fp16

游雁

2024-02-23 8827e26b8d487f123f8d7d5cbd8d00b81dcefcff

 funasr/models/paraformer/cif_predictor.py

@@ -1,20 +1,25 @@
#!/usr/bin/env python3

# -*- encoding: utf-8 -*-

# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.

#  MIT License  (https://opensource.org/licenses/MIT)



import torch

from torch import nn

from torch import Tensor

import logging

import numpy as np



from funasr.register import tables

from funasr.train_utils.device_funcs import to_device

from funasr.models.transformer.utils.nets_utils import make_pad_mask

from funasr.models.scama.utils import sequence_mask

from typing import Optional, Tuple



class CifPredictor(nn.Module):



@tables.register("predictor_classes", "CifPredictor")

class CifPredictor(torch.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__()

        super().__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.pad = torch.nn.ConstantPad1d((l_order, r_order), 0)

        self.cif_conv1d = torch.nn.Conv1d(idim, idim, l_order + r_order + 1, groups=idim)

        self.cif_output = torch.nn.Linear(idim, 1)

        self.dropout = torch.nn.Dropout(p=dropout)

        self.threshold = threshold

        self.smooth_factor = smooth_factor

@@ -133,8 +138,8 @@
        predictor_alignments_length = predictor_alignments.sum(-1).type(encoder_sequence_length.dtype)

        return predictor_alignments.detach(), predictor_alignments_length.detach()





class CifPredictorV2(nn.Module):

@tables.register("predictor_classes", "CifPredictorV2")

class CifPredictorV2(torch.nn.Module):

    def __init__(self,

                 idim,

                 l_order,

@@ -150,9 +155,9 @@
                 ):

        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.pad = torch.nn.ConstantPad1d((l_order, r_order), 0)

        self.cif_conv1d = torch.nn.Conv1d(idim, idim, l_order + r_order + 1)

        self.cif_output = torch.nn.Linear(idim, 1)

        self.dropout = torch.nn.Dropout(p=dropout)

        self.threshold = threshold

        self.smooth_factor = smooth_factor

@@ -181,7 +186,7 @@
        alphas = alphas.squeeze(-1)

        mask = mask.squeeze(-1)

        if target_label_length is not None:

            target_length = target_label_length

            target_length = target_label_length.squeeze(-1)

        elif target_label is not None:

            target_length = (target_label != ignore_id).float().sum(-1)

        else:

@@ -202,7 +207,8 @@


        return acoustic_embeds, token_num, alphas, cif_peak



    def forward_chunk(self, hidden, cache=None):

    def forward_chunk(self, hidden, cache=None, **kwargs):

        is_final = kwargs.get("is_final", False)

        batch_size, len_time, hidden_size = hidden.shape

        h = hidden

        context = h.transpose(1, 2)

@@ -223,14 +229,14 @@


        if cache is not None and "chunk_size" in cache:

            alphas[:, :cache["chunk_size"][0]] = 0.0

            if "is_final" in cache and not cache["is_final"]:

            if not is_final:

                alphas[:, sum(cache["chunk_size"][:2]):] = 0.0

        if cache is not None and "cif_alphas" in cache and "cif_hidden" in cache:

            cache["cif_hidden"] = to_device(cache["cif_hidden"], device=hidden.device)

            cache["cif_alphas"] = to_device(cache["cif_alphas"], device=alphas.device)

            hidden = torch.cat((cache["cif_hidden"], hidden), dim=1)

            alphas = torch.cat((cache["cif_alphas"], alphas), dim=1)

        if cache is not None and "is_final" in cache and cache["is_final"]:

        if cache is not None and is_final:

            tail_hidden = torch.zeros((batch_size, 1, hidden_size), device=hidden.device)

            tail_alphas = torch.tensor([[self.tail_threshold]], device=alphas.device)

            tail_alphas = torch.tile(tail_alphas, (batch_size, 1))

@@ -274,7 +280,7 @@


        max_token_len = max(token_length)

        if max_token_len == 0:

             return hidden, torch.stack(token_length, 0)

             return hidden, torch.stack(token_length, 0), None, None

        list_ls = []

        for b in range(batch_size):

            pad_frames = torch.zeros((max_token_len - token_length[b], hidden_size), device=alphas.device)

@@ -288,7 +294,7 @@
        cache["cif_alphas"] = torch.unsqueeze(cache["cif_alphas"], axis=0)

        cache["cif_hidden"] = torch.stack(cache_hiddens, axis=0)

        cache["cif_hidden"] = torch.unsqueeze(cache["cif_hidden"], axis=0)

        return torch.stack(list_ls, 0), torch.stack(token_length, 0)

        return torch.stack(list_ls, 0), torch.stack(token_length, 0), None, None





    def tail_process_fn(self, hidden, alphas, token_num=None, mask=None):

@@ -422,7 +428,7 @@
        return var_dict_torch_update





class mae_loss(nn.Module):

class mae_loss(torch.nn.Module):



    def __init__(self, normalize_length=False):

        super(mae_loss, self).__init__()

