export onnx (#1455)

zhifu gao

2024-03-11 f2d8ded57f6403696001d39dd07a1396e5a03658

 funasr/models/fsmn_vad_streaming/model.py

@@ -284,6 +284,7 @@
      encoder_class = tables.encoder_classes.get(encoder)
      encoder = encoder_class(**encoder_conf)
      self.encoder = encoder
      self.encoder_conf = encoder_conf
   
   def ResetDetection(self, cache: dict = {}):
      cache["stats"].continous_silence_frame_count = 0
@@ -642,6 +643,45 @@
      
      return results, meta_data
   
   def export(self, **kwargs):
      self.forward = self._export_forward
		
      return self
		
   def _export_forward(self, feats: torch.Tensor, *args, **kwargs):
      scores, out_caches = self.encoder.export_forward(feats, *args)
      return scores, out_caches
	
   def export_dummy_inputs(self, data_in=None, frame=30):
      if data_in is None:
         speech = torch.randn(1, frame, self.encoder_conf.get("input_dim"))
      else:
         speech = None # Undo
		
      cache_frames = self.encoder_conf.get("lorder") + self.encoder_conf.get("rorder") - 1
      in_cache0 = torch.randn(1, self.encoder_conf.get("proj_dim"), cache_frames, 1)
      in_cache1 = torch.randn(1, self.encoder_conf.get("proj_dim"), cache_frames, 1)
      in_cache2 = torch.randn(1, self.encoder_conf.get("proj_dim"), cache_frames, 1)
      in_cache3 = torch.randn(1, self.encoder_conf.get("proj_dim"), cache_frames, 1)
		
      return (speech, in_cache0, in_cache1, in_cache2, in_cache3)
	
   def export_input_names(self):
      return ['speech', 'in_cache0', 'in_cache1', 'in_cache2', 'in_cache3']
	
   def export_output_names(self):
      return ['logits', 'out_cache0', 'out_cache1', 'out_cache2', 'out_cache3']
	
   def export_dynamic_axes(self):
      return {
         'speech': {
            1: 'feats_length'
         },
      }
	
   def export_name(self, ):
      return "model.onnx"
	
   def DetectCommonFrames(self, cache: dict = {}) -> int:
      if cache["stats"].vad_state_machine == VadStateMachine.kVadInStateEndPointDetected:
         return 0