add encoder

encoder/audio.py

@@ -0,0 +1,117 @@
+from scipy.ndimage.morphology import binary_dilation
+from encoder.params_data import *
+from pathlib import Path
+from typing import Optional, Union
+from warnings import warn
+import numpy as np
+import librosa
+import struct
+
+try:
+    import webrtcvad
+except:
+    warn("Unable to import 'webrtcvad'. This package enables noise removal and is recommended.")
+    webrtcvad=None
+
+int16_max = (2 ** 15) - 1
+
+
+def preprocess_wav(fpath_or_wav: Union[str, Path, np.ndarray],
+                   source_sr: Optional[int] = None,
+                   normalize: Optional[bool] = True,
+                   trim_silence: Optional[bool] = True):
+    """
+    Applies the preprocessing operations used in training the Speaker Encoder to a waveform 
+    either on disk or in memory. The waveform will be resampled to match the data hyperparameters.
+
+    :param fpath_or_wav: either a filepath to an audio file (many extensions are supported, not 
+    just .wav), either the waveform as a numpy array of floats.
+    :param source_sr: if passing an audio waveform, the sampling rate of the waveform before 
+    preprocessing. After preprocessing, the waveform's sampling rate will match the data 
+    hyperparameters. If passing a filepath, the sampling rate will be automatically detected and 
+    this argument will be ignored.
+    """
+    # Load the wav from disk if needed
+    if isinstance(fpath_or_wav, str) or isinstance(fpath_or_wav, Path):
+        wav, source_sr = librosa.load(str(fpath_or_wav), sr=None)
+    else:
+        wav = fpath_or_wav
+    
+    # Resample the wav if needed
+    if source_sr is not None and source_sr != sampling_rate:
+        wav = librosa.resample(wav, source_sr, sampling_rate)
+        
+    # Apply the preprocessing: normalize volume and shorten long silences 
+    if normalize:
+        wav = normalize_volume(wav, audio_norm_target_dBFS, increase_only=True)
+    if webrtcvad and trim_silence:
+        wav = trim_long_silences(wav)
+    
+    return wav
+
+
+def wav_to_mel_spectrogram(wav):
+    """
+    Derives a mel spectrogram ready to be used by the encoder from a preprocessed audio waveform.
+    Note: this not a log-mel spectrogram.
+    """
+    frames = librosa.feature.melspectrogram(
+        wav,
+        sampling_rate,
+        n_fft=int(sampling_rate * mel_window_length / 1000),
+        hop_length=int(sampling_rate * mel_window_step / 1000),
+        n_mels=mel_n_channels
+    )
+    return frames.astype(np.float32).T
+
+
+def trim_long_silences(wav):
+    """
+    Ensures that segments without voice in the waveform remain no longer than a 
+    threshold determined by the VAD parameters in params.py.
+
+    :param wav: the raw waveform as a numpy array of floats 
+    :return: the same waveform with silences trimmed away (length <= original wav length)
+    """
+    # Compute the voice detection window size
+    samples_per_window = (vad_window_length * sampling_rate) // 1000
+    
+    # Trim the end of the audio to have a multiple of the window size
+    wav = wav[:len(wav) - (len(wav) % samples_per_window)]
+    
+    # Convert the float waveform to 16-bit mono PCM
+    pcm_wave = struct.pack("%dh" % len(wav), *(np.round(wav * int16_max)).astype(np.int16))
+    
+    # Perform voice activation detection
+    voice_flags = []
+    vad = webrtcvad.Vad(mode=3)
+    for window_start in range(0, len(wav), samples_per_window):
+        window_end = window_start + samples_per_window
+        voice_flags.append(vad.is_speech(pcm_wave[window_start * 2:window_end * 2],
+                                         sample_rate=sampling_rate))
+    voice_flags = np.array(voice_flags)
+    
+    # Smooth the voice detection with a moving average
+    def moving_average(array, width):
+        array_padded = np.concatenate((np.zeros((width - 1) // 2), array, np.zeros(width // 2)))
+        ret = np.cumsum(array_padded, dtype=float)
+        ret[width:] = ret[width:] - ret[:-width]
+        return ret[width - 1:] / width
+    
+    audio_mask = moving_average(voice_flags, vad_moving_average_width)
+    audio_mask = np.round(audio_mask).astype(np.bool)
+    
+    # Dilate the voiced regions
+    audio_mask = binary_dilation(audio_mask, np.ones(vad_max_silence_length + 1))
+    audio_mask = np.repeat(audio_mask, samples_per_window)
+    
+    return wav[audio_mask == True]
+
+
+def normalize_volume(wav, target_dBFS, increase_only=False, decrease_only=False):
+    if increase_only and decrease_only:
+        raise ValueError("Both increase only and decrease only are set")
+    dBFS_change = target_dBFS - 10 * np.log10(np.mean(wav ** 2))
+    if (dBFS_change < 0 and increase_only) or (dBFS_change > 0 and decrease_only):
+        return wav
+    return wav * (10 ** (dBFS_change / 20))

encoder/config.py

@@ -0,0 +1,45 @@
+librispeech_datasets = {
+    "train": {
+        "clean": ["LibriSpeech/train-clean-100", "LibriSpeech/train-clean-360"],
+        "other": ["LibriSpeech/train-other-500"]
+    },
+    "test": {
+        "clean": ["LibriSpeech/test-clean"],
+        "other": ["LibriSpeech/test-other"]
+    },
+    "dev": {
+        "clean": ["LibriSpeech/dev-clean"],
+        "other": ["LibriSpeech/dev-other"]
+    },
+}
+libritts_datasets = {
+    "train": {
+        "clean": ["LibriTTS/train-clean-100", "LibriTTS/train-clean-360"],
+        "other": ["LibriTTS/train-other-500"]
+    },
+    "test": {
+        "clean": ["LibriTTS/test-clean"],
+        "other": ["LibriTTS/test-other"]
+    },
+    "dev": {
+        "clean": ["LibriTTS/dev-clean"],
+        "other": ["LibriTTS/dev-other"]
+    },
+}
+voxceleb_datasets = {
+    "voxceleb1" : {
+        "train": ["VoxCeleb1/wav"],
+        "test": ["VoxCeleb1/test_wav"]
+    },
+    "voxceleb2" : {
+        "train": ["VoxCeleb2/dev/aac"],
+        "test": ["VoxCeleb2/test_wav"]
+    }
+}
+
+other_datasets = [
+    "LJSpeech-1.1",
+    "VCTK-Corpus/wav48",
+]
+
+anglophone_nationalites = ["australia", "canada", "ireland", "uk", "usa"]

