speaker_encoder/model.py

from speaker_encoder.params_model import *
from speaker_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,     # 40
                            hidden_size=model_hidden_size, # 256 
                            num_layers=model_num_layers,   # 3 
                            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)
        
        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)

        # 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)

        # 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