update funasr-onnx

shixian.shi

2024-01-10 e30a17cf4e715b3d139fa1e0ba01cda1bcf0f884

 funasr/models/paraformer/model.py

@@ -22,13 +22,13 @@

from torch.cuda.amp import autocast

from funasr.datasets.audio_datasets.load_audio_extract_fbank import load_audio, extract_fbank
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
from funasr.utils import postprocess_utils
from funasr.utils.datadir_writer import DatadirWriter
from funasr.utils.register import register_class, registry_tables
from funasr.register import tables
from funasr.models.ctc.ctc import CTC

@register_class("model_classes", "Paraformer")
@tables.register("model_classes", "Paraformer")
class Paraformer(nn.Module):
   """
   Author: Speech Lab of DAMO Academy, Alibaba Group
@@ -79,17 +79,17 @@
      super().__init__()

      if specaug is not None:
         specaug_class = registry_tables.specaug_classes.get(specaug.lower())
         specaug_class = tables.specaug_classes.get(specaug.lower())
         specaug = specaug_class(**specaug_conf)
      if normalize is not None:
         normalize_class = registry_tables.normalize_classes.get(normalize.lower())
         normalize_class = tables.normalize_classes.get(normalize.lower())
         normalize = normalize_class(**normalize_conf)
      encoder_class = registry_tables.encoder_classes.get(encoder.lower())
      encoder_class = tables.encoder_classes.get(encoder.lower())
      encoder = encoder_class(input_size=input_size, **encoder_conf)
      encoder_output_size = encoder.output_size()

      if decoder is not None:
         decoder_class = registry_tables.decoder_classes.get(decoder.lower())
         decoder_class = tables.decoder_classes.get(decoder.lower())
         decoder = decoder_class(
            vocab_size=vocab_size,
            encoder_output_size=encoder_output_size,
@@ -104,7 +104,7 @@
            odim=vocab_size, encoder_output_size=encoder_output_size, **ctc_conf
         )
      if predictor is not None:
         predictor_class = registry_tables.predictor_classes.get(predictor.lower())
         predictor_class = tables.predictor_classes.get(predictor.lower())
         predictor = predictor_class(**predictor_conf)
      
      # note that eos is the same as sos (equivalent ID)
@@ -466,7 +466,7 @@
      else:
         # extract fbank feats
         time1 = time.perf_counter()
         audio_sample_list = load_audio(data_in, fs=frontend.fs, audio_fs=kwargs.get("fs", 16000))
         audio_sample_list = load_audio_text_image_video(data_in, fs=frontend.fs, audio_fs=kwargs.get("fs", 16000), data_type=kwargs.get("data_type", "sound"), tokenizer=tokenizer)
         time2 = time.perf_counter()
         meta_data["load_data"] = f"{time2 - time1:0.3f}"
         speech, speech_lengths = extract_fbank(audio_sample_list, data_type=kwargs.get("data_type", "sound"), frontend=frontend)
@@ -495,6 +495,8 @@

      results = []
      b, n, d = decoder_out.size()
      if isinstance(key[0], (list, tuple)):
         key = key[0]
      for i in range(b):
         x = encoder_out[i, :encoder_out_lens[i], :]
         am_scores = decoder_out[i, :pre_token_length[i], :]
@@ -535,13 +537,14 @@
               text = tokenizer.tokens2text(token)
               
               text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
               result_i = {"key": key[i], "token": token, "text": text, "text_postprocessed": text_postprocessed}
					
               result_i = {"key": key[i], "text": text_postprocessed}

               
               if ibest_writer is not None:
                  ibest_writer["token"][key[i]] = " ".join(token)
                  ibest_writer["text"][key[i]] = text
                  ibest_writer["text_postprocessed"][key[i]] = text_postprocessed
                  # ibest_writer["text"][key[i]] = text
                  ibest_writer["text"][key[i]] = text_postprocessed
            else:
               result_i = {"key": key[i], "token_int": token_int}
            results.append(result_i)