Merge pull request #204 from alibaba-damo-academy/dev_wjm

hnluo

2023-03-10 3fb2ca8378fc21b8f8dc3a451797a54ed42132d2

Merge pull request #204 from alibaba-damo-academy/dev_wjm

Dev wjm

 funasr/modules/eend_ola/__init__.py

 funasr/modules/eend_ola/encoder.py

New file
@@ -0,0 +1,127 @@
import math
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn


class MultiHeadSelfAttention(nn.Module):
    def __init__(self, n_units, h=8, dropout_rate=0.1):
        super(MultiHeadSelfAttention, self).__init__()
        self.linearQ = nn.Linear(n_units, n_units)
        self.linearK = nn.Linear(n_units, n_units)
        self.linearV = nn.Linear(n_units, n_units)
        self.linearO = nn.Linear(n_units, n_units)
        self.d_k = n_units // h
        self.h = h
        self.dropout = nn.Dropout(dropout_rate)

    def __call__(self, x, batch_size, x_mask):
        q = self.linearQ(x).view(batch_size, -1, self.h, self.d_k)
        k = self.linearK(x).view(batch_size, -1, self.h, self.d_k)
        v = self.linearV(x).view(batch_size, -1, self.h, self.d_k)
        scores = torch.matmul(
            q.permute(0, 2, 1, 3), k.permute(0, 2, 3, 1)) / math.sqrt(self.d_k)
        if x_mask is not None:
            x_mask = x_mask.unsqueeze(1)
            scores = scores.masked_fill(x_mask == 0, -1e9)
        self.att = F.softmax(scores, dim=3)
        p_att = self.dropout(self.att)
        x = torch.matmul(p_att, v.permute(0, 2, 1, 3))
        x = x.permute(0, 2, 1, 3).contiguous().view(-1, self.h * self.d_k)
        return self.linearO(x)


class PositionwiseFeedForward(nn.Module):
    def __init__(self, n_units, d_units, dropout_rate):
        super(PositionwiseFeedForward, self).__init__()
        self.linear1 = nn.Linear(n_units, d_units)
        self.linear2 = nn.Linear(d_units, n_units)
        self.dropout = nn.Dropout(dropout_rate)

    def __call__(self, x):
        return self.linear2(self.dropout(F.relu(self.linear1(x))))


class PositionalEncoding(torch.nn.Module):
    def __init__(self, d_model, dropout_rate, max_len=5000, reverse=False):
        super(PositionalEncoding, self).__init__()
        self.d_model = d_model
        self.reverse = reverse
        self.xscale = math.sqrt(self.d_model)
        self.dropout = torch.nn.Dropout(p=dropout_rate)
        self.pe = None
        self.extend_pe(torch.tensor(0.0).expand(1, max_len))

    def extend_pe(self, x):
        if self.pe is not None:
            if self.pe.size(1) >= x.size(1):
                if self.pe.dtype != x.dtype or self.pe.device != x.device:
                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
                return
        pe = torch.zeros(x.size(1), self.d_model)
        if self.reverse:
            position = torch.arange(
                x.size(1) - 1, -1, -1.0, dtype=torch.float32
            ).unsqueeze(1)
        else:
            position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, self.d_model, 2, dtype=torch.float32)
            * -(math.log(10000.0) / self.d_model)
        )
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)
        self.pe = pe.to(device=x.device, dtype=x.dtype)

    def forward(self, x: torch.Tensor):
        self.extend_pe(x)
        x = x * self.xscale + self.pe[:, : x.size(1)]
        return self.dropout(x)