encoder/data_objects/__init__.py

@@ -0,0 +1,2 @@
+from encoder.data_objects.speaker_verification_dataset import SpeakerVerificationDataset
+from encoder.data_objects.speaker_verification_dataset import SpeakerVerificationDataLoader

encoder/data_objects/random_cycler.py

@@ -0,0 +1,37 @@
+import random
+
+class RandomCycler:
+    """
+    Creates an internal copy of a sequence and allows access to its items in a constrained random 
+    order. For a source sequence of n items and one or several consecutive queries of a total 
+    of m items, the following guarantees hold (one implies the other):
+        - Each item will be returned between m // n and ((m - 1) // n) + 1 times.
+        - Between two appearances of the same item, there may be at most 2 * (n - 1) other items.
+    """
+    
+    def __init__(self, source):
+        if len(source) == 0:
+            raise Exception("Can't create RandomCycler from an empty collection")
+        self.all_items = list(source)
+        self.next_items = []
+    
+    def sample(self, count: int):
+        shuffle = lambda l: random.sample(l, len(l))
+        
+        out = []
+        while count > 0:
+            if count >= len(self.all_items):
+                out.extend(shuffle(list(self.all_items)))
+                count -= len(self.all_items)
+                continue
+            n = min(count, len(self.next_items))
+            out.extend(self.next_items[:n])
+            count -= n
+            self.next_items = self.next_items[n:]
+            if len(self.next_items) == 0:
+                self.next_items = shuffle(list(self.all_items))
+        return out
+    
+    def __next__(self):
+        return self.sample(1)[0]
+

encoder/data_objects/speaker.py

@@ -0,0 +1,40 @@
+from encoder.data_objects.random_cycler import RandomCycler
+from encoder.data_objects.utterance import Utterance
+from pathlib import Path
+
+# Contains the set of utterances of a single speaker
+class Speaker:
+    def __init__(self, root: Path):
+        self.root = root
+        self.name = root.name
+        self.utterances = None
+        self.utterance_cycler = None
+        
+    def _load_utterances(self):
+        with self.root.joinpath("_sources.txt").open("r") as sources_file:
+            sources = [l.split(",") for l in sources_file]
+        sources = {frames_fname: wave_fpath for frames_fname, wave_fpath in sources}
+        self.utterances = [Utterance(self.root.joinpath(f), w) for f, w in sources.items()]
+        self.utterance_cycler = RandomCycler(self.utterances)
+               
+    def random_partial(self, count, n_frames):
+        """
+        Samples a batch of <count> unique partial utterances from the disk in a way that all 
+        utterances come up at least once every two cycles and in a random order every time.
+        
+        :param count: The number of partial utterances to sample from the set of utterances from 
+        that speaker. Utterances are guaranteed not to be repeated if <count> is not larger than 
+        the number of utterances available.
+        :param n_frames: The number of frames in the partial utterance.
+        :return: A list of tuples (utterance, frames, range) where utterance is an Utterance, 
+        frames are the frames of the partial utterances and range is the range of the partial 
+        utterance with regard to the complete utterance.
+        """
+        if self.utterances is None:
+            self._load_utterances()
+
+        utterances = self.utterance_cycler.sample(count)
+
+        a = [(u,) + u.random_partial(n_frames) for u in utterances]
+
+        return a

encoder/data_objects/speaker_batch.py

@@ -0,0 +1,13 @@
+import numpy as np
+from typing import List
+from encoder.data_objects.speaker import Speaker
+
+
+class SpeakerBatch:
+    def __init__(self, speakers: List[Speaker], utterances_per_speaker: int, n_frames: int):
+        self.speakers = speakers
+        self.partials = {s: s.random_partial(utterances_per_speaker, n_frames) for s in speakers}
+
+        # Array of shape (n_speakers * n_utterances, n_frames, mel_n), e.g. for 3 speakers with
+        # 4 utterances each of 160 frames of 40 mel coefficients: (12, 160, 40)
+        self.data = np.array([frames for s in speakers for _, frames, _ in self.partials[s]])

encoder/data_objects/speaker_verification_dataset.py

@@ -0,0 +1,56 @@
+from encoder.data_objects.random_cycler import RandomCycler
+from encoder.data_objects.speaker_batch import SpeakerBatch
+from encoder.data_objects.speaker import Speaker
+from encoder.params_data import partials_n_frames
+from torch.utils.data import Dataset, DataLoader
+from pathlib import Path
+
+# TODO: improve with a pool of speakers for data efficiency
+
+class SpeakerVerificationDataset(Dataset):
+    def __init__(self, datasets_root: Path):
+        self.root = datasets_root
+        speaker_dirs = [f for f in self.root.glob("*") if f.is_dir()]
+        if len(speaker_dirs) == 0:
+            raise Exception("No speakers found. Make sure you are pointing to the directory "
+                            "containing all preprocessed speaker directories.")
+        self.speakers = [Speaker(speaker_dir) for speaker_dir in speaker_dirs]
+        self.speaker_cycler = RandomCycler(self.speakers)
+
+    def __len__(self):
+        return int(1e10)
+        
+    def __getitem__(self, index):
+        return next(self.speaker_cycler)
+    
+    def get_logs(self):
+        log_string = ""
+        for log_fpath in self.root.glob("*.txt"):
+            with log_fpath.open("r") as log_file:
+                log_string += "".join(log_file.readlines())
+        return log_string
+    
+    
+class SpeakerVerificationDataLoader(DataLoader):
+    def __init__(self, dataset, speakers_per_batch, utterances_per_speaker, sampler=None, 
+                 batch_sampler=None, num_workers=0, pin_memory=False, timeout=0, 
+                 worker_init_fn=None):
+        self.utterances_per_speaker = utterances_per_speaker
+
+        super().__init__(
+            dataset=dataset, 
+            batch_size=speakers_per_batch, 
+            shuffle=False, 
+            sampler=sampler, 
+            batch_sampler=batch_sampler, 
+            num_workers=num_workers,
+            collate_fn=self.collate, 
+            pin_memory=pin_memory, 
+            drop_last=False, 
+            timeout=timeout, 
+            worker_init_fn=worker_init_fn
+        )
+
+    def collate(self, speakers):
+        return SpeakerBatch(speakers, self.utterances_per_speaker, partials_n_frames) 
+    

