export

funasr/export/models/encoder/sanm_encoder.py

@@ -163,9 +163,9 @@ class SANMVadEncoder(nn.Module):
         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]
-        sub_masks = subsequent_mask(mask.size(-1)).type(torch.float32)
+        # sub_masks = subsequent_mask(mask.size(-1)).type(torch.float32)
         if len(mask.shape) == 2:
             mask_4d_bhlt = 1 - sub_masks[:, None, None, :]
         elif len(mask.shape) == 3:
@@ -178,6 +178,7 @@ class SANMVadEncoder(nn.Module):
                 speech: torch.Tensor,
                 speech_lengths: torch.Tensor,
                 vad_mask: torch.Tensor,
+                sub_masks: torch.Tensor,
                 ):
         speech = speech * self._output_size ** 0.5
         mask = self.make_pad_mask(speech_lengths)

funasr/export/models/vad_realtime_transformer.py

@@ -11,7 +11,7 @@ from funasr.punctuation.abs_model import AbsPunctuation
 from funasr.punctuation.sanm_encoder import SANMVadEncoder
 from funasr.export.models.encoder.sanm_encoder import SANMVadEncoder as SANMVadEncoder_export
 
-class VadRealtimeTransformer(AbsPunctuation):
+class VadRealtimeTransformer(nn.Module):
 
     def __init__(
         self,
@@ -36,8 +36,11 @@ class VadRealtimeTransformer(AbsPunctuation):
 
 
 
-    def forward(self, input: torch.Tensor, text_lengths: torch.Tensor,
-                vad_indexes: torch.Tensor) -> Tuple[torch.Tensor, None]:
+    def forward(self, input: torch.Tensor,
+                text_lengths: torch.Tensor,
+                vad_indexes: torch.Tensor,
+                sub_masks: torch.Tensor,
+                ) -> Tuple[torch.Tensor, None]:
         """Compute loss value from buffer sequences.
 
         Args:
@@ -47,7 +50,7 @@ class VadRealtimeTransformer(AbsPunctuation):
         """
         x = self.embed(input)
         # mask = self._target_mask(input)
-        h, _ = self.encoder(x, text_lengths, vad_indexes)
+        h, _ = self.encoder(x, text_lengths, vad_indexes, sub_masks)
         y = self.decoder(h)
         return y
 
@@ -59,7 +62,9 @@ class VadRealtimeTransformer(AbsPunctuation):
         text_indexes = torch.randint(0, self.embed.num_embeddings, (1, length))
         text_lengths = torch.tensor([length], dtype=torch.int32)
         vad_mask = torch.ones(length, length, dtype=torch.float32)[None, None, :, :]
-        return (text_indexes, text_lengths, vad_mask)
+        sub_masks = torch.ones(length, length, dtype=torch.float32)
+        sub_masks = torch.tril(sub_masks)
+        return (text_indexes, text_lengths, vad_mask, sub_masks)
 
     def get_input_names(self):
         return ['input', 'text_lengths', 'vad_mask']

funasr/export/test/test_onnx_punc_vadrealtime.py

@@ -9,7 +9,11 @@ if __name__ == '__main__':
     output_name = [nd.name for nd in sess.get_outputs()]
 
     def _get_feed_dict(text_length):
-        return {'input': np.ones((1, text_length), dtype=np.int64), 'text_lengths': np.array([text_length,], dtype=np.int32), 'vad_mask': np.ones((1, 1, text_length, text_length), dtype=np.float32)}
+        return {'input': np.ones((1, text_length), dtype=np.int64),
+                'text_lengths': np.array([text_length,], dtype=np.int32),
+                'vad_mask': np.ones((1, 1, text_length, text_length), dtype=np.float32),
+                'sub_masks': np.tril(np.ones((text_length, text_length), dtype=np.float32))
+                }
 
     def _run(feed_dict):
         output = sess.run(output_name, input_feed=feed_dict)