class TransformerEncoder(nn.Module):
    def __init__(self, idim, n_layers, n_units,
                 e_units=2048, h=8, dropout_rate=0.1, use_pos_emb=False):
        super(TransformerEncoder, self).__init__()
        self.lnorm_in = nn.LayerNorm(n_units)
        self.n_layers = n_layers
        self.dropout = nn.Dropout(dropout_rate)
        for i in range(n_layers):
            setattr(self, '{}{:d}'.format("lnorm1_", i),
                    nn.LayerNorm(n_units))
            setattr(self, '{}{:d}'.format("self_att_", i),
                    MultiHeadSelfAttention(n_units, h))
            setattr(self, '{}{:d}'.format("lnorm2_", i),
                    nn.LayerNorm(n_units))
            setattr(self, '{}{:d}'.format("ff_", i),
                    PositionwiseFeedForward(n_units, e_units, dropout_rate))
        self.lnorm_out = nn.LayerNorm(n_units)
        if use_pos_emb:
            self.pos_enc = torch.nn.Sequential(
                torch.nn.Linear(idim, n_units),
                torch.nn.LayerNorm(n_units),
                torch.nn.Dropout(dropout_rate),
                torch.nn.ReLU(),
                PositionalEncoding(n_units, dropout_rate),
            )
        else:
            self.linear_in = nn.Linear(idim, n_units)
            self.pos_enc = None

    def __call__(self, x, x_mask=None):
        BT_size = x.shape[0] * x.shape[1]
        if self.pos_enc is not None:
            e = self.pos_enc(x)
            e = e.view(BT_size, -1)
        else:
            e = self.linear_in(x.reshape(BT_size, -1))
        for i in range(self.n_layers):
            e = getattr(self, '{}{:d}'.format("lnorm1_", i))(e)
            s = getattr(self, '{}{:d}'.format("self_att_", i))(e, x.shape[0], x_mask)
            e = e + self.dropout(s)
            e = getattr(self, '{}{:d}'.format("lnorm2_", i))(e)
            s = getattr(self, '{}{:d}'.format("ff_", i))(e)
            e = e + self.dropout(s)
        return self.lnorm_out(e)

 funasr/modules/eend_ola/encoder_decoder_attractor.py

New file
@@ -0,0 +1,50 @@
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn


class EncoderDecoderAttractor(nn.Module):

    def __init__(self, n_units, encoder_dropout=0.1, decoder_dropout=0.1):
        super(EncoderDecoderAttractor, self).__init__()
        self.enc0_dropout = nn.Dropout(encoder_dropout)
        self.encoder = nn.LSTM(n_units, n_units, 1, batch_first=True, dropout=encoder_dropout)
        self.dec0_dropout = nn.Dropout(decoder_dropout)
        self.decoder = nn.LSTM(n_units, n_units, 1, batch_first=True, dropout=decoder_dropout)
        self.counter = nn.Linear(n_units, 1)
        self.n_units = n_units

    def forward_core(self, xs, zeros):
        ilens = torch.from_numpy(np.array([x.shape[0] for x in xs])).to(torch.float32).to(xs[0].device)
        xs = [self.enc0_dropout(x) for x in xs]
        xs = nn.utils.rnn.pad_sequence(xs, batch_first=True, padding_value=-1)
        xs = nn.utils.rnn.pack_padded_sequence(xs, ilens, batch_first=True, enforce_sorted=False)
        _, (hx, cx) = self.encoder(xs)
        zlens = torch.from_numpy(np.array([z.shape[0] for z in zeros])).to(torch.float32).to(zeros[0].device)
        max_zlen = torch.max(zlens).to(torch.int).item()
        zeros = [self.enc0_dropout(z) for z in zeros]
        zeros = nn.utils.rnn.pad_sequence(zeros, batch_first=True, padding_value=-1)
        zeros = nn.utils.rnn.pack_padded_sequence(zeros, zlens, batch_first=True, enforce_sorted=False)
        attractors, (_, _) = self.decoder(zeros, (hx, cx))
        attractors = nn.utils.rnn.pad_packed_sequence(attractors, batch_first=True, padding_value=-1,
                                                      total_length=max_zlen)[0]
        attractors = [att[:zlens[i].to(torch.int).item()] for i, att in enumerate(attractors)]
        return attractors

