From dee1354d0d984df21d16a2eba5bacec31bfb0b4b Mon Sep 17 00:00:00 2001
From: 维石 <shixian.shi@alibaba-inc.com>
Date: 星期五, 19 四月 2024 14:57:31 +0800
Subject: [PATCH] empty result bug fix
---
funasr/models/bicif_paraformer/cif_predictor.py | 201 +++++++++++++++++++++++++++++++++++++++++++++-----
1 files changed, 181 insertions(+), 20 deletions(-)
diff --git a/funasr/models/bicif_paraformer/cif_predictor.py b/funasr/models/bicif_paraformer/cif_predictor.py
index 5a1488e..2cdbc16 100644
--- a/funasr/models/bicif_paraformer/cif_predictor.py
+++ b/funasr/models/bicif_paraformer/cif_predictor.py
@@ -1,17 +1,15 @@
+#!/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.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
from funasr.register import tables
+from funasr.models.transformer.utils.nets_utils import make_pad_mask
-class mae_loss(nn.Module):
+class mae_loss(torch.nn.Module):
def __init__(self, normalize_length=False):
super(mae_loss, self).__init__()
@@ -95,7 +93,7 @@
return fires
@tables.register("predictor_classes", "CifPredictorV3")
-class CifPredictorV3(nn.Module):
+class CifPredictorV3(torch.nn.Module):
def __init__(self,
idim,
l_order,
@@ -116,9 +114,9 @@
):
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.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
@@ -131,14 +129,14 @@
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)
+ self.upsample_cnn = torch.nn.ConvTranspose1d(idim, idim, self.upsample_times, self.upsample_times)
+ self.cif_output2 = torch.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)
+ self.upsample_cnn = torch.nn.ConvTranspose1d(idim, idim, self.upsample_times, self.upsample_times)
+ self.blstm = torch.nn.LSTM(idim, idim, 1, bias=True, batch_first=True, dropout=0.0, bidirectional=True)
+ self.cif_output2 = torch.nn.Linear(idim*2, 1)
elif self.upsample_type == 'cnn_attn':
- self.upsample_cnn = nn.ConvTranspose1d(idim, idim, self.upsample_times, self.upsample_times)
+ self.upsample_cnn = torch.nn.ConvTranspose1d(idim, idim, self.upsample_times, self.upsample_times)
from funasr.models.transformer.encoder import EncoderLayer as TransformerEncoderLayer
from funasr.models.transformer.attention import MultiHeadedAttention
from funasr.models.transformer.positionwise_feed_forward import PositionwiseFeedForward
@@ -157,7 +155,7 @@
True, #normalize_before,
False, #concat_after,
)
- self.cif_output2 = nn.Linear(idim, 1)
+ self.cif_output2 = torch.nn.Linear(idim, 1)
self.smooth_factor2 = smooth_factor2
self.noise_threshold2 = noise_threshold2
@@ -338,3 +336,166 @@
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()
+
+@tables.register("predictor_classes", "CifPredictorV3Export")
+class CifPredictorV3Export(torch.nn.Module):
+ def __init__(self, model, **kwargs):
+ 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_export(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_export(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_export(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], 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], 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], dtype=alphas.dtype, 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)
+
+ 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):
+ 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
+
+
+@torch.jit.script
+def cif_wo_hidden_export(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) * threshold,
+ integrate)
+
+ fires = torch.stack(list_fires, 1)
+ return fires
\ No newline at end of file
--
Gitblit v1.9.1