update

嘉渊

2023-04-24 6427c834dfd97b1f05c6659cdc7ccf010bf82fe1

 funasr/export/models/predictor/cif.py

@@ -1,9 +1,8 @@
#!/usr/bin/env python3

# -*- coding: utf-8 -*-



import torch

from torch import nn

import logging

import numpy as np





def sequence_mask(lengths, maxlen=None, dtype=torch.float32, device=None):

@@ -16,6 +15,11 @@
   

   return mask.type(dtype).to(device) if device is not None else mask.type(dtype)



def sequence_mask_scripts(lengths, maxlen:int):

   row_vector = torch.arange(0, maxlen, 1).type(lengths.dtype).to(lengths.device)

   matrix = torch.unsqueeze(lengths, dim=-1)

   mask = row_vector < matrix

   return mask.type(torch.float32).to(lengths.device)



class CifPredictorV2(nn.Module):

   def __init__(self, model):

@@ -43,11 +47,11 @@
      alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)

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

      alphas = alphas * mask

		
      alphas = alphas.squeeze(-1)

		
      token_num = alphas.sum(-1)

      

      mask = mask.squeeze(-1)

      hidden, alphas, token_num = self.tail_process_fn(hidden, alphas, mask=mask)

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

      

      return acoustic_embeds, token_num, alphas, cif_peak

@@ -58,12 +62,14 @@
      

      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, tail_threshold], dim=1)

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

      alphas = torch.add(alphas, tail_threshold)



      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)

@@ -71,28 +77,76 @@
      

      return hidden, alphas, token_num_floor





# @torch.jit.script

# def cif(hidden, alphas, threshold: float):

#    batch_size, len_time, hidden_size = hidden.size()

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

#

#    # 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.floor(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([int(max_label_len - l.size(0)), int(hidden_size)], device=hidden.device)

#       list_ls.append(torch.cat([l, pad_l], 0))

#    return torch.stack(list_ls, 0), fires





@torch.jit.script

def cif(hidden, alphas, threshold: float):

   batch_size, len_time, hidden_size = hidden.size()

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

   

   # loop varss

   integrate = torch.zeros([batch_size], device=hidden.device)

   frame = torch.zeros([batch_size, hidden_size], device=hidden.device)

   integrate = torch.zeros([batch_size], dtype=alphas.dtype, device=hidden.device)

   frame = torch.zeros([batch_size, hidden_size], dtype=hidden.dtype, 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

      distribution_completion = torch.ones([batch_size], dtype=alphas.dtype, 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 - torch.ones([batch_size], dtype=alphas.dtype, device=hidden.device),

                              integrate)

      cur = torch.where(fire_place,

                        distribution_completion,

@@ -107,13 +161,128 @@
   

   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().item()

   print("type: {}".format(type(max_label_len)))



   fire_idxs = fires >= threshold

   frame_fires = torch.zeros_like(hidden)

   max_label_len = frames[0, fire_idxs[0]].size(0)

   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([int(max_label_len - l.size(0)), int(hidden_size)], device=hidden.device)

      list_ls.append(torch.cat([l, pad_l], 0))

   return torch.stack(list_ls, 0), fires

      frame_fire = frames[b, fire_idxs[b]]

      frame_len = frame_fire.size(0)

      frame_fires[b, :frame_len, :] = frame_fire

	
      if frame_len >= max_label_len:

         max_label_len = frame_len

   frame_fires = frame_fires[:, :max_label_len, :]

   return frame_fires, fires





class CifPredictorV3(nn.Module):

   def __init__(self, model):

      super().__init__()

		
      self.pad = model.pad

      self.cif_conv1d = model.cif_conv1d

      self.cif_output = model.cif_output

      self.threshold = model.threshold

      self.smooth_factor = model.smooth_factor

      self.noise_threshold = model.noise_threshold

      self.tail_threshold = model.tail_threshold



      self.upsample_times = model.upsample_times

      self.upsample_cnn = model.upsample_cnn

      self.blstm = model.blstm

      self.cif_output2 = model.cif_output2

      self.smooth_factor2 = model.smooth_factor2

      self.noise_threshold2 = model.noise_threshold2

	
   def forward(self, hidden: torch.Tensor,

               mask: torch.Tensor,

               ):

      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)

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

      alphas = alphas * mask

      alphas = alphas.squeeze(-1)

      token_num = alphas.sum(-1)

		
      mask = mask.squeeze(-1)

      hidden, alphas, token_num = self.tail_process_fn(hidden, alphas, mask=mask)

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

		
      return acoustic_embeds, token_num, alphas, cif_peak

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

      h = hidden

      b = hidden.shape[0]

      context = h.transpose(1, 2)



      # generate alphas2

      _output = context

      output2 = self.upsample_cnn(_output)

      output2 = output2.transpose(1, 2)

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

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

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

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

      mask = mask.unsqueeze(-1)

      alphas2 = alphas2 * mask

      alphas2 = alphas2.squeeze(-1)

      _token_num = alphas2.sum(-1)

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

      # upsampled alphas and cif_peak

      us_alphas = alphas2

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

      return 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

		
      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)



      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





@torch.jit.script

def cif_wo_hidden(alphas, threshold: float):

    batch_size, len_time = alphas.size()



    # loop varss

    integrate = torch.zeros([batch_size], dtype=alphas.dtype, device=alphas.device)

    # intermediate vars along time

    list_fires = []



    for t in range(len_time):

        alpha = alphas[:, t]



        integrate += alpha

        list_fires.append(integrate)



        fire_place = integrate >= threshold

        integrate = torch.where(fire_place,

                                integrate - torch.ones([batch_size], device=alphas.device),

                                integrate)



    fires = torch.stack(list_fires, 1)

    return fires