From e8f80e96f99cb856423d030c7d055c302a6d3278 Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期三, 17 四月 2024 21:23:20 +0800
Subject: [PATCH] Dev gzf exp (#1627)
---
funasr/auto/auto_model.py | 12 +++++++++++-
1 files changed, 11 insertions(+), 1 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index d8ac5ca..d173a53 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -21,6 +21,7 @@
from funasr.utils.timestamp_tools import timestamp_sentence
from funasr.download.download_from_hub import download_model
from funasr.utils.vad_utils import slice_padding_audio_samples
+from funasr.utils.vad_utils import merge_vad
from funasr.utils.load_utils import load_audio_text_image_video
from funasr.train_utils.set_all_random_seed import set_all_random_seed
from funasr.train_utils.load_pretrained_model import load_pretrained_model
@@ -174,6 +175,8 @@
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
+ if vocab_size == -1 and hasattr(tokenizer, "get_vocab_size"):
+ vocab_size = tokenizer.get_vocab_size()
else:
vocab_size = -1
kwargs["tokenizer"] = tokenizer
@@ -202,7 +205,7 @@
load_pretrained_model(
model=model,
path=init_param,
- ignore_init_mismatch=kwargs.get("ignore_init_mismatch", False),
+ ignore_init_mismatch=kwargs.get("ignore_init_mismatch", True),
oss_bucket=kwargs.get("oss_bucket", None),
scope_map=kwargs.get("scope_map", []),
excludes=kwargs.get("excludes", None),
@@ -210,6 +213,9 @@
else:
print(f"error, init_param does not exist!: {init_param}")
+ # fp16
+ if kwargs.get("fp16", False):
+ model.to(torch.float16)
return model, kwargs
def __call__(self, *args, **cfg):
@@ -295,6 +301,10 @@
res = self.inference(input, input_len=input_len, model=self.vad_model, kwargs=self.vad_kwargs, **cfg)
end_vad = time.time()
+ # FIX(gcf): concat the vad clips for sense vocie model for better aed
+ if kwargs.get("merge_vad", False):
+ for i in range(len(res)):
+ res[i]['value'] = merge_vad(res[i]['value'], kwargs.get("merge_length", 15000))
# step.2 compute asr model
model = self.model
--
Gitblit v1.9.1