From f2ed4b3856eaed8abe568e6904ffd8dc3a799f5f Mon Sep 17 00:00:00 2001
From: 彭震东 <zhendong.peng@qq.com>
Date: 星期一, 15 七月 2024 17:54:27 +0800
Subject: [PATCH] fix progress bar for batch_size (#1917)
---
funasr/auto/auto_model.py | 160 ++++++++++++++++++++++++++++++++---------------------
1 files changed, 97 insertions(+), 63 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 32fd560..1c2d2e4 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -19,13 +19,15 @@
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.download.download_from_hub import download_model
+from funasr.utils.timestamp_tools import timestamp_sentence_en
+from funasr.download.download_model_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
from funasr.utils import export_utils
+from funasr.utils import misc
try:
from funasr.models.campplus.utils import sv_chunk, postprocess, distribute_spk
@@ -35,21 +37,15 @@
def prepare_data_iterator(data_in, input_len=None, data_type=None, key=None):
- """
-
- :param input:
- :param input_len:
- :param data_type:
- :param frontend:
- :return:
- """
+ """ """
data_list = []
key_list = []
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
@@ -73,7 +69,8 @@
key_list.append(key)
else:
if key is None:
- key = "rand_key_" + "".join(random.choice(chars) for _ in range(13))
+ # key = "rand_key_" + "".join(random.choice(chars) for _ in range(13))
+ key = misc.extract_filename_without_extension(data_in)
data_list = [data_in]
key_list = [key]
elif isinstance(data_in, (list, tuple)):
@@ -90,10 +87,15 @@
else:
# [audio sample point, fbank, text]
data_list = data_in
- key_list = [
- "rand_key_" + "".join(random.choice(chars) for _ in range(13))
- for _ in range(len(data_in))
- ]
+ key_list = []
+ for data_i in data_in:
+ if isinstance(data_i, str) and os.path.exists(data_i):
+ key = misc.extract_filename_without_extension(data_i)
+ else:
+ if key is None:
+ key = "rand_key_" + "".join(random.choice(chars) for _ in range(13))
+ key_list.append(key)
+
else: # raw text; audio sample point, fbank; bytes
if isinstance(data_in, bytes): # audio bytes
data_in = load_bytes(data_in)
@@ -108,8 +110,16 @@
class AutoModel:
def __init__(self, **kwargs):
- if not kwargs.get("disable_log", True):
- tables.print()
+
+ try:
+ from funasr.utils.version_checker import check_for_update
+
+ check_for_update()
+ except:
+ pass
+
+ log_level = getattr(logging, kwargs.get("log_level", "INFO").upper())
+ logging.basicConfig(level=log_level)
model, kwargs = self.build_model(**kwargs)
@@ -158,7 +168,8 @@
self.spk_kwargs = spk_kwargs
self.model_path = kwargs.get("model_path")
- def build_model(self, **kwargs):
+ @staticmethod
+ def build_model(**kwargs):
assert "model" in kwargs
if "model_conf" not in kwargs:
logging.info("download models from model hub: {}".format(kwargs.get("hub", