import torch
|
from torch import nn
|
|
from funasr.modules.nets_utils import make_pad_mask
|
|
class CifPredictor(nn.Module):
|
def __init__(self, idim, l_order, r_order, threshold=1.0, dropout=0.1, smooth_factor=1.0, noise_threshold=0, tail_threshold=0.45):
|
super(CifPredictor, self).__init__()
|
|
self.pad = nn.ConstantPad1d((l_order, r_order), 0)
|
self.cif_conv1d = nn.Conv1d(idim, idim, l_order + r_order + 1, groups=idim)
|
self.cif_output = nn.Linear(idim, 1)
|
self.dropout = torch.nn.Dropout(p=dropout)
|
self.threshold = threshold
|
self.smooth_factor = smooth_factor
|
self.noise_threshold = noise_threshold
|
|
def forward(self, hidden, target_label=None, mask=None, ignore_id=-1, mask_chunk_predictor=None,
|
target_label_length=None):
|
h = hidden
|
context = h.transpose(1, 2)
|
queries = self.pad(context)
|
memory = self.cif_conv1d(queries)
|
output = memory + context
|
output = self.dropout(output)
|
output = output.transpose(1, 2)
|
output = torch.relu(output)
|
output = self.cif_output(output)
|
alphas = torch.sigmoid(output)
|
alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)
|
if mask is not None:
|
alphas = alphas * mask.transpose(-1, -2).float()
|
if mask_chunk_predictor is not None:
|
alphas = alphas * mask_chunk_predictor
|
alphas = alphas.squeeze(-1)
|
if target_label_length is not None:
|
target_length = target_label_length
|
elif target_label is not None:
|
target_length = (target_label != ignore_id).float().sum(-1)
|
else:
|
target_length = None
|
token_num = alphas.sum(-1)
|
if target_length is not None:
|
alphas *= (target_length / token_num)[:, None].repeat(1, alphas.size(1))
|
acoustic_embeds, cif_peak = cif(hidden, alphas, self.threshold)
|
return acoustic_embeds, token_num, alphas, cif_peak
|
|
def gen_frame_alignments(self,
|
alphas: torch.Tensor = None,
|
encoder_sequence_length: torch.Tensor = None):
|
batch_size, maximum_length = alphas.size()
|
int_type = torch.int32
|
|
is_training = self.training
|
if is_training:
|
token_num = torch.round(torch.sum(alphas, dim=1)).type(int_type)
|
else:
|
token_num = torch.floor(torch.sum(alphas, dim=1)).type(int_type)
|
|
max_token_num = torch.max(token_num).item()
|
|
alphas_cumsum = torch.cumsum(alphas, dim=1)
|
alphas_cumsum = torch.floor(alphas_cumsum).type(int_type)
|
alphas_cumsum = alphas_cumsum[:, None, :].repeat(1, max_token_num, 1)
|
|
index = torch.ones([batch_size, max_token_num], dtype=int_type)
|
index = torch.cumsum(index, dim=1)
|
index = index[:, :, None].repeat(1, 1, maximum_length).to(alphas_cumsum.device)
|
|
index_div = torch.floor(torch.true_divide(alphas_cumsum, index)).type(int_type)
|
index_div_bool_zeros = index_div.eq(0)
|
index_div_bool_zeros_count = torch.sum(index_div_bool_zeros, dim=-1) + 1
|
index_div_bool_zeros_count = torch.clamp(index_div_bool_zeros_count, 0, encoder_sequence_length.max())
|
token_num_mask = (~make_pad_mask(token_num, maxlen=max_token_num)).to(token_num.device)
|
index_div_bool_zeros_count *= token_num_mask
|
|
index_div_bool_zeros_count_tile = index_div_bool_zeros_count[:, :, None].repeat(1, 1, maximum_length)
|
ones = torch.ones_like(index_div_bool_zeros_count_tile)
|
zeros = torch.zeros_like(index_div_bool_zeros_count_tile)
|
ones = torch.cumsum(ones, dim=2)
|
cond = index_div_bool_zeros_count_tile == ones
|
index_div_bool_zeros_count_tile = torch.where(cond, zeros, ones)
|
|
index_div_bool_zeros_count_tile_bool = index_div_bool_zeros_count_tile.type(torch.bool)
|
index_div_bool_zeros_count_tile = 1 - index_div_bool_zeros_count_tile_bool.type(int_type)
|
index_div_bool_zeros_count_tile_out = torch.sum(index_div_bool_zeros_count_tile, dim=1)
|
index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out.type(int_type)
|
predictor_mask = (~make_pad_mask(encoder_sequence_length, maxlen=encoder_sequence_length.max())).type(
|
int_type).to(encoder_sequence_length.device)
|
index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out * predictor_mask
|
|
predictor_alignments = index_div_bool_zeros_count_tile_out
|
predictor_alignments_length = predictor_alignments.sum(-1).type(encoder_sequence_length.dtype)
|
return predictor_alignments.detach(), predictor_alignments_length.detach()
|
|
|
class CifPredictorV2(nn.Module):
|
def __init__(self, idim, l_order, r_order, threshold=1.0, dropout=0.1, smooth_factor=1.0, noise_threshold=0,
|
tail_threshold=0.0):
|
super(CifPredictorV2, self).__init__()
|
|
self.pad = nn.ConstantPad1d((l_order, r_order), 0)
|
self.cif_conv1d = nn.Conv1d(idim, idim, l_order + r_order + 1)
|
self.cif_output = nn.Linear(idim, 1)
|
self.dropout = torch.nn.Dropout(p=dropout)
|
self.threshold = threshold
|
self.smooth_factor = smooth_factor
|
self.noise_threshold = noise_threshold
|
self.tail_threshold = tail_threshold
|
|
def forward(self, hidden, target_label=None, mask=None, ignore_id=-1, mask_chunk_predictor=None,
|
target_label_length=None):
|
h = hidden
|
context = h.transpose(1, 2)
|
queries = self.pad(context)
|
output = torch.relu(self.cif_conv1d(queries))
|
output = output.transpose(1, 2)
|
|
output = self.cif_output(output)
|
alphas = torch.sigmoid(output)
|
alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)
|
if mask is not None:
|
alphas = alphas * mask.transpose(-1, -2).float()
|
if mask_chunk_predictor is not None:
|
alphas = alphas * mask_chunk_predictor
|
alphas = alphas.squeeze(-1)
|
if target_label_length is not None:
|
target_length = target_label_length
|
elif target_label is not None:
|
target_length = (target_label != ignore_id).float().sum(-1)
|
else:
|
target_length = None
|
token_num = alphas.sum(-1)
|
if target_length is not None:
|
alphas *= (target_length / token_num)[:, None].repeat(1, alphas.size(1))
|
elif self.tail_threshold > 0.0:
|
hidden, alphas, token_num = self.tail_process_fn(hidden, alphas, token_num)
|
|
acoustic_embeds, cif_peak = cif(hidden, alphas, self.threshold)
|
if target_length is None and self.tail_threshold > 0.0:
|
token_num_int = torch.max(token_num).type(torch.int32).item()
|
acoustic_embeds = acoustic_embeds[:, :token_num_int, :]
|
|
return acoustic_embeds, token_num, alphas, cif_peak
|
|
def tail_process_fn(self, hidden, alphas, token_num=None):
|
b, t, d = hidden.size()
|
tail_threshold = self.tail_threshold
|
tail_threshold = torch.tensor([tail_threshold], dtype=alphas.dtype).to(alphas.device)
|
tail_threshold = tail_threshold.unsqueeze(0).repeat(b, 1)
|
alphas = torch.cat([alphas, tail_threshold], dim=1)
|
zeros = torch.zeros((b, 1, d), dtype=hidden.dtype).to(hidden.device)
|
hidden = torch.cat([hidden, zeros], dim=1)
|
token_num = alphas.sum(dim=-1)
|
token_num_floor = torch.floor(token_num)
|
|
return hidden, alphas, token_num_floor
|
|
def gen_frame_alignments(self,
|
alphas: torch.Tensor = None,
|
encoder_sequence_length: torch.Tensor = None):
|
batch_size, maximum_length = alphas.size()
|
int_type = torch.int32
|
|
is_training = self.training
|
if is_training:
|
token_num = torch.round(torch.sum(alphas, dim=1)).type(int_type)
|
else:
|
token_num = torch.floor(torch.sum(alphas, dim=1)).type(int_type)
|
|
max_token_num = torch.max(token_num).item()
|
|
alphas_cumsum = torch.cumsum(alphas, dim=1)
|
alphas_cumsum = torch.floor(alphas_cumsum).type(int_type)
|
alphas_cumsum = alphas_cumsum[:, None, :].repeat(1, max_token_num, 1)
|
|
index = torch.ones([batch_size, max_token_num], dtype=int_type)
|
index = torch.cumsum(index, dim=1)
