From 4ace5a95b052d338947fc88809a440ccd55cf6b4 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 16 十一月 2023 16:39:52 +0800
Subject: [PATCH] funasr pages
---
funasr/export/models/encoder/sanm_encoder.py | 76 +++++++++++++++++++++----------------
1 files changed, 43 insertions(+), 33 deletions(-)
diff --git a/funasr/export/models/encoder/sanm_encoder.py b/funasr/export/models/encoder/sanm_encoder.py
index 3b7b414..d1b4b1e 100644
--- a/funasr/export/models/encoder/sanm_encoder.py
+++ b/funasr/export/models/encoder/sanm_encoder.py
@@ -8,6 +8,8 @@
from funasr.export.models.modules.encoder_layer import EncoderLayerSANM as EncoderLayerSANM_export
from funasr.modules.positionwise_feed_forward import PositionwiseFeedForward
from funasr.export.models.modules.feedforward import PositionwiseFeedForward as PositionwiseFeedForward_export
+from funasr.modules.embedding import StreamSinusoidalPositionEncoder
+
class SANMEncoder(nn.Module):
def __init__(
@@ -20,6 +22,8 @@
):
super().__init__()
self.embed = model.embed
+ if isinstance(self.embed, StreamSinusoidalPositionEncoder):
+ self.embed = None
self.model = model
self.feats_dim = feats_dim
self._output_size = model._output_size
@@ -62,8 +66,10 @@
def forward(self,
speech: torch.Tensor,
speech_lengths: torch.Tensor,
+ online: bool = False
):
- speech = speech * self._output_size ** 0.5
+ if not online:
+ speech = speech * self._output_size ** 0.5
mask = self.make_pad_mask(speech_lengths)
mask = self.prepare_mask(mask)
if self.embed is None:
@@ -148,23 +154,23 @@
self.num_heads = model.encoders[0].self_attn.h
self.hidden_size = model.encoders[0].self_attn.linear_out.out_features
- def prepare_mask(self, mask):
+ def prepare_mask(self, mask, sub_masks):
mask_3d_btd = mask[:, :, None]
- if len(mask.shape) == 2:
- mask_4d_bhlt = 1 - mask[:, None, None, :]
- elif len(mask.shape) == 3:
- mask_4d_bhlt = 1 - mask[:, None, :]
- mask_4d_bhlt = mask_4d_bhlt * -10000.0
+ mask_4d_bhlt = (1 - sub_masks) * -10000.0
return mask_3d_btd, mask_4d_bhlt
def forward(self,
speech: torch.Tensor,
speech_lengths: torch.Tensor,
+ vad_masks: torch.Tensor,
+ sub_masks: torch.Tensor,
):
speech = speech * self._output_size ** 0.5
mask = self.make_pad_mask(speech_lengths)
- mask = self.prepare_mask(mask)
+ vad_masks = self.prepare_mask(mask, vad_masks)
+ mask = self.prepare_mask(mask, sub_masks)
+
if self.embed is None:
xs_pad = speech
else:
@@ -173,8 +179,12 @@
encoder_outs = self.model.encoders0(xs_pad, mask)
xs_pad, masks = encoder_outs[0], encoder_outs[1]
- encoder_outs = self.model.encoders(xs_pad, mask)
- xs_pad, masks = encoder_outs[0], encoder_outs[1]
+ # encoder_outs = self.model.encoders(xs_pad, mask)
+ for layer_idx, encoder_layer in enumerate(self.model.encoders):
+ if layer_idx == len(self.model.encoders) - 1:
+ mask = vad_masks
+ encoder_outs = encoder_layer(xs_pad, mask)
+ xs_pad, masks = encoder_outs[0], encoder_outs[1]
xs_pad = self.model.after_norm(xs_pad)
@@ -183,26 +193,26 @@
def get_output_size(self):
return self.model.encoders[0].size
- def get_dummy_inputs(self):
- feats = torch.randn(1, 100, self.feats_dim)
- return (feats)
-
- def get_input_names(self):
- return ['feats']
-
- def get_output_names(self):
- return ['encoder_out', 'encoder_out_lens', 'predictor_weight']
-
- def get_dynamic_axes(self):
- return {
- 'feats': {
- 1: 'feats_length'
- },
- 'encoder_out': {
- 1: 'enc_out_length'
- },
- 'predictor_weight': {
- 1: 'pre_out_length'
- }
-
- }
+ # def get_dummy_inputs(self):
+ # feats = torch.randn(1, 100, self.feats_dim)
+ # return (feats)
+ #
+ # def get_input_names(self):
+ # return ['feats']
+ #
+ # def get_output_names(self):
+ # return ['encoder_out', 'encoder_out_lens', 'predictor_weight']
+ #
+ # def get_dynamic_axes(self):
+ # return {
+ # 'feats': {
+ # 1: 'feats_length'
+ # },
+ # 'encoder_out': {
+ # 1: 'enc_out_length'
+ # },
+ # 'predictor_weight': {
+ # 1: 'pre_out_length'
+ # }
+ #
+ # }
--
Gitblit v1.9.1