    def forward(self, xs, n_speakers):
        zeros = [torch.zeros(n_spk + 1, self.n_units).to(torch.float32).to(xs[0].device) for n_spk in n_speakers]
        attractors = self.forward_core(xs, zeros)
        labels = torch.cat([torch.from_numpy(np.array([[1] * n_spk + [0]], np.float32)) for n_spk in n_speakers], dim=1)
        labels = labels.to(xs[0].device)
        logit = torch.cat([self.counter(att).view(-1, n_spk + 1) for att, n_spk in zip(attractors, n_speakers)], dim=1)
        loss = F.binary_cross_entropy(torch.sigmoid(logit), labels)

        attractors = [att[slice(0, att.shape[0] - 1)] for att in attractors]
        return loss, attractors

    def estimate(self, xs, max_n_speakers=15):
        zeros = [torch.zeros(max_n_speakers, self.n_units).to(torch.float32).to(xs[0].device) for _ in xs]
        attractors = self.forward_core(xs, zeros)
        probs = [torch.sigmoid(torch.flatten(self.counter(att))) for att in attractors]
        return attractors, probs

 funasr/modules/eend_ola/utils/losses.py

New file
@@ -0,0 +1,67 @@
import numpy as np
import torch
import torch.nn.functional as F
from itertools import permutations
from torch import nn


def standard_loss(ys, ts, label_delay=0):
    losses = [F.binary_cross_entropy(torch.sigmoid(y), t) * len(y) for y, t in zip(ys, ts)]
    loss = torch.sum(torch.stack(losses))
    n_frames = torch.from_numpy(np.array(np.sum([t.shape[0] for t in ts]))).to(torch.float32).to(ys[0].device)
    loss = loss / n_frames
    return loss


def batch_pit_n_speaker_loss(ys, ts, n_speakers_list):
    max_n_speakers = ts[0].shape[1]
    olens = [y.shape[0] for y in ys]
    ys = nn.utils.rnn.pad_sequence(ys, batch_first=True, padding_value=-1)
    ys_mask = [torch.ones(olen).to(ys.device) for olen in olens]
    ys_mask = torch.nn.utils.rnn.pad_sequence(ys_mask, batch_first=True, padding_value=0).unsqueeze(-1)

    losses = []
    for shift in range(max_n_speakers):
        ts_roll = [torch.roll(t, -shift, dims=1) for t in ts]
        ts_roll = nn.utils.rnn.pad_sequence(ts_roll, batch_first=True, padding_value=-1)
        loss = F.binary_cross_entropy(torch.sigmoid(ys), ts_roll, reduction='none')
        if ys_mask is not None:
            loss = loss * ys_mask
        loss = torch.sum(loss, dim=1)
        losses.append(loss)
    losses = torch.stack(losses, dim=2)

    perms = np.array(list(permutations(range(max_n_speakers)))).astype(np.float32)
    perms = torch.from_numpy(perms).to(losses.device)
    y_ind = torch.arange(max_n_speakers, dtype=torch.float32, device=losses.device)
    t_inds = torch.fmod(perms - y_ind, max_n_speakers).to(torch.long)

    losses_perm = []
    for t_ind in t_inds:
        losses_perm.append(
            torch.mean(losses[:, y_ind.to(torch.long), t_ind], dim=1))
    losses_perm = torch.stack(losses_perm, dim=1)

    def select_perm_indices(num, max_num):
        perms = list(permutations(range(max_num)))
        sub_perms = list(permutations(range(num)))
        return [
            [x[:num] for x in perms].index(perm)
            for perm in sub_perms]

    masks = torch.full_like(losses_perm, device=losses.device, fill_value=float('inf'))
    for i, t in enumerate(ts):
        n_speakers = n_speakers_list[i]
        indices = select_perm_indices(n_speakers, max_n_speakers)
        masks[i, indices] = 0
    losses_perm += masks

    min_loss = torch.sum(torch.min(losses_perm, dim=1)[0])
    n_frames = torch.from_numpy(np.array(np.sum([t.shape[0] for t in ts]))).to(losses.device)
    min_loss = min_loss / n_frames

