Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

commit

语帆920331972a13
diff --cc examples/industrial_data_pretraining/contextual_paraformer/demo.sh
index 8fc66f34f,1bd4f7f5b..1bd4f7f5b
mode 100644,100755..100755
--- a/examples/industrial_data_pretraining/contextual_paraformer/demo.sh
+++ b/examples/industrial_data_pretraining/contextual_paraformer/demo.sh
diff --cc funasr/auto/auto_model.py
index 921ede809,9bb9ce07d..ec3c3f370
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@@ -162,16 -158,13 +163,15 @@@ class AutoModel
tokenizer_class = tables.tokenizer_classes.get(tokenizer)
tokenizer = tokenizer_class(**kwargs["tokenizer_conf"])
kwargs["tokenizer"] = tokenizer
- kwargs["token_list"] = tokenizer.token_list
- vocab_size = len(tokenizer.token_list)
+
+ kwargs["token_list"] = tokenizer.token_list if hasattr(tokenizer, "token_list") else None
+ kwargs["token_list"] = tokenizer.get_vocab() if hasattr(tokenizer, "get_vocab") else kwargs["token_list"]
+ vocab_size = len(kwargs["token_list"]) if kwargs["token_list"] is not None else -1
else:
vocab_size = -1
-
# build frontend
frontend = kwargs.get("frontend", None)
-
+ kwargs["input_size"] = None
if frontend is not None:
frontend_class = tables.frontend_classes.get(frontend)
frontend = frontend_class(**kwargs["frontend_conf"])
diff --cc funasr/models/contextual_paraformer/model.py
index 49868a8f4,1c0805ab0..7d6f729a8
--- a/funasr/models/contextual_paraformer/model.py
+++ b/funasr/models/contextual_paraformer/model.py
@@@ -194,8 -194,10 +193,8 @@@ class ContextualParaformer(Paraformer)
sematic_embeds, decoder_out_1st = self.sampler(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens,
pre_acoustic_embeds, contextual_info)
else:
- if self.step_cur < 2:
- logging.info("disable sampler in paraformer, sampling_ratio: {}".format(self.sampling_ratio))
sematic_embeds = pre_acoustic_embeds
-
+
# 1. Forward decoder
decoder_outs = self.decoder(
encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, contextual_info=contextual_info
diff --cc funasr/models/seaco_paraformer/model.py
index 20b0cc838,e0467b3c4..a8b1f1fb1
--- a/funasr/models/seaco_paraformer/model.py
+++ b/funasr/models/seaco_paraformer/model.py
@@@ -128,8 -130,9 +128,8 @@@ class SeacoParaformer(BiCifParaformer,
hotword_pad = kwargs.get("hotword_pad")
hotword_lengths = kwargs.get("hotword_lengths")
dha_pad = kwargs.get("dha_pad")
-
+
batch_size = speech.shape[0]
- self.step_cur += 1
# for data-parallel
text = text[:, : text_lengths.max()]
speech = speech[:, :speech_lengths.max()]
@@@ -219,11 -225,17 +222,11 @@@
contextual_info = selected.squeeze(0).repeat(encoder_out.shape[0], 1, 1).to(encoder_out.device)
num_hot_word = contextual_info.shape[1]
_contextual_length = torch.Tensor([num_hot_word]).int().repeat(encoder_out.shape[0]).to(encoder_out.device)
-
+
# ASF Core
if nfilter > 0 and nfilter < num_hot_word:
- for dec in self.seaco_decoder.decoders:
- dec.reserve_attn = True
-
- # cif_attended, _ = self.decoder2(contextual_info, _contextual_length, sematic_embeds, ys_pad_lens)
- dec_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
- # cif_filter = torch.topk(self.decoder2.decoders[-1].attn_mat[0][0].sum(0).sum(0)[:-1], min(nfilter, num_hot_word-1))[1].tolist()
-
- hotword_scores = self.seaco_decoder.decoders[-1].attn_mat[0][0].sum(0).sum(0)[:-1]
+ hotword_scores = self.seaco_decoder.forward_asf6(contextual_info, _contextual_length, decoder_hidden, ys_pad_lens)
+ hotword_scores = hotword_scores[0].sum(0).sum(0)
# hotword_scores /= torch.sqrt(torch.tensor(hw_lengths)[:-1].float()).to(hotword_scores.device)
dec_filter = torch.topk(hotword_scores, min(nfilter, num_hot_word-1))[1].tolist()
add_filter = dec_filter
diff --cc funasr/train_utils/load_pretrained_model.py
index ea2372537,9127e2fe1..0c46449e4
--- a/funasr/train_utils/load_pretrained_model.py
+++ b/funasr/train_utils/load_pretrained_model.py
@@@ -61,8 -96,8 +61,9 @@@ def load_pretrained_model
obj = model
dst_state = obj.state_dict()
+
print(f"ckpt: {path}")
+
if oss_bucket is None:
src_state = torch.load(path, map_location=map_location)
else: