From 33d3d2084403fd34b79c835d2f2fe04f6cd8f738 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期三, 13 九月 2023 09:33:54 +0800
Subject: [PATCH] Merge branch 'main' of github.com:alibaba-damo-academy/FunASR add
---
funasr/export/models/predictor/cif.py | 209 +++++++++++++++++++++++++++++++++------------------
1 files changed, 135 insertions(+), 74 deletions(-)
diff --git a/funasr/export/models/predictor/cif.py b/funasr/export/models/predictor/cif.py
index 034e233..03c4433 100644
--- a/funasr/export/models/predictor/cif.py
+++ b/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):
@@ -37,6 +36,17 @@
def forward(self, hidden: torch.Tensor,
mask: torch.Tensor,
):
+ alphas, token_num = self.forward_cnn(hidden, mask)
+ mask = mask.transpose(-1, -2).float()
+ 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 forward_cnn(self, hidden: torch.Tensor,
+ mask: torch.Tensor,
+ ):
h = hidden
context = h.transpose(1, 2)
queries = self.pad(context)
@@ -48,14 +58,10 @@
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)
-
- acoustic_embeds, cif_peak = cif(hidden, alphas, self.threshold)
-
- return acoustic_embeds, token_num, alphas, cif_peak
+
+ return alphas, token_num
def tail_process_fn(self, hidden, alphas, token_num=None, mask=None):
b, t, d = hidden.size()
@@ -63,18 +69,21 @@
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)
token_num_floor = torch.floor(token_num)
return hidden, alphas, token_num_floor
+
# @torch.jit.script
# def cif(hidden, alphas, threshold: float):
@@ -113,70 +122,14 @@
# 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)))
+# 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)], dtype=l.dtype, device=hidden.device)
+# 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], 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)
-# len_labels = torch.floor(torch.sum(alphas, dim=1)).int()
-# max_label_len = torch.max(len_labels)
-# pad_num = max_label_len - len_labels
-# pad_num_max = torch.max(pad_num).item()
-# frames_pad_tensor = torch.zeros([int(batch_size), int(pad_num_max), int(hidden_size)], dtype=frames.dtype,
-# device=frames.device)
-# fires_pad_tensor = torch.ones([int(batch_size), int(pad_num_max)], dtype=fires.dtype, device=fires.device)
-# fires_pad_tensor_mask = sequence_mask_scripts(pad_num, maxlen=int(pad_num_max))
-# fires_pad_tensor *= fires_pad_tensor_mask
-# frames_pad = torch.cat([frames, frames_pad_tensor], dim=1)
-# fires_pad = torch.cat([fires, fires_pad_tensor], dim=1)
-# index_bool = fires_pad >= threshold
-# frames_fire = frames_pad[index_bool]
-# frames_fire = torch.reshape(frames_fire, (int(batch_size), -1, int(hidden_size)))
-# frames_fire_mask = sequence_mask_scripts(len_labels, maxlen=int(max_label_len))
-# frames_fire *= frames_fire_mask[:, :, None]
-#
-# return frames_fire, fires
@torch.jit.script
@@ -215,15 +168,11 @@
fires = torch.stack(list_fires, 1)
frames = torch.stack(list_frames, 1)
- # list_ls = []
- len_labels = torch.round(alphas.sum(-1)).type(torch.int32)
- # max_label_len = int(torch.max(len_labels).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, :]
frame_fire = frames[b, fire_idxs[b]]
frame_len = frame_fire.size(0)
frame_fires[b, :frame_len, :] = frame_fire
@@ -232,3 +181,115 @@
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)*threshold,
+ integrate)
+
+ fires = torch.stack(list_fires, 1)
+ return fires
--
Gitblit v1.9.1