@@ -505,372 +511,3 @@
    fires = torch.stack(list_fires, 1)

    return fires





class CifPredictorV3(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,

                 tf2torch_tensor_name_prefix_torch="predictor",

                 tf2torch_tensor_name_prefix_tf="seq2seq/cif",

                 smooth_factor2=1.0,

                 noise_threshold2=0,

                 upsample_times=5,

                 upsample_type="cnn",

                 use_cif1_cnn=True,

                 tail_mask=True,

                 ):

        super(CifPredictorV3, 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

        self.tf2torch_tensor_name_prefix_torch = tf2torch_tensor_name_prefix_torch

        self.tf2torch_tensor_name_prefix_tf = tf2torch_tensor_name_prefix_tf



        self.upsample_times = upsample_times

        self.upsample_type = upsample_type

        self.use_cif1_cnn = use_cif1_cnn

        if self.upsample_type == 'cnn':

            self.upsample_cnn = nn.ConvTranspose1d(idim, idim, self.upsample_times, self.upsample_times)

            self.cif_output2 = nn.Linear(idim, 1)

        elif self.upsample_type == 'cnn_blstm':

            self.upsample_cnn = nn.ConvTranspose1d(idim, idim, self.upsample_times, self.upsample_times)

            self.blstm = nn.LSTM(idim, idim, 1, bias=True, batch_first=True, dropout=0.0, bidirectional=True)

            self.cif_output2 = nn.Linear(idim*2, 1)

        elif self.upsample_type == 'cnn_attn':

            self.upsample_cnn = nn.ConvTranspose1d(idim, idim, self.upsample_times, self.upsample_times)

            from funasr.models.encoder.transformer_encoder import EncoderLayer as TransformerEncoderLayer

            from funasr.models.transformer.attention import MultiHeadedAttention

            from funasr.models.transformer.positionwise_feed_forward import PositionwiseFeedForward

            positionwise_layer_args = (

                idim,

                idim*2,

                0.1,

            )

            self.self_attn = TransformerEncoderLayer(

                idim,

                MultiHeadedAttention(

                    4, idim, 0.1

                ),

                PositionwiseFeedForward(*positionwise_layer_args),

                0.1,

                True, #normalize_before,

                False, #concat_after,

            )

            self.cif_output2 = nn.Linear(idim, 1)

        self.smooth_factor2 = smooth_factor2

        self.noise_threshold2 = noise_threshold2



    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))



        # alphas2 is an extra head for timestamp prediction

        if not self.use_cif1_cnn:

            _output = context

        else:

            _output = output

        if self.upsample_type == 'cnn':

            output2 = self.upsample_cnn(_output)

            output2 = output2.transpose(1,2)

        elif self.upsample_type == 'cnn_blstm':

            output2 = self.upsample_cnn(_output)

            output2 = output2.transpose(1,2)

            output2, (_, _) = self.blstm(output2)

        elif self.upsample_type == 'cnn_attn':

            output2 = self.upsample_cnn(_output)

            output2 = output2.transpose(1,2)

            output2, _ = self.self_attn(output2, mask)

        # import pdb; pdb.set_trace()

        alphas2 = torch.sigmoid(self.cif_output2(output2))

        alphas2 = torch.nn.functional.relu(alphas2 * self.smooth_factor2 - self.noise_threshold2)

        # repeat the mask in T demension to match the upsampled length

        if mask is not None:

            mask2 = mask.repeat(1, self.upsample_times, 1).transpose(-1, -2).reshape(alphas2.shape[0], -1)

            mask2 = mask2.unsqueeze(-1)

            alphas2 = alphas2 * mask2

        alphas2 = alphas2.squeeze(-1)

        token_num2 = alphas2.sum(-1)



        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:

            mask = mask.transpose(-1, -2).float()

            alphas = alphas * mask

        if mask_chunk_predictor is not None:

            alphas = alphas * mask_chunk_predictor

        alphas = alphas.squeeze(-1)

        mask = mask.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, mask=mask)



        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, token_num2



    def get_upsample_timestamp(self, hidden, mask=None, token_num=None):

        h = hidden

        b = hidden.shape[0]

        context = h.transpose(1, 2)

        queries = self.pad(context)

        output = torch.relu(self.cif_conv1d(queries))



        # alphas2 is an extra head for timestamp prediction

        if not self.use_cif1_cnn:

            _output = context

        else:

            _output = output

        if self.upsample_type == 'cnn':

            output2 = self.upsample_cnn(_output)

            output2 = output2.transpose(1,2)

        elif self.upsample_type == 'cnn_blstm':

            output2 = self.upsample_cnn(_output)

            output2 = output2.transpose(1,2)

            output2, (_, _) = self.blstm(output2)

        elif self.upsample_type == 'cnn_attn':

            output2 = self.upsample_cnn(_output)

            output2 = output2.transpose(1,2)

            output2, _ = self.self_attn(output2, mask)

        alphas2 = torch.sigmoid(self.cif_output2(output2))

        alphas2 = torch.nn.functional.relu(alphas2 * self.smooth_factor2 - self.noise_threshold2)

