From eb82674d880b1bad0319339b2036644a538e99e8 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 07 四月 2023 14:08:34 +0800
Subject: [PATCH] onnx
---
funasr/export/models/vad_realtime_transformer.py | 8 ++--
funasr/export/models/encoder/sanm_encoder.py | 53 +++++++++++++-------------
2 files changed, 31 insertions(+), 30 deletions(-)
diff --git a/funasr/export/models/encoder/sanm_encoder.py b/funasr/export/models/encoder/sanm_encoder.py
index 44a48ff..f583f56 100644
--- a/funasr/export/models/encoder/sanm_encoder.py
+++ b/funasr/export/models/encoder/sanm_encoder.py
@@ -158,13 +158,14 @@
def forward(self,
speech: torch.Tensor,
speech_lengths: torch.Tensor,
- vad_mask: torch.Tensor,
+ vad_masks: torch.Tensor,
sub_masks: torch.Tensor,
):
speech = speech * self._output_size ** 0.5
mask = self.make_pad_mask(speech_lengths)
+ vad_masks = self.prepare_mask(mask, vad_masks)
mask = self.prepare_mask(mask, sub_masks)
- vad_mask = self.prepare_mask(mask, vad_mask)
+
if self.embed is None:
xs_pad = speech
else:
@@ -176,7 +177,7 @@
# 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_mask
+ mask = vad_masks
encoder_outs = encoder_layer(xs_pad, mask)
xs_pad, masks = encoder_outs[0], encoder_outs[1]
@@ -187,26 +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'
+ # }
+ #
+ # }
diff --git a/funasr/export/models/vad_realtime_transformer.py b/funasr/export/models/vad_realtime_transformer.py
index c8f5364..c34525b 100644
--- a/funasr/export/models/vad_realtime_transformer.py
+++ b/funasr/export/models/vad_realtime_transformer.py
@@ -66,13 +66,13 @@
length = 10
text_indexes = torch.tensor([[266757, 266757, 266757, 266757, 266757, 266757, 266757, 266757, 266757, 266757]], dtype=torch.int32)
text_lengths = torch.tensor([length], dtype=torch.int32)
- vad_mask = vad_mask(10, 3, dtype=torch.float32)[None, None, :, :]
+ vad_masks = vad_mask(10, 3, dtype=torch.float32)[None, None, :, :]
sub_masks = torch.ones(length, length, dtype=torch.float32)
sub_masks = torch.tril(sub_masks).type(torch.float32)
- return (text_indexes, text_lengths, vad_mask, sub_masks[None, None, :, :])
+ return (text_indexes, text_lengths, vad_masks, sub_masks[None, None, :, :])
def get_input_names(self):
- return ['input', 'text_lengths', 'vad_mask', 'sub_masks']
+ return ['input', 'text_lengths', 'vad_masks', 'sub_masks']
def get_output_names(self):
return ['logits']
@@ -82,7 +82,7 @@
'input': {
1: 'feats_length'
},
- 'vad_mask': {
+ 'vad_masks': {
2: 'feats_length1',
3: 'feats_length2'
},
--
Gitblit v1.9.1