encoder/data_objects/utterance.py

@@ -0,0 +1,26 @@
+import numpy as np
+
+
+class Utterance:
+    def __init__(self, frames_fpath, wave_fpath):
+        self.frames_fpath = frames_fpath
+        self.wave_fpath = wave_fpath
+        
+    def get_frames(self):
+        return np.load(self.frames_fpath)
+
+    def random_partial(self, n_frames):
+        """
+        Crops the frames into a partial utterance of n_frames
+        
+        :param n_frames: The number of frames of the partial utterance
+        :return: the partial utterance frames and a tuple indicating the start and end of the 
+        partial utterance in the complete utterance.
+        """
+        frames = self.get_frames()
+        if frames.shape[0] == n_frames:
+            start = 0
+        else:
+            start = np.random.randint(0, frames.shape[0] - n_frames)
+        end = start + n_frames
+        return frames[start:end], (start, end)

encoder/inference.py

@@ -0,0 +1,178 @@
+from encoder.params_data import *
+from encoder.model import SpeakerEncoder
+from encoder.audio import preprocess_wav   # We want to expose this function from here
+from matplotlib import cm
+from encoder import audio
+from pathlib import Path
+import numpy as np
+import torch
+
+_model = None # type: SpeakerEncoder
+_device = None # type: torch.device
+
+
+def load_model(weights_fpath: Path, device=None):
+    """
+    Loads the model in memory. If this function is not explicitely called, it will be run on the
+    first call to embed_frames() with the default weights file.
+
+    :param weights_fpath: the path to saved model weights.
+    :param device: either a torch device or the name of a torch device (e.g. "cpu", "cuda"). The
+    model will be loaded and will run on this device. Outputs will however always be on the cpu.
+    If None, will default to your GPU if it"s available, otherwise your CPU.
+    """
+    # TODO: I think the slow loading of the encoder might have something to do with the device it
+    #   was saved on. Worth investigating.
+    global _model, _device
+    if device is None:
+        _device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    elif isinstance(device, str):
+        _device = torch.device(device)
+    _model = SpeakerEncoder(_device, torch.device("cpu"))
+    checkpoint = torch.load(weights_fpath, _device)
+    _model.load_state_dict(checkpoint["model_state"])
+    _model.eval()
+    print("Loaded encoder \"%s\" trained to step %d" % (weights_fpath.name, checkpoint["step"]))
+
+
+def is_loaded():
+    return _model is not None
+
+
+def embed_frames_batch(frames_batch):
+    """
+    Computes embeddings for a batch of mel spectrogram.
+
+    :param frames_batch: a batch mel of spectrogram as a numpy array of float32 of shape
+    (batch_size, n_frames, n_channels)
+    :return: the embeddings as a numpy array of float32 of shape (batch_size, model_embedding_size)
+    """
+    if _model is None:
+        raise Exception("Model was not loaded. Call load_model() before inference.")
+
+    frames = torch.from_numpy(frames_batch).to(_device)
+    embed = _model.forward(frames).detach().cpu().numpy()
+    return embed
+
+
+def compute_partial_slices(n_samples, partial_utterance_n_frames=partials_n_frames,
+                           min_pad_coverage=0.75, overlap=0.5):
+    """
+    Computes where to split an utterance waveform and its corresponding mel spectrogram to obtain
+    partial utterances of <partial_utterance_n_frames> each. Both the waveform and the mel
+    spectrogram slices are returned, so as to make each partial utterance waveform correspond to
+    its spectrogram. This function assumes that the mel spectrogram parameters used are those
+    defined in params_data.py.
+
+    The returned ranges may be indexing further than the length of the waveform. It is
+    recommended that you pad the waveform with zeros up to wave_slices[-1].stop.
+
+    :param n_samples: the number of samples in the waveform
+    :param partial_utterance_n_frames: the number of mel spectrogram frames in each partial
+    utterance
+    :param min_pad_coverage: when reaching the last partial utterance, it may or may not have
+    enough frames. If at least <min_pad_coverage> of <partial_utterance_n_frames> are present,
+    then the last partial utterance will be considered, as if we padded the audio. Otherwise,
+    it will be discarded, as if we trimmed the audio. If there aren't enough frames for 1 partial
+    utterance, this parameter is ignored so that the function always returns at least 1 slice.
+    :param overlap: by how much the partial utterance should overlap. If set to 0, the partial
+    utterances are entirely disjoint.
+    :return: the waveform slices and mel spectrogram slices as lists of array slices. Index
+    respectively the waveform and the mel spectrogram with these slices to obtain the partial
+    utterances.
+    """
+    assert 0 <= overlap < 1
+    assert 0 < min_pad_coverage <= 1
+
+    samples_per_frame = int((sampling_rate * mel_window_step / 1000))
+    n_frames = int(np.ceil((n_samples + 1) / samples_per_frame))
+    frame_step = max(int(np.round(partial_utterance_n_frames * (1 - overlap))), 1)
+
+    # Compute the slices
+    wav_slices, mel_slices = [], []
+    steps = max(1, n_frames - partial_utterance_n_frames + frame_step + 1)
+    for i in range(0, steps, frame_step):
+        mel_range = np.array([i, i + partial_utterance_n_frames])
+        wav_range = mel_range * samples_per_frame
+        mel_slices.append(slice(*mel_range))
+        wav_slices.append(slice(*wav_range))
+
+    # Evaluate whether extra padding is warranted or not
+    last_wav_range = wav_slices[-1]
+    coverage = (n_samples - last_wav_range.start) / (last_wav_range.stop - last_wav_range.start)
+    if coverage < min_pad_coverage and len(mel_slices) > 1:
+        mel_slices = mel_slices[:-1]
+        wav_slices = wav_slices[:-1]
+
+    return wav_slices, mel_slices
+
+
+def embed_utterance(wav, using_partials=True, return_partials=False, **kwargs):
+    """
+    Computes an embedding for a single utterance.
+
+    # TODO: handle multiple wavs to benefit from batching on GPU
+    :param wav: a preprocessed (see audio.py) utterance waveform as a numpy array of float32
+    :param using_partials: if True, then the utterance is split in partial utterances of
+    <partial_utterance_n_frames> frames and the utterance embedding is computed from their
+    normalized average. If False, the utterance is instead computed from feeding the entire
+    spectogram to the network.
+    :param return_partials: if True, the partial embeddings will also be returned along with the
+    wav slices that correspond to the partial embeddings.
+    :param kwargs: additional arguments to compute_partial_splits()
+    :return: the embedding as a numpy array of float32 of shape (model_embedding_size,). If
+    <return_partials> is True, the partial utterances as a numpy array of float32 of shape
+    (n_partials, model_embedding_size) and the wav partials as a list of slices will also be
+    returned. If <using_partials> is simultaneously set to False, both these values will be None
+    instead.
+    """
+    # Process the entire utterance if not using partials
+    if not using_partials:
+        frames = audio.wav_to_mel_spectrogram(wav)
+        embed = embed_frames_batch(frames[None, ...])[0]
+        if return_partials:
+            return embed, None, None
+        return embed
+
+    # Compute where to split the utterance into partials and pad if necessary
+    wave_slices, mel_slices = compute_partial_slices(len(wav), **kwargs)
+    max_wave_length = wave_slices[-1].stop
+    if max_wave_length >= len(wav):
+        wav = np.pad(wav, (0, max_wave_length - len(wav)), "constant")
+
+    # Split the utterance into partials
+    frames = audio.wav_to_mel_spectrogram(wav)
+    frames_batch = np.array([frames[s] for s in mel_slices])
+    partial_embeds = embed_frames_batch(frames_batch)
+
+    # Compute the utterance embedding from the partial embeddings
+    raw_embed = np.mean(partial_embeds, axis=0)
+    embed = raw_embed / np.linalg.norm(raw_embed, 2)
+
+    if return_partials:
+        return embed, partial_embeds, wave_slices
+    return embed
+
+
+def embed_speaker(wavs, **kwargs):
+    raise NotImplemented()
+
+
+def plot_embedding_as_heatmap(embed, ax=None, title="", shape=None, color_range=(0, 0.30)):
+    import matplotlib.pyplot as plt
+    if ax is None:
+        ax = plt.gca()
+
+    if shape is None:
+        height = int(np.sqrt(len(embed)))
+        shape = (height, -1)
+    embed = embed.reshape(shape)
+
+    cmap = cm.get_cmap()
+    mappable = ax.imshow(embed, cmap=cmap)
+    cbar = plt.colorbar(mappable, ax=ax, fraction=0.046, pad=0.04)
+    sm = cm.ScalarMappable(cmap=cmap)
+    sm.set_clim(*color_range)
+
+    ax.set_xticks([]), ax.set_yticks([])
+    ax.set_title(title)

encoder/model.py

@@ -0,0 +1,135 @@
+from encoder.params_model import *
+from encoder.params_data import *
+from scipy.interpolate import interp1d
+from sklearn.metrics import roc_curve
+from torch.nn.utils import clip_grad_norm_
+from scipy.optimize import brentq
+from torch import nn
+import numpy as np
+import torch
+
+
+class SpeakerEncoder(nn.Module):
+    def __init__(self, device, loss_device):
+        super().__init__()
+        self.loss_device = loss_device
+        
+        # Network defition
+        self.lstm = nn.LSTM(input_size=mel_n_channels,
+                            hidden_size=model_hidden_size, 
+                            num_layers=model_num_layers, 
+                            batch_first=True).to(device)
+        self.linear = nn.Linear(in_features=model_hidden_size, 
+                                out_features=model_embedding_size).to(device)
+        self.relu = torch.nn.ReLU().to(device)
+        
+        # Cosine similarity scaling (with fixed initial parameter values)
+        self.similarity_weight = nn.Parameter(torch.tensor([10.])).to(loss_device)
+        self.similarity_bias = nn.Parameter(torch.tensor([-5.])).to(loss_device)
+
+        # Loss
+        self.loss_fn = nn.CrossEntropyLoss().to(loss_device)
+        
+    def do_gradient_ops(self):
+        # Gradient scale
+        self.similarity_weight.grad *= 0.01
+        self.similarity_bias.grad *= 0.01
+            
+        # Gradient clipping
+        clip_grad_norm_(self.parameters(), 3, norm_type=2)
+    
+    def forward(self, utterances, hidden_init=None):
+        """
+        Computes the embeddings of a batch of utterance spectrograms.
+        
+        :param utterances: batch of mel-scale filterbanks of same duration as a tensor of shape 
+        (batch_size, n_frames, n_channels) 
+        :param hidden_init: initial hidden state of the LSTM as a tensor of shape (num_layers, 
+        batch_size, hidden_size). Will default to a tensor of zeros if None.
+        :return: the embeddings as a tensor of shape (batch_size, embedding_size)
+        """
+        # Pass the input through the LSTM layers and retrieve all outputs, the final hidden state
+        # and the final cell state.
+        out, (hidden, cell) = self.lstm(utterances, hidden_init)
+        
+        # We take only the hidden state of the last layer
+        embeds_raw = self.relu(self.linear(hidden[-1]))
+        
+        # L2-normalize it
+        embeds = embeds_raw / (torch.norm(embeds_raw, dim=1, keepdim=True) + 1e-5)        
+
+        return embeds
+    
+    def similarity_matrix(self, embeds):
+        """
+        Computes the similarity matrix according the section 2.1 of GE2E.
+
+        :param embeds: the embeddings as a tensor of shape (speakers_per_batch, 
+        utterances_per_speaker, embedding_size)
+        :return: the similarity matrix as a tensor of shape (speakers_per_batch,
+        utterances_per_speaker, speakers_per_batch)
+        """
+        speakers_per_batch, utterances_per_speaker = embeds.shape[:2]
+        
+        # Inclusive centroids (1 per speaker). Cloning is needed for reverse differentiation
+        centroids_incl = torch.mean(embeds, dim=1, keepdim=True)
+        centroids_incl = centroids_incl.clone() / (torch.norm(centroids_incl, dim=2, keepdim=True) + 1e-5)
+
+        # Exclusive centroids (1 per utterance)
+        centroids_excl = (torch.sum(embeds, dim=1, keepdim=True) - embeds)
+        centroids_excl /= (utterances_per_speaker - 1)
+        centroids_excl = centroids_excl.clone() / (torch.norm(centroids_excl, dim=2, keepdim=True) + 1e-5)
+
+        # Similarity matrix. The cosine similarity of already 2-normed vectors is simply the dot
+        # product of these vectors (which is just an element-wise multiplication reduced by a sum).
+        # We vectorize the computation for efficiency.
+        sim_matrix = torch.zeros(speakers_per_batch, utterances_per_speaker,
+                                 speakers_per_batch).to(self.loss_device)
+        mask_matrix = 1 - np.eye(speakers_per_batch, dtype=np.int)
+        for j in range(speakers_per_batch):
+            mask = np.where(mask_matrix[j])[0]
+            sim_matrix[mask, :, j] = (embeds[mask] * centroids_incl[j]).sum(dim=2)
+            sim_matrix[j, :, j] = (embeds[j] * centroids_excl[j]).sum(dim=1)
+        
+        ## Even more vectorized version (slower maybe because of transpose)
+        # sim_matrix2 = torch.zeros(speakers_per_batch, speakers_per_batch, utterances_per_speaker
+        #                           ).to(self.loss_device)
+        # eye = np.eye(speakers_per_batch, dtype=np.int)
+        # mask = np.where(1 - eye)
+        # sim_matrix2[mask] = (embeds[mask[0]] * centroids_incl[mask[1]]).sum(dim=2)
+        # mask = np.where(eye)
+        # sim_matrix2[mask] = (embeds * centroids_excl).sum(dim=2)
+        # sim_matrix2 = sim_matrix2.transpose(1, 2)
+        
+        sim_matrix = sim_matrix * self.similarity_weight + self.similarity_bias
+        return sim_matrix
+    
+    def loss(self, embeds):
+        """
+        Computes the softmax loss according the section 2.1 of GE2E.
+        
+        :param embeds: the embeddings as a tensor of shape (speakers_per_batch, 
+        utterances_per_speaker, embedding_size)
+        :return: the loss and the EER for this batch of embeddings.
+        """
+        speakers_per_batch, utterances_per_speaker = embeds.shape[:2]
+        
+        # Loss
+        sim_matrix = self.similarity_matrix(embeds)
+        sim_matrix = sim_matrix.reshape((speakers_per_batch * utterances_per_speaker, 
+                                         speakers_per_batch))
+        ground_truth = np.repeat(np.arange(speakers_per_batch), utterances_per_speaker)
+        target = torch.from_numpy(ground_truth).long().to(self.loss_device)
+        loss = self.loss_fn(sim_matrix, target)
+        
+        # EER (not backpropagated)
+        with torch.no_grad():
+            inv_argmax = lambda i: np.eye(1, speakers_per_batch, i, dtype=np.int)[0]
+            labels = np.array([inv_argmax(i) for i in ground_truth])
+            preds = sim_matrix.detach().cpu().numpy()
+
+            # Snippet from https://yangcha.github.io/EER-ROC/
+            fpr, tpr, thresholds = roc_curve(labels.flatten(), preds.flatten())           
+            eer = brentq(lambda x: 1. - x - interp1d(fpr, tpr)(x), 0., 1.)
+            
+        return loss, eer

encoder/params_data.py

@@ -0,0 +1,29 @@
+
+## Mel-filterbank
+mel_window_length = 25  # In milliseconds
+mel_window_step = 10    # In milliseconds
+mel_n_channels = 40
+
+
+## Audio
+sampling_rate = 16000
+# Number of spectrogram frames in a partial utterance
+partials_n_frames = 160     # 1600 ms
+# Number of spectrogram frames at inference
+inference_n_frames = 80     #  800 ms
+
+
+## Voice Activation Detection
+# Window size of the VAD. Must be either 10, 20 or 30 milliseconds.
+# This sets the granularity of the VAD. Should not need to be changed.
+vad_window_length = 30  # In milliseconds
+# Number of frames to average together when performing the moving average smoothing.
+# The larger this value, the larger the VAD variations must be to not get smoothed out. 
+vad_moving_average_width = 8
+# Maximum number of consecutive silent frames a segment can have.
+vad_max_silence_length = 6
+
+
+## Audio volume normalization
+audio_norm_target_dBFS = -30
+

encoder/params_model.py

@@ -0,0 +1,11 @@
+
+## Model parameters
+model_hidden_size = 256
+model_embedding_size = 256
+model_num_layers = 3
+
+
+## Training parameters
+learning_rate_init = 1e-4
+speakers_per_batch = 64
+utterances_per_speaker = 10

encoder/preprocess.py

@@ -0,0 +1,184 @@
+from datetime import datetime
+from functools import partial
+from multiprocessing import Pool
+from pathlib import Path
+
+import numpy as np
+from tqdm import tqdm
+
+from encoder import audio
+from encoder.config import librispeech_datasets, anglophone_nationalites
+from encoder.params_data import *
+
+
+_AUDIO_EXTENSIONS = ("wav", "flac", "m4a", "mp3")
+
+class DatasetLog:
+    """
+    Registers metadata about the dataset in a text file.
+    """
+    def __init__(self, root, name):
+        self.text_file = open(Path(root, "Log_%s.txt" % name.replace("/", "_")), "w")
+        self.sample_data = dict()
+
+        start_time = str(datetime.now().strftime("%A %d %B %Y at %H:%M"))
+        self.write_line("Creating dataset %s on %s" % (name, start_time))
+        self.write_line("-----")
+        self._log_params()
+
+    def _log_params(self):
+        from encoder import params_data
+        self.write_line("Parameter values:")
+        for param_name in (p for p in dir(params_data) if not p.startswith("__")):
+            value = getattr(params_data, param_name)
+            self.write_line("\t%s: %s" % (param_name, value))
+        self.write_line("-----")
+
+    def write_line(self, line):
+        self.text_file.write("%s\n" % line)
+
+    def add_sample(self, **kwargs):
+        for param_name, value in kwargs.items():
+            if not param_name in self.sample_data:
+                self.sample_data[param_name] = []
+            self.sample_data[param_name].append(value)
+
+    def finalize(self):
+        self.write_line("Statistics:")
+        for param_name, values in self.sample_data.items():
+            self.write_line("\t%s:" % param_name)
+            self.write_line("\t\tmin %.3f, max %.3f" % (np.min(values), np.max(values)))
+            self.write_line("\t\tmean %.3f, median %.3f" % (np.mean(values), np.median(values)))
+        self.write_line("-----")
+        end_time = str(datetime.now().strftime("%A %d %B %Y at %H:%M"))
+        self.write_line("Finished on %s" % end_time)
+        self.text_file.close()
+
+
+def _init_preprocess_dataset(dataset_name, datasets_root, out_dir) -> (Path, DatasetLog):
+    dataset_root = datasets_root.joinpath(dataset_name)
+    if not dataset_root.exists():
+        print("Couldn\'t find %s, skipping this dataset." % dataset_root)
+        return None, None
+    return dataset_root, DatasetLog(out_dir, dataset_name)
+
+
+def _preprocess_speaker(speaker_dir: Path, datasets_root: Path, out_dir: Path, skip_existing: bool):
+    # Give a name to the speaker that includes its dataset
+    speaker_name = "_".join(speaker_dir.relative_to(datasets_root).parts)
+
+    # Create an output directory with that name, as well as a txt file containing a
+    # reference to each source file.
+    speaker_out_dir = out_dir.joinpath(speaker_name)
+    speaker_out_dir.mkdir(exist_ok=True)
+    sources_fpath = speaker_out_dir.joinpath("_sources.txt")
+
+    # There's a possibility that the preprocessing was interrupted earlier, check if
+    # there already is a sources file.
+    if sources_fpath.exists():
+        try:
+            with sources_fpath.open("r") as sources_file:
+                existing_fnames = {line.split(",")[0] for line in sources_file}
+        except:
+            existing_fnames = {}
+    else:
+        existing_fnames = {}
+
+    # Gather all audio files for that speaker recursively
+    sources_file = sources_fpath.open("a" if skip_existing else "w")
+    audio_durs = []
+    for extension in _AUDIO_EXTENSIONS:
+        for in_fpath in speaker_dir.glob("**/*.%s" % extension):
+            # Check if the target output file already exists
+            out_fname = "_".join(in_fpath.relative_to(speaker_dir).parts)
+            out_fname = out_fname.replace(".%s" % extension, ".npy")
+            if skip_existing and out_fname in existing_fnames:
+                continue
+
+            # Load and preprocess the waveform
+            wav = audio.preprocess_wav(in_fpath)
+            if len(wav) == 0:
+                continue
+
+            # Create the mel spectrogram, discard those that are too short
+            frames = audio.wav_to_mel_spectrogram(wav)
+            if len(frames) < partials_n_frames:
+                continue
+
+            out_fpath = speaker_out_dir.joinpath(out_fname)
+            np.save(out_fpath, frames)
+            sources_file.write("%s,%s\n" % (out_fname, in_fpath))
+            audio_durs.append(len(wav) / sampling_rate)
+
+    sources_file.close()
+
+    return audio_durs
+
+
+def _preprocess_speaker_dirs(speaker_dirs, dataset_name, datasets_root, out_dir, skip_existing, logger):
+    print("%s: Preprocessing data for %d speakers." % (dataset_name, len(speaker_dirs)))
+
+    # Process the utterances for each speaker
+    work_fn = partial(_preprocess_speaker, datasets_root=datasets_root, out_dir=out_dir, skip_existing=skip_existing)
+    with Pool(4) as pool:
+        tasks = pool.imap(work_fn, speaker_dirs)
+        for sample_durs in tqdm(tasks, dataset_name, len(speaker_dirs), unit="speakers"):
+            for sample_dur in sample_durs:
+                logger.add_sample(duration=sample_dur)
+
+    logger.finalize()
+    print("Done preprocessing %s.\n" % dataset_name)
+
+
+def preprocess_librispeech(datasets_root: Path, out_dir: Path, skip_existing=False):
+    for dataset_name in librispeech_datasets["train"]["other"]:
+        # Initialize the preprocessing
+        dataset_root, logger = _init_preprocess_dataset(dataset_name, datasets_root, out_dir)
+        if not dataset_root:
+            return
+
+        # Preprocess all speakers
+        speaker_dirs = list(dataset_root.glob("*"))
+        _preprocess_speaker_dirs(speaker_dirs, dataset_name, datasets_root, out_dir, skip_existing, logger)
+
+
+def preprocess_voxceleb1(datasets_root: Path, out_dir: Path, skip_existing=False):
+    # Initialize the preprocessing
+    dataset_name = "VoxCeleb1"
+    dataset_root, logger = _init_preprocess_dataset(dataset_name, datasets_root, out_dir)
+    if not dataset_root:
+        return
+
+    # Get the contents of the meta file
+    with dataset_root.joinpath("vox1_meta.csv").open("r") as metafile:
+        metadata = [line.split("\t") for line in metafile][1:]
+
+    # Select the ID and the nationality, filter out non-anglophone speakers
+    nationalities = {line[0]: line[3] for line in metadata}
+    keep_speaker_ids = [speaker_id for speaker_id, nationality in nationalities.items() if
+                        nationality.lower() in anglophone_nationalites]
+    print("VoxCeleb1: using samples from %d (presumed anglophone) speakers out of %d." %
+          (len(keep_speaker_ids), len(nationalities)))
+
+    # Get the speaker directories for anglophone speakers only
+    speaker_dirs = dataset_root.joinpath("wav").glob("*")
+    speaker_dirs = [speaker_dir for speaker_dir in speaker_dirs if
+                    speaker_dir.name in keep_speaker_ids]
+    print("VoxCeleb1: found %d anglophone speakers on the disk, %d missing (this is normal)." %
+          (len(speaker_dirs), len(keep_speaker_ids) - len(speaker_dirs)))
+
+    # Preprocess all speakers
+    _preprocess_speaker_dirs(speaker_dirs, dataset_name, datasets_root, out_dir, skip_existing, logger)
+
+
+def preprocess_voxceleb2(datasets_root: Path, out_dir: Path, skip_existing=False):
+    # Initialize the preprocessing
+    dataset_name = "VoxCeleb2"
+    dataset_root, logger = _init_preprocess_dataset(dataset_name, datasets_root, out_dir)
+    if not dataset_root:
+        return
+
+    # Get the speaker directories
+    # Preprocess all speakers
+    speaker_dirs = list(dataset_root.joinpath("dev", "aac").glob("*"))
+    _preprocess_speaker_dirs(speaker_dirs, dataset_name, datasets_root, out_dir, skip_existing, logger)

encoder/train.py

@@ -0,0 +1,125 @@
+from pathlib import Path
+
+import torch
+
+from encoder.data_objects import SpeakerVerificationDataLoader, SpeakerVerificationDataset
+from encoder.model import SpeakerEncoder
+from encoder.params_model import *
+from encoder.visualizations import Visualizations
+from utils.profiler import Profiler
+
+
+def sync(device: torch.device):
+    # For correct profiling (cuda operations are async)
+    if device.type == "cuda":
+        torch.cuda.synchronize(device)
+
+
+def train(run_id: str, clean_data_root: Path, models_dir: Path, umap_every: int, save_every: int,
+          backup_every: int, vis_every: int, force_restart: bool, visdom_server: str,
+          no_visdom: bool):
+    # Create a dataset and a dataloader
+    dataset = SpeakerVerificationDataset(clean_data_root)
+    loader = SpeakerVerificationDataLoader(
+        dataset,
+        speakers_per_batch,
+        utterances_per_speaker,
+        num_workers=4,
+    )
+
+    # Setup the device on which to run the forward pass and the loss. These can be different,
+    # because the forward pass is faster on the GPU whereas the loss is often (depending on your
+    # hyperparameters) faster on the CPU.
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    # FIXME: currently, the gradient is None if loss_device is cuda
+    loss_device = torch.device("cpu")
+
+    # Create the model and the optimizer
+    model = SpeakerEncoder(device, loss_device)
+    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate_init)
+    init_step = 1
+
+    # Configure file path for the model
+    model_dir = models_dir / run_id
+    model_dir.mkdir(exist_ok=True, parents=True)
+    state_fpath = model_dir / "encoder.pt"
+
+    # Load any existing model
+    if not force_restart:
+        if state_fpath.exists():
+            print("Found existing model \"%s\", loading it and resuming training." % run_id)
+            checkpoint = torch.load(state_fpath)
+            init_step = checkpoint["step"]
+            model.load_state_dict(checkpoint["model_state"])
+            optimizer.load_state_dict(checkpoint["optimizer_state"])
+            optimizer.param_groups[0]["lr"] = learning_rate_init
+        else:
+            print("No model \"%s\" found, starting training from scratch." % run_id)
+    else:
+        print("Starting the training from scratch.")
+    model.train()
+
+    # Initialize the visualization environment
+    vis = Visualizations(run_id, vis_every, server=visdom_server, disabled=no_visdom)
+    vis.log_dataset(dataset)
+    vis.log_params()
+    device_name = str(torch.cuda.get_device_name(0) if torch.cuda.is_available() else "CPU")
+    vis.log_implementation({"Device": device_name})
+
+    # Training loop
+    profiler = Profiler(summarize_every=10, disabled=False)
+    for step, speaker_batch in enumerate(loader, init_step):
+        profiler.tick("Blocking, waiting for batch (threaded)")
+
+        # Forward pass
+        inputs = torch.from_numpy(speaker_batch.data).to(device)
+        sync(device)
+        profiler.tick("Data to %s" % device)
+        embeds = model(inputs)
+        sync(device)
+        profiler.tick("Forward pass")
+        embeds_loss = embeds.view((speakers_per_batch, utterances_per_speaker, -1)).to(loss_device)
+        loss, eer = model.loss(embeds_loss)
+        sync(loss_device)
+        profiler.tick("Loss")
+
+        # Backward pass
+        model.zero_grad()
+        loss.backward()
+        profiler.tick("Backward pass")
+        model.do_gradient_ops()
+        optimizer.step()
+        profiler.tick("Parameter update")
+
+        # Update visualizations
+        # learning_rate = optimizer.param_groups[0]["lr"]
+        vis.update(loss.item(), eer, step)
+
+        # Draw projections and save them to the backup folder
+        if umap_every != 0 and step % umap_every == 0:
+            print("Drawing and saving projections (step %d)" % step)
+            projection_fpath = model_dir / f"umap_{step:06d}.png"
+            embeds = embeds.detach().cpu().numpy()
+            vis.draw_projections(embeds, utterances_per_speaker, step, projection_fpath)
+            vis.save()
+
+        # Overwrite the latest version of the model
+        if save_every != 0 and step % save_every == 0:
+            print("Saving the model (step %d)" % step)
+            torch.save({
+                "step": step + 1,
+                "model_state": model.state_dict(),
+                "optimizer_state": optimizer.state_dict(),
+            }, state_fpath)
+
+        # Make a backup
+        if backup_every != 0 and step % backup_every == 0:
+            print("Making a backup (step %d)" % step)
+            backup_fpath = model_dir / f"encoder_{step:06d}.bak"
+            torch.save({
+                "step": step + 1,
+                "model_state": model.state_dict(),
+                "optimizer_state": optimizer.state_dict(),
+            }, backup_fpath)
+
+        profiler.tick("Extras (visualizations, saving)")