    min_indices = torch.argmin(losses_perm, dim=1)
    labels_perm = [t[:, perms[idx].to(torch.long)] for t, idx in zip(ts, min_indices)]
    labels_perm = [t[:, :n_speakers] for t, n_speakers in zip(labels_perm, n_speakers_list)]

    return min_loss, labels_perm

 funasr/modules/eend_ola/utils/power.py

New file
@@ -0,0 +1,95 @@
import numpy as np
import torch
import torch.multiprocessing
import torch.nn.functional as F
from itertools import combinations
from itertools import permutations


def generate_mapping_dict(max_speaker_num=6, max_olp_speaker_num=3):
    all_kinds = []
    all_kinds.append(0)
    for i in range(max_olp_speaker_num):
        selected_num = i + 1
        coms = np.array(list(combinations(np.arange(max_speaker_num), selected_num)))
        for com in coms:
            tmp = np.zeros(max_speaker_num)
            tmp[com] = 1
            item = int(raw_dec_trans(tmp.reshape(1, -1), max_speaker_num)[0])
            all_kinds.append(item)
    all_kinds_order = sorted(all_kinds)

    mapping_dict = {}
    mapping_dict['dec2label'] = {}
    mapping_dict['label2dec'] = {}
    for i in range(len(all_kinds_order)):
        dec = all_kinds_order[i]
        mapping_dict['dec2label'][dec] = i
        mapping_dict['label2dec'][i] = dec
    oov_id = len(all_kinds_order)
    mapping_dict['oov'] = oov_id
    return mapping_dict


def raw_dec_trans(x, max_speaker_num):
    num_list = []
    for i in range(max_speaker_num):
        num_list.append(x[:, i])
    base = 1
    T = x.shape[0]
    res = np.zeros((T))
    for num in num_list:
        res += num * base
        base = base * 2
    return res


def mapping_func(num, mapping_dict):
    if num in mapping_dict['dec2label'].keys():
        label = mapping_dict['dec2label'][num]
    else:
        label = mapping_dict['oov']
    return label


def dec_trans(x, max_speaker_num, mapping_dict):
    num_list = []
    for i in range(max_speaker_num):
        num_list.append(x[:, i])
    base = 1
    T = x.shape[0]
    res = np.zeros((T))
    for num in num_list:
        res += num * base
        base = base * 2
    res = np.array([mapping_func(i, mapping_dict) for i in res])
    return res


def create_powerlabel(label, mapping_dict, max_speaker_num=6, max_olp_speaker_num=3):
    T, C = label.shape
    padding_label = np.zeros((T, max_speaker_num))
    padding_label[:, :C] = label
    out_label = dec_trans(padding_label, max_speaker_num, mapping_dict)
    out_label = torch.from_numpy(out_label)
    return out_label


def generate_perm_pse(label, n_speaker, mapping_dict, max_speaker_num, max_olp_speaker_num=3):
    perms = np.array(list(permutations(range(n_speaker)))).astype(np.float32)
    perms = torch.from_numpy(perms).to(label.device).to(torch.int64)
    perm_labels = [label[:, perm] for perm in perms]
    perm_pse_labels = [create_powerlabel(perm_label.cpu().numpy(), mapping_dict, max_speaker_num).
                           to(perm_label.device, non_blocking=True) for perm_label in perm_labels]
    return perm_labels, perm_pse_labels


def generate_min_pse(label, n_speaker, mapping_dict, max_speaker_num, pse_logit, max_olp_speaker_num=3):
    perm_labels, perm_pse_labels = generate_perm_pse(label, n_speaker, mapping_dict, max_speaker_num,
                                                     max_olp_speaker_num=max_olp_speaker_num)
    losses = [F.cross_entropy(input=pse_logit, target=perm_pse_label.to(torch.long)) * len(pse_logit)
              for perm_pse_label in perm_pse_labels]
    loss = torch.stack(losses)
    min_index = torch.argmin(loss)
    selected_perm_label, selected_pse_label = perm_labels[min_index], perm_pse_labels[min_index]
    return selected_perm_label, selected_pse_label

 funasr/modules/eend_ola/utils/report.py

New file
@@ -0,0 +1,159 @@
import copy
import numpy as np
import time
import torch
from eend.utils.power import create_powerlabel
from itertools import combinations

metrics = [
    ('diarization_error', 'speaker_scored', 'DER'),
    ('speech_miss', 'speech_scored', 'SAD_MR'),
    ('speech_falarm', 'speech_scored', 'SAD_FR'),
    ('speaker_miss', 'speaker_scored', 'MI'),
    ('speaker_falarm', 'speaker_scored', 'FA'),
    ('speaker_error', 'speaker_scored', 'CF'),
    ('correct', 'frames', 'accuracy')
]


def recover_prediction(y, n_speaker):
    if n_speaker <= 1:
        return y
    elif n_speaker == 2:
        com_index = torch.from_numpy(
            np.array(list(combinations(np.arange(n_speaker), 2)))).to(
            y.dtype)
        num_coms = com_index.shape[0]
        y_single = y[:, :-num_coms]
        y_olp = y[:, -num_coms:]
        olp_map_index = torch.where(y_olp > 0.5)
        olp_map_index = torch.stack(olp_map_index, dim=1)
        com_map_index = com_index[olp_map_index[:, -1]]
        speaker_map_index = torch.from_numpy(np.array(com_map_index)).view(-1).to(torch.int64)
        frame_map_index = olp_map_index[:, 0][:, None].repeat([1, 2]).view(-1).to(
            torch.int64)
        y_single[frame_map_index] = 0
        y_single[frame_map_index, speaker_map_index] = 1
        return y_single
    else:
        olp2_com_index = torch.from_numpy(np.array(list(combinations(np.arange(n_speaker), 2)))).to(y.dtype)
        olp2_num_coms = olp2_com_index.shape[0]
        olp3_com_index = torch.from_numpy(np.array(list(combinations(np.arange(n_speaker), 3)))).to(y.dtype)
        olp3_num_coms = olp3_com_index.shape[0]
        y_single = y[:, :n_speaker]
        y_olp2 = y[:, n_speaker:n_speaker + olp2_num_coms]
        y_olp3 = y[:, -olp3_num_coms:]

        olp3_map_index = torch.where(y_olp3 > 0.5)
        olp3_map_index = torch.stack(olp3_map_index, dim=1)
        olp3_com_map_index = olp3_com_index[olp3_map_index[:, -1]]
        olp3_speaker_map_index = torch.from_numpy(np.array(olp3_com_map_index)).view(-1).to(torch.int64)
        olp3_frame_map_index = olp3_map_index[:, 0][:, None].repeat([1, 3]).view(-1).to(torch.int64)
        y_single[olp3_frame_map_index] = 0
        y_single[olp3_frame_map_index, olp3_speaker_map_index] = 1
        y_olp2[olp3_frame_map_index] = 0

        olp2_map_index = torch.where(y_olp2 > 0.5)
        olp2_map_index = torch.stack(olp2_map_index, dim=1)
        olp2_com_map_index = olp2_com_index[olp2_map_index[:, -1]]
        olp2_speaker_map_index = torch.from_numpy(np.array(olp2_com_map_index)).view(-1).to(torch.int64)
        olp2_frame_map_index = olp2_map_index[:, 0][:, None].repeat([1, 2]).view(-1).to(torch.int64)
        y_single[olp2_frame_map_index] = 0
        y_single[olp2_frame_map_index, olp2_speaker_map_index] = 1
        return y_single


class PowerReporter():
    def __init__(self, valid_data_loader, mapping_dict, max_n_speaker):
        valid_data_loader_cp = copy.deepcopy(valid_data_loader)
        self.valid_data_loader = valid_data_loader_cp
        del valid_data_loader
        self.mapping_dict = mapping_dict
        self.max_n_speaker = max_n_speaker

    def report(self, model, eidx, device):
        self.report_val(model, eidx, device)

    def report_val(self, model, eidx, device):
        model.eval()
        ud_valid_start = time.time()
        valid_res, valid_loss, stats_keys, vad_valid_accuracy = self.report_core(model, self.valid_data_loader, device)

