From d79287c37e4e7ae2694a992cbbfb03a5ca4f7670 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期二, 20 二月 2024 14:05:58 +0800
Subject: [PATCH] Merge branch 'main' of github.com:alibaba-damo-academy/FunASR merge
---
funasr/models/seaco_paraformer/model.py | 14 ++++++--------
1 files changed, 6 insertions(+), 8 deletions(-)
diff --git a/funasr/models/seaco_paraformer/model.py b/funasr/models/seaco_paraformer/model.py
index a1ce310..2f55e6e 100644
--- a/funasr/models/seaco_paraformer/model.py
+++ b/funasr/models/seaco_paraformer/model.py
@@ -66,7 +66,6 @@
# bias encoder
if self.bias_encoder_type == 'lstm':
- logging.warning("enable bias encoder sampling and contextual training")
self.bias_encoder = torch.nn.LSTM(self.inner_dim,
self.inner_dim,
2,
@@ -79,7 +78,6 @@
self.lstm_proj = None
self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim)
elif self.bias_encoder_type == 'mean':
- logging.warning("enable bias encoder sampling and contextual training")
self.bias_embed = torch.nn.Embedding(self.vocab_size, self.inner_dim)
else:
logging.error("Unsupport bias encoder type: {}".format(self.bias_encoder_type))
@@ -212,7 +210,7 @@
ys_pad_lens,
hw_list,
nfilter=50,
- seaco_weight=1.0):
+ seaco_weight=1.0):
# decoder forward
decoder_out, decoder_hidden, _ = self.decoder(encoder_out, encoder_out_lens, sematic_embeds, ys_pad_lens, return_hidden=True, return_both=True)
decoder_pred = torch.log_softmax(decoder_out, dim=-1)
@@ -254,10 +252,9 @@
dha_output = self.hotword_output_layer(merged) # remove the last token in loss calculation
dha_pred = torch.log_softmax(dha_output, dim=-1)
- # import pdb; pdb.set_trace()
def _merge_res(dec_output, dha_output):
lmbd = torch.Tensor([seaco_weight] * dha_output.shape[0])
- dha_ids = dha_output.max(-1)[-1][0]
+ dha_ids = dha_output.max(-1)[-1]# [0]
dha_mask = (dha_ids == 8377).int().unsqueeze(-1)
a = (1 - lmbd) / lmbd
b = 1 / lmbd
@@ -267,6 +264,7 @@
logits = dec_output * dha_mask + dha_output[:,:,:] * (1-dha_mask)
return logits
merged_pred = _merge_res(decoder_pred, dha_pred)
+ # import pdb; pdb.set_trace()
return merged_pred
else:
return decoder_pred
@@ -337,7 +335,7 @@
speech = speech.to(device=kwargs["device"])
speech_lengths = speech_lengths.to(device=kwargs["device"])
-
+
# hotword
self.hotword_list = self.generate_hotwords_list(kwargs.get("hotword", None), tokenizer=tokenizer, frontend=frontend)
@@ -415,12 +413,12 @@
token, timestamp)
result_i = {"key": key[i], "text": text_postprocessed,
- "timestamp": time_stamp_postprocessed,
+ "timestamp": time_stamp_postprocessed, "raw_text": copy.copy(text_postprocessed)
}
if ibest_writer is not None:
ibest_writer["token"][key[i]] = " ".join(token)
- # ibest_writer["text"][key[i]] = text
+ # ibest_writer["raw_text"][key[i]] = text
ibest_writer["timestamp"][key[i]] = time_stamp_postprocessed
ibest_writer["text"][key[i]] = text_postprocessed
else:
--
Gitblit v1.9.1