encoder/visualizations.py

@@ -0,0 +1,179 @@
+from datetime import datetime
+from time import perf_counter as timer
+
+import numpy as np
+import umap
+import visdom
+
+from encoder.data_objects.speaker_verification_dataset import SpeakerVerificationDataset
+
+
+colormap = np.array([
+    [76, 255, 0],
+    [0, 127, 70],
+    [255, 0, 0],
+    [255, 217, 38],
+    [0, 135, 255],
+    [165, 0, 165],
+    [255, 167, 255],
+    [0, 255, 255],
+    [255, 96, 38],
+    [142, 76, 0],
+    [33, 0, 127],
+    [0, 0, 0],
+    [183, 183, 183],
+], dtype=np.float) / 255
+
+
+class Visualizations:
+    def __init__(self, env_name=None, update_every=10, server="http://localhost", disabled=False):
+        # Tracking data
+        self.last_update_timestamp = timer()
+        self.update_every = update_every
+        self.step_times = []
+        self.losses = []
+        self.eers = []
+        print("Updating the visualizations every %d steps." % update_every)
+
+        # If visdom is disabled TODO: use a better paradigm for that
+        self.disabled = disabled
+        if self.disabled:
+            return
+
+        # Set the environment name
+        now = str(datetime.now().strftime("%d-%m %Hh%M"))
+        if env_name is None:
+            self.env_name = now
+        else:
+            self.env_name = "%s (%s)" % (env_name, now)
+
+        # Connect to visdom and open the corresponding window in the browser
+        try:
+            self.vis = visdom.Visdom(server, env=self.env_name, raise_exceptions=True)
+        except ConnectionError:
+            raise Exception("No visdom server detected. Run the command \"visdom\" in your CLI to "
+                            "start it.")
+        # webbrowser.open("http://localhost:8097/env/" + self.env_name)
+
+        # Create the windows
+        self.loss_win = None
+        self.eer_win = None
+        # self.lr_win = None
+        self.implementation_win = None
+        self.projection_win = None
+        self.implementation_string = ""
+
+    def log_params(self):
+        if self.disabled:
+            return
+        from encoder import params_data
+        from encoder import params_model
+        param_string = "<b>Model parameters</b>:<br>"
+        for param_name in (p for p in dir(params_model) if not p.startswith("__")):
+            value = getattr(params_model, param_name)
+            param_string += "\t%s: %s<br>" % (param_name, value)
+        param_string += "<b>Data parameters</b>:<br>"
+        for param_name in (p for p in dir(params_data) if not p.startswith("__")):
+            value = getattr(params_data, param_name)
+            param_string += "\t%s: %s<br>" % (param_name, value)
+        self.vis.text(param_string, opts={"title": "Parameters"})
+
+    def log_dataset(self, dataset: SpeakerVerificationDataset):
+        if self.disabled:
+            return
+        dataset_string = ""
+        dataset_string += "<b>Speakers</b>: %s\n" % len(dataset.speakers)
+        dataset_string += "\n" + dataset.get_logs()
+        dataset_string = dataset_string.replace("\n", "<br>")
+        self.vis.text(dataset_string, opts={"title": "Dataset"})
+
+    def log_implementation(self, params):
+        if self.disabled:
+            return
+        implementation_string = ""
+        for param, value in params.items():
+            implementation_string += "<b>%s</b>: %s\n" % (param, value)
+            implementation_string = implementation_string.replace("\n", "<br>")
+        self.implementation_string = implementation_string
+        self.implementation_win = self.vis.text(
+            implementation_string,
+            opts={"title": "Training implementation"}
+        )
+
+    def update(self, loss, eer, step):
+        # Update the tracking data
+        now = timer()
+        self.step_times.append(1000 * (now - self.last_update_timestamp))
+        self.last_update_timestamp = now
+        self.losses.append(loss)
+        self.eers.append(eer)
+        print(".", end="")
+
+        # Update the plots every <update_every> steps
+        if step % self.update_every != 0:
+            return
+        time_string = "Step time:  mean: %5dms  std: %5dms" % \
+                      (int(np.mean(self.step_times)), int(np.std(self.step_times)))
+        print("\nStep %6d   Loss: %.4f   EER: %.4f   %s" %
+              (step, np.mean(self.losses), np.mean(self.eers), time_string))
+        if not self.disabled:
+            self.loss_win = self.vis.line(
+                [np.mean(self.losses)],
+                [step],
+                win=self.loss_win,
+                update="append" if self.loss_win else None,
+                opts=dict(
+                    legend=["Avg. loss"],
+                    xlabel="Step",
+                    ylabel="Loss",
+                    title="Loss",
+                )
+            )
+            self.eer_win = self.vis.line(
+                [np.mean(self.eers)],
+                [step],
+                win=self.eer_win,
+                update="append" if self.eer_win else None,
+                opts=dict(
+                    legend=["Avg. EER"],
+                    xlabel="Step",
+                    ylabel="EER",
+                    title="Equal error rate"
+                )
+            )
+            if self.implementation_win is not None:
+                self.vis.text(
+                    self.implementation_string + ("<b>%s</b>" % time_string),
+                    win=self.implementation_win,
+                    opts={"title": "Training implementation"},
+                )
+
+        # Reset the tracking
+        self.losses.clear()
+        self.eers.clear()
+        self.step_times.clear()
+
+    def draw_projections(self, embeds, utterances_per_speaker, step, out_fpath=None, max_speakers=10):
+        import matplotlib.pyplot as plt
+
+        max_speakers = min(max_speakers, len(colormap))
+        embeds = embeds[:max_speakers * utterances_per_speaker]
+
+        n_speakers = len(embeds) // utterances_per_speaker
+        ground_truth = np.repeat(np.arange(n_speakers), utterances_per_speaker)
+        colors = [colormap[i] for i in ground_truth]
+
+        reducer = umap.UMAP()
+        projected = reducer.fit_transform(embeds)
+        plt.scatter(projected[:, 0], projected[:, 1], c=colors)
+        plt.gca().set_aspect("equal", "datalim")
+        plt.title("UMAP projection (step %d)" % step)
+        if not self.disabled:
+            self.projection_win = self.vis.matplot(plt, win=self.projection_win)
+        if out_fpath is not None:
+            plt.savefig(out_fpath)
+        plt.clf()
+
+    def save(self):
+        if not self.disabled:
+            self.vis.save([self.env_name])