From 08114ae27d85949106aeab03b3fa5d764d100b33 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 14 六月 2024 15:16:40 +0800
Subject: [PATCH] decoding
---
funasr/auto/auto_model.py | 79 +++++++++++++++++++++++++++------------
1 files changed, 54 insertions(+), 25 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 577c328..e30eb09 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -19,6 +19,7 @@
from funasr.utils.load_utils import load_bytes
from funasr.download.file import download_from_url
from funasr.utils.timestamp_tools import timestamp_sentence
+from funasr.utils.timestamp_tools import timestamp_sentence_en
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
@@ -42,8 +43,9 @@
filelist = [".scp", ".txt", ".json", ".jsonl", ".text"]
chars = string.ascii_letters + string.digits
- if isinstance(data_in, str) and data_in.startswith("http"): # url
- data_in = download_from_url(data_in)
+ if isinstance(data_in, str):
+ if data_in.startswith("http://") or data_in.startswith("https://"): # url
+ data_in = download_from_url(data_in)
if isinstance(data_in, str) and os.path.exists(
data_in
@@ -211,7 +213,6 @@
deep_update(model_conf, kwargs.get("model_conf", {}))
deep_update(model_conf, kwargs)
model = model_class(**model_conf, vocab_size=vocab_size)
- model.to(device)
# init_param
init_param = kwargs.get("init_param", None)
@@ -232,6 +233,9 @@
# fp16
if kwargs.get("fp16", False):
model.to(torch.float16)
+ elif kwargs.get("bf16", False):
+ model.to(torch.bfloat16)
+ model.to(device)
return model, kwargs
def __call__(self, *args, **cfg):
@@ -284,7 +288,7 @@
with torch.no_grad():
res = model.inference(**batch, **kwargs)
if isinstance(res, (list, tuple)):
- results = res[0]
+ results = res[0] if len(res) > 0 else [{"text": ""}]
meta_data = res[1] if len(res) > 1 else {}
time2 = time.perf_counter()
@@ -320,7 +324,7 @@
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)):
@@ -358,13 +362,13 @@
results_sorted = []
if not len(sorted_data):
+ results_ret_list.append({"key": key, "text": "", "timestamp": []})
logging.info("decoding, utt: {}, empty speech".format(key))
continue
if len(sorted_data) > 0 and len(sorted_data[0]) > 0:
batch_size = max(batch_size, sorted_data[0][0][1] - sorted_data[0][0][0])
- batch_size_ms_cum = 0
beg_idx = 0
beg_asr_total = time.time()
time_speech_total_per_sample = speech_lengths / 16000
@@ -373,19 +377,22 @@
# pbar_sample = tqdm(colour="blue", total=n, dynamic_ncols=True)
all_segments = []
+ max_len_in_batch = 0
+ end_idx = 1
for j, _ in enumerate(range(0, n)):
# pbar_sample.update(1)
- batch_size_ms_cum += sorted_data[j][0][1] - sorted_data[j][0][0]
+ sample_length = sorted_data[j][0][1] - sorted_data[j][0][0]
+ potential_batch_length = max(max_len_in_batch, sample_length) * (j + 1 - beg_idx)
+ # batch_size_ms_cum += sorted_data[j][0][1] - sorted_data[j][0][0]
if (
j < n - 1
- and (batch_size_ms_cum + sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0])
- < batch_size
- and (sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0])
- < batch_size_threshold_ms
+ and sample_length < batch_size_threshold_ms
+ and potential_batch_length < batch_size
):
+ max_len_in_batch = max(max_len_in_batch, sample_length)
+ end_idx += 1
continue
- batch_size_ms_cum = 0
- end_idx = j + 1
+
speech_j, speech_lengths_j = slice_padding_audio_samples(
speech, speech_lengths, sorted_data[beg_idx:end_idx]
)
@@ -410,6 +417,8 @@
)
results[_b]["spk_embedding"] = spk_res[0]["spk_embedding"]
beg_idx = end_idx
+ end_idx += 1
+ max_len_in_batch = sample_length
if len(results) < 1:
continue
results_sorted.extend(results)
@@ -421,6 +430,10 @@
# f"time_speech_total_per_sample: {time_speech_total_per_sample: 0.3f}, "
# f"time_escape_total_per_sample: {time_escape_total_per_sample:0.3f}")
+ if len(results_sorted) != n:
+ results_ret_list.append({"key": key, "text": "", "timestamp": []})
+ logging.info("decoding, utt: {}, empty result".format(key))
+ continue
restored_data = [0] * n
for j in range(n):
index = sorted_data[j][1]
@@ -507,24 +520,40 @@
and 'iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch'\
can predict timestamp, and speaker diarization relies on timestamps."
)
- sentence_list = timestamp_sentence(
- punc_res[0]["punc_array"],
- result["timestamp"],
- raw_text,
- return_raw_text=return_raw_text,
- )
+ if kwargs.get("en_post_proc", False):
+ sentence_list = timestamp_sentence_en(
+ punc_res[0]["punc_array"],
+ result["timestamp"],
+ raw_text,
+ return_raw_text=return_raw_text,
+ )
+ else:
+ sentence_list = timestamp_sentence(
+ punc_res[0]["punc_array"],
+ result["timestamp"],
+ raw_text,
+ return_raw_text=return_raw_text,
+ )
distribute_spk(sentence_list, sv_output)
result["sentence_info"] = sentence_list
elif kwargs.get("sentence_timestamp", False):
if not len(result["text"].strip()):
sentence_list = []
else:
- sentence_list = timestamp_sentence(
- punc_res[0]["punc_array"],
- result["timestamp"],
- raw_text,
- return_raw_text=return_raw_text,
- )
+ if kwargs.get("en_post_proc", False):
+ sentence_list = timestamp_sentence_en(
+ punc_res[0]["punc_array"],
+ result["timestamp"],
+ raw_text,
+ return_raw_text=return_raw_text,
+ )
+ else:
+ sentence_list = timestamp_sentence(
+ punc_res[0]["punc_array"],
+ result["timestamp"],
+ raw_text,
+ return_raw_text=return_raw_text,
+ )
result["sentence_info"] = sentence_list
if "spk_embedding" in result:
del result["spk_embedding"]
--
Gitblit v1.9.1