        # repeat the mask in T demension to match the upsampled length

        if mask is not None:

            mask2 = mask.repeat(1, self.upsample_times, 1).transpose(-1, -2).reshape(alphas2.shape[0], -1)

            mask2 = mask2.unsqueeze(-1)

            alphas2 = alphas2 * mask2

        alphas2 = alphas2.squeeze(-1)

        _token_num = alphas2.sum(-1)

        if token_num is not None:

            alphas2 *= (token_num / _token_num)[:, None].repeat(1, alphas2.size(1))

        # re-downsample

        ds_alphas = alphas2.reshape(b, -1, self.upsample_times).sum(-1)

        ds_cif_peak = cif_wo_hidden(ds_alphas, self.threshold - 1e-4)

        # upsampled alphas and cif_peak

        us_alphas = alphas2

        us_cif_peak = cif_wo_hidden(us_alphas, self.threshold - 1e-4)

        return ds_alphas, ds_cif_peak, us_alphas, us_cif_peak



    def tail_process_fn(self, hidden, alphas, token_num=None, mask=None):

        b, t, d = hidden.size()

        tail_threshold = self.tail_threshold

        if mask is not None:

            zeros_t = torch.zeros((b, 1), dtype=torch.float32, device=alphas.device)

            ones_t = torch.ones_like(zeros_t)

            mask_1 = torch.cat([mask, zeros_t], dim=1)

            mask_2 = torch.cat([ones_t, mask], dim=1)

            mask = mask_2 - mask_1

            tail_threshold = mask * tail_threshold

            alphas = torch.cat([alphas, zeros_t], dim=1)

            alphas = torch.add(alphas, tail_threshold)

        else:

            tail_threshold = torch.tensor([tail_threshold], dtype=alphas.dtype).to(alphas.device)

            tail_threshold = torch.reshape(tail_threshold, (1, 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 BATPredictor(nn.Module):

    def __init__(self, idim, l_order, r_order, threshold=1.0, dropout=0.1, smooth_factor=1.0, noise_threshold=0, return_accum=False):

        super(BATPredictor, 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

        self.return_accum = return_accum



    def cif(

        self,

        input: Tensor,

        alpha: Tensor,

        beta: float = 1.0,

        return_accum: bool = False,

    ) -> Tuple[Tensor, Tensor, Tensor, Tensor]:

        B, S, C = input.size()

        assert tuple(alpha.size()) == (B, S), f"{alpha.size()} != {(B, S)}"



        dtype = alpha.dtype

        alpha = alpha.float()



        alpha_sum = alpha.sum(1)

        feat_lengths = (alpha_sum / beta).floor().long()

        T = feat_lengths.max()



        # aggregate and integrate

        csum = alpha.cumsum(-1)

        with torch.no_grad():

            # indices used for scattering

            right_idx = (csum / beta).floor().long().clip(max=T)

            left_idx = right_idx.roll(1, dims=1)

            left_idx[:, 0] = 0



            # count # of fires from each source

            fire_num = right_idx - left_idx

            extra_weights = (fire_num - 1).clip(min=0)

            # The extra entry in last dim is for

            output = input.new_zeros((B, T + 1, C))

            source_range = torch.arange(1, 1 + S).unsqueeze(0).type_as(input)

            zero = alpha.new_zeros((1,))



        # right scatter

        fire_mask = fire_num > 0

        right_weight = torch.where(

            fire_mask,

            csum - right_idx.type_as(alpha) * beta,

            zero

        ).type_as(input)

        # assert right_weight.ge(0).all(), f"{right_weight} should be non-negative."

        output.scatter_add_(

            1,

            right_idx.unsqueeze(-1).expand(-1, -1, C),

            right_weight.unsqueeze(-1) * input

        )



        # left scatter

        left_weight = (

            alpha - right_weight - extra_weights.type_as(alpha) * beta

        ).type_as(input)

        output.scatter_add_(

            1,

            left_idx.unsqueeze(-1).expand(-1, -1, C),

            left_weight.unsqueeze(-1) * input

        )



         # extra scatters

        if extra_weights.ge(0).any():

            extra_steps = extra_weights.max().item()

            tgt_idx = left_idx

            src_feats = input * beta

            for _ in range(extra_steps):

                tgt_idx = (tgt_idx + 1).clip(max=T)

                # (B, S, 1)

                src_mask = (extra_weights > 0)

                output.scatter_add_(

                    1,

                    tgt_idx.unsqueeze(-1).expand(-1, -1, C),

                    src_feats * src_mask.unsqueeze(2)

                )

                extra_weights -= 1



        output = output[:, :T, :]



        if return_accum:

            return output, csum

        else:

            return output, alpha



    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)

            # logging.info("target_length: {}".format(target_length))

        else:

            target_length = None

        token_num = alphas.sum(-1)

        if target_length is not None:

            # length_noise = torch.rand(alphas.size(0), device=alphas.device) - 0.5

            # target_length = length_noise + target_length

            alphas *= ((target_length + 1e-4) / token_num)[:, None].repeat(1, alphas.size(1))

        acoustic_embeds, cif_peak = self.cif(hidden, alphas, self.threshold, self.return_accum)

        return acoustic_embeds, token_num, alphas, cif_peak