|
index = index[:, :, None].repeat(1, 1, maximum_length).to(alphas_cumsum.device)
|
|
index_div = torch.floor(torch.true_divide(alphas_cumsum, index)).type(int_type)
|
index_div_bool_zeros = index_div.eq(0)
|
index_div_bool_zeros_count = torch.sum(index_div_bool_zeros, dim=-1) + 1
|
index_div_bool_zeros_count = torch.clamp(index_div_bool_zeros_count, 0, encoder_sequence_length.max())
|
token_num_mask = (~make_pad_mask(token_num, maxlen=max_token_num)).to(token_num.device)
|
index_div_bool_zeros_count *= token_num_mask
|
|
index_div_bool_zeros_count_tile = index_div_bool_zeros_count[:, :, None].repeat(1, 1, maximum_length)
|
ones = torch.ones_like(index_div_bool_zeros_count_tile)
|
zeros = torch.zeros_like(index_div_bool_zeros_count_tile)
|
ones = torch.cumsum(ones, dim=2)
|
cond = index_div_bool_zeros_count_tile == ones
|
index_div_bool_zeros_count_tile = torch.where(cond, zeros, ones)
|
|
index_div_bool_zeros_count_tile_bool = index_div_bool_zeros_count_tile.type(torch.bool)
|
index_div_bool_zeros_count_tile = 1 - index_div_bool_zeros_count_tile_bool.type(int_type)
|
index_div_bool_zeros_count_tile_out = torch.sum(index_div_bool_zeros_count_tile, dim=1)
|
index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out.type(int_type)
|
predictor_mask = (~make_pad_mask(encoder_sequence_length, maxlen=encoder_sequence_length.max())).type(
|
int_type).to(encoder_sequence_length.device)
|
index_div_bool_zeros_count_tile_out = index_div_bool_zeros_count_tile_out * predictor_mask
|
|
predictor_alignments = index_div_bool_zeros_count_tile_out
|
predictor_alignments_length = predictor_alignments.sum(-1).type(encoder_sequence_length.dtype)
|
return predictor_alignments.detach(), predictor_alignments_length.detach()
|
|
|
class mae_loss(nn.Module):
|
|
def __init__(self, normalize_length=False):
|
super(mae_loss, self).__init__()
|
self.normalize_length = normalize_length
|
self.criterion = torch.nn.L1Loss(reduction='sum')
|
|
def forward(self, token_length, pre_token_length):
|
loss_token_normalizer = token_length.size(0)
|
if self.normalize_length:
|
loss_token_normalizer = token_length.sum().type(torch.float32)
|
loss = self.criterion(token_length, pre_token_length)
|
loss = loss / loss_token_normalizer
|
return loss
|
|
|
def cif(hidden, alphas, threshold):
|
batch_size, len_time, hidden_size = hidden.size()
|
|
# loop varss
|
integrate = torch.zeros([batch_size], device=hidden.device)
|
frame = torch.zeros([batch_size, hidden_size], device=hidden.device)
|
# intermediate vars along time
|
list_fires = []
|
list_frames = []
|
|
for t in range(len_time):
|
alpha = alphas[:, t]
|
distribution_completion = torch.ones([batch_size], device=hidden.device) - integrate
|
|
integrate += alpha
|
list_fires.append(integrate)
|
|
fire_place = integrate >= threshold
|
integrate = torch.where(fire_place,
|
integrate - torch.ones([batch_size], device=hidden.device),
|
integrate)
|
cur = torch.where(fire_place,
|
distribution_completion,
|
alpha)
|
remainds = alpha - cur
|
|
frame += cur[:, None] * hidden[:, t, :]
|
list_frames.append(frame)
|
frame = torch.where(fire_place[:, None].repeat(1, hidden_size),
|
remainds[:, None] * hidden[:, t, :],
|
frame)
|
|
fires = torch.stack(list_fires, 1)
|
frames = torch.stack(list_frames, 1)
|
list_ls = []
|
len_labels = torch.round(alphas.sum(-1)).int()
|
max_label_len = len_labels.max()
|
for b in range(batch_size):
|
fire = fires[b, :]
|
l = torch.index_select(frames[b, :, :], 0, torch.nonzero(fire >= threshold).squeeze())
|
pad_l = torch.zeros([max_label_len - l.size(0), hidden_size], device=hidden.device)
|
list_ls.append(torch.cat([l, pad_l], 0))
|
return torch.stack(list_ls, 0), fires
|
