From 4137f5cf26e7c4b40853959cd2574edfde03aa60 Mon Sep 17 00:00:00 2001
From: 志浩 <neo.dzh@alibaba-inc.com>
Date: 星期五, 07 四月 2023 21:03:34 +0800
Subject: [PATCH] Merge branch 'main' of github.com:alibaba-damo-academy/FunASR into dev_dzh
---
funasr/export/models/encoder/sanm_encoder.py | 104 ++++++++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 104 insertions(+), 0 deletions(-)
diff --git a/funasr/export/models/encoder/sanm_encoder.py b/funasr/export/models/encoder/sanm_encoder.py
index 8a50538..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,
@@ -107,3 +108,106 @@
}
}
+
+
+class SANMVadEncoder(nn.Module):
+ def __init__(
+ self,
+ model,
+ max_seq_len=512,
+ feats_dim=560,
+ model_name='encoder',
+ onnx: bool = True,
+ ):
+ super().__init__()
+ self.embed = model.embed
+ self.model = model
+ self.feats_dim = feats_dim
+ self._output_size = model._output_size
+
+ if onnx:
+ self.make_pad_mask = MakePadMask(max_seq_len, flip=False)
+ else:
+ self.make_pad_mask = sequence_mask(max_seq_len, flip=False)
+
+ if hasattr(model, 'encoders0'):
+ for i, d in enumerate(self.model.encoders0):
+ if isinstance(d.self_attn, MultiHeadedAttentionSANM):
+ d.self_attn = MultiHeadedAttentionSANM_export(d.self_attn)
+ if isinstance(d.feed_forward, PositionwiseFeedForward):
+ d.feed_forward = PositionwiseFeedForward_export(d.feed_forward)
+ self.model.encoders0[i] = EncoderLayerSANM_export(d)
+
+ for i, d in enumerate(self.model.encoders):
+ if isinstance(d.self_attn, MultiHeadedAttentionSANM):
+ d.self_attn = MultiHeadedAttentionSANM_export(d.self_attn)
+ if isinstance(d.feed_forward, PositionwiseFeedForward):
+ d.feed_forward = PositionwiseFeedForward_export(d.feed_forward)
+ self.model.encoders[i] = EncoderLayerSANM_export(d)
+
+ self.model_name = model_name
+ 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, sub_masks):
+ mask_3d_btd = mask[:, :, None]
+ 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)
+ vad_masks = self.prepare_mask(mask, vad_masks)
+ mask = self.prepare_mask(mask, sub_masks)
+
+ if self.embed is None:
+ xs_pad = speech
+ else:
+ xs_pad = self.embed(speech)
+
+ 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)
+ 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)
+
+ return xs_pad, speech_lengths
+
+ 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'
+ # }
+ #
+ # }
--
Gitblit v1.9.1