        # Epoch Display
        valid_der = valid_res['diarization_error'] / valid_res['speaker_scored']
        valid_accuracy = valid_res['correct'].to(torch.float32) / valid_res['frames'] * 100
        vad_valid_accuracy = vad_valid_accuracy * 100
        print('Epoch ', eidx + 1, 'Valid Loss ', valid_loss, 'Valid_DER %.5f' % valid_der,
              'Valid_Accuracy %.5f%% ' % valid_accuracy, 'VAD_Valid_Accuracy %.5f%% ' % vad_valid_accuracy)
        ud_valid = (time.time() - ud_valid_start) / 60.
        print('Valid cost time ... ', ud_valid)

    def inv_mapping_func(self, label, mapping_dict):
        if not isinstance(label, int):
            label = int(label)
        if label in mapping_dict['label2dec'].keys():
            num = mapping_dict['label2dec'][label]
        else:
            num = -1
        return num

    def report_core(self, model, data_loader, device):
        res = {}
        for item in metrics:
            res[item[0]] = 0.
            res[item[1]] = 0.
        with torch.no_grad():
            loss_s = 0.
            uidx = 0
            for xs, ts, orders in data_loader:
                xs = [x.to(device) for x in xs]
                ts = [t.to(device) for t in ts]
                orders = [o.to(device) for o in orders]
                loss, pit_loss, mpit_loss, att_loss, ys, logits, labels, attractors = model(xs, ts, orders)
                loss_s += loss.item()
                uidx += 1

                for logit, t, att in zip(logits, labels, attractors):
                    pred = torch.argmax(torch.softmax(logit, dim=-1), dim=-1)  # (T, )
                    oov_index = torch.where(pred == self.mapping_dict['oov'])[0]
                    for i in oov_index:
                        if i > 0:
                            pred[i] = pred[i - 1]
                        else:
                            pred[i] = 0
                    pred = [self.inv_mapping_func(i, self.mapping_dict) for i in pred]
                    decisions = [bin(num)[2:].zfill(self.max_n_speaker)[::-1] for num in pred]
                    decisions = torch.from_numpy(
                        np.stack([np.array([int(i) for i in dec]) for dec in decisions], axis=0)).to(att.device).to(
                        torch.float32)
                    decisions = decisions[:, :att.shape[0]]

                    stats = self.calc_diarization_error(decisions, t)
                    res['speaker_scored'] += stats['speaker_scored']
                    res['speech_scored'] += stats['speech_scored']
                    res['frames'] += stats['frames']
                    for item in metrics:
                        res[item[0]] += stats[item[0]]
                loss_s /= uidx
                vad_acc = 0

        return res, loss_s, stats.keys(), vad_acc

    def calc_diarization_error(self, decisions, label, label_delay=0):
        label = label[:len(label) - label_delay, ...]
        n_ref = torch.sum(label, dim=-1)
        n_sys = torch.sum(decisions, dim=-1)
        res = {}
        res['speech_scored'] = torch.sum(n_ref > 0)
        res['speech_miss'] = torch.sum((n_ref > 0) & (n_sys == 0))
        res['speech_falarm'] = torch.sum((n_ref == 0) & (n_sys > 0))
        res['speaker_scored'] = torch.sum(n_ref)
        res['speaker_miss'] = torch.sum(torch.max(n_ref - n_sys, torch.zeros_like(n_ref)))
        res['speaker_falarm'] = torch.sum(torch.max(n_sys - n_ref, torch.zeros_like(n_ref)))
        n_map = torch.sum(((label == 1) & (decisions == 1)), dim=-1).to(torch.float32)
        res['speaker_error'] = torch.sum(torch.min(n_ref, n_sys) - n_map)
        res['correct'] = torch.sum(label == decisions) / label.shape[1]
        res['diarization_error'] = (
                res['speaker_miss'] + res['speaker_falarm'] + res['speaker_error'])
        res['frames'] = len(label)
        return res