From 6427c834dfd97b1f05c6659cdc7ccf010bf82fe1 Mon Sep 17 00:00:00 2001
From: 嘉渊 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期一, 24 四月 2023 19:50:07 +0800
Subject: [PATCH] update
---
funasr/export/models/encoder/sanm_encoder.py | 69 ++++++++++++++++++----------------
1 files changed, 37 insertions(+), 32 deletions(-)
diff --git a/funasr/export/models/encoder/sanm_encoder.py b/funasr/export/models/encoder/sanm_encoder.py
index 3b7b414..f583f56 100644
--- a/funasr/export/models/encoder/sanm_encoder.py
+++ b/funasr/export/models/encoder/sanm_encoder.py
@@ -9,6 +9,7 @@
from funasr.modules.positionwise_feed_forward import PositionwiseFeedForward
from funasr.export.models.modules.feedforward import PositionwiseFeedForward as PositionwiseFeedForward_export
+
class SANMEncoder(nn.Module):
def __init__(
self,
@@ -148,23 +149,23 @@
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 +174,12 @@
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 +188,26 @@
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