.expanduser(root)
+        self.transform = transform
+        self.target_transform = target_transform
+        self.dev_mode = dev_mode
+        self.data = []
+        self.labels = []
+        self.num_samples = 0
+        self.max_len = 0
+
+        if download:
+            self.download()
+
+        if not self._check_exists():
+            raise RuntimeError('Dataset not found.' +
+                               ' You can use download=True to download it')
+        self.data, self.labels = torch.load(os.path.join(self.root, self.processed_folder, self.processed_file))
+
+    def __getitem__(self, index):
+        """
+        Args:
+            index (int): Index
+
+        Returns:
+            tuple: (image, target) where target is index of the target class.
+        """
+        audio, target = self.data[index], self.labels[index]
+
+        if self.transform is not None:
+            audio = self.transform(audio)
+
+        if self.target_transform is not None:
+            target = self.target_transform(target)
+
+        return audio, target
+
+    def __len__(self):
+        return len(self.data)
+
+    def _check_exists(self):
+        return os.path.exists(os.path.join(self.root, self.processed_folder, self.processed_file))
+
+    def download(self):
+        """Download the yesno data if it doesn't exist in processed_folder already."""
+        from six.moves import urllib
+        import tarfile
+
+        if self._check_exists():
+            return
+
+        raw_abs_dir = os.path.join(self.root, self.raw_folder)
+        processed_abs_dir = os.path.join(self.root, self.processed_folder)
+        dset_abs_path = os.path.join(self.root, self.raw_folder, self.dset_path)
+
+        # download files
+        try:
+            os.makedirs(os.path.join(self.root, self.raw_folder))
+            os.makedirs(os.path.join(self.root, self.processed_folder))
+        except OSError as e:
+            if e.errno == errno.EEXIST:
+                pass
+            else:
+                raise
+
+        url = self.url
+        print('Downloading ' + url)
+        filename = url.rpartition('/')[2]
+        file_path = os.path.join(self.root, self.raw_folder, filename)
+        if not os.path.isfile(file_path):
+            data = urllib.request.urlopen(url)
+            with open(file_path, 'wb') as f:
+                f.write(data.read())
+        else:
+            print("Tar file already downloaded")
+        if not os.path.exists(dset_abs_path):
+            with tarfile.open(file_path) as zip_f:
+                zip_f.extractall(raw_abs_dir)
+        else:
+            print("Tar file already extracted")
+
+        if not self.dev_mode:
+            os.unlink(file_path)
+
+        # process and save as torch files
+        print('Processing...')
+        shutil.copyfile(
+            os.path.join(dset_abs_path, "README"),
+            os.path.join(processed_abs_dir, "YESNO_README")
+        )
+        audios = [x for x in os.listdir(dset_abs_path) if ".wav" in x]
+        print("Found {} audio files".format(len(audios)))
+        tensors = []
+        labels = []
+        lengths = []
+        for i, f in enumerate(audios):
+            full_path = os.path.join(dset_abs_path, f)
+            sig, sr = torchaudio.load(full_path)
+            tensors.append(sig)
+            lengths.append(sig.size(0))
+            labels.append(os.path.basename(f).split(".", 1)[0].split("_"))
+        # sort sigs/labels: longest -> shortest
+        tensors, labels = zip(*[(b, c) for (a,b,c) in sorted(zip(lengths, tensors, labels), key=lambda x: x[0], reverse=True)])
+        self.max_len = tensors[0].size(0)
+        torch.save(
+            (tensors, labels),
+            os.path.join(
+                self.root,
+                self.processed_folder,
+                self.processed_file
+            )
+        )
+        if not self.dev_mode:
+            shutil.rmtree(raw_abs_dir, ignore_errors=True)
+
+        print('Done!')

torchaudio/transforms.py

+from __future__ import division
+import torch
+import numpy as np
+
+class Compose(object):
+    """Composes several transforms together.
+
+    Args:
+        transforms (list of ``Transform`` objects): list of transforms to compose.
+
+    Example:
+        >>> transforms.Compose([
+        >>>     transforms.Scale(),
+        >>>     transforms.PadTrim(max_len=16000),
+        >>> ])
+    """
+
+    def __init__(self, transforms):
+        self.transforms = transforms
+
+    def __call__(self, audio):
+        for t in self.transforms:
+            audio = t(audio)
+        return audio
+
+class Scale(object):
+    """Scale audio tensor from a 16-bit integer (represented as a FloatTensor)
+    to a floating point number between -1.0 and 1.0.  Note the 16-bit number is
+    called the "bit depth" or "precision", not to be confused with "bit rate".
+
+    Args:
+        factor (float): maximum value of input tensor. default: 16-bit depth
+
+    """
+
+    def __init__(self, factor=2**31):
+        self.factor = factor
+
+    def __call__(self, tensor):
+        """
+
+        Args:
+            tensor (Tensor): Tensor of audio of size (Samples x Channels)
+
+        Returns:
+            Tensor: Scaled by the scale factor. (default between -1.0 and 1.0)
+
+        """
+        if isinstance(tensor, (torch.LongTensor, torch.IntTensor)):
+            tensor = tensor.float()
+
+        return tensor / self.factor
+
+class PadTrim(object):
+    """Pad/Trim a 1d-Tensor (Signal or Labels)
+
+    """
+
+    def __init__(self, max_len, fill_value=0):
+        self.max_len = max_len
+        self.fill_value = fill_value
+
+    def __call__(self, tensor):
+        """
+
+        Args:
+            tensor (Tensor): Tensor of audio of size (Samples x Channels)
+            max_len (int): Length to which the tensor will be padded
+
+        Returns:
+            Tensor: (max_len x Channels)
+
+        """
+        if self.max_len > tensor.size(0):
+            pad = torch.ones((self.max_len-tensor.size(0),
+                              tensor.size(1))) * self.fill_value
+            pad = pad.type_as(tensor)
+            tensor = torch.cat((tensor, pad), dim=0)
+        elif self.max_len < tensor.size(0):
+            tensor = tensor[:self.max_len, :]
+        return tensor
+
+
+class DownmixMono(object):
+    """Downmix any stereo signals to mono
+
+    """
+
+    def __init__(self):
+        pass
+
+    def __call__(self, tensor):
+        """
+
+        Args:
+            tensor (Tensor): Tensor of audio of size (Samples x Channels)
+
+        Returns:
+            Tensor: (Samples x 1)
+
+        """
+        if isinstance(tensor, (torch.LongTensor, torch.IntTensor)):
+            tensor = tensor.float()
+
+        if tensor.size(1) > 1:
+            tensor = torch.mean(tensor.float(), 1, True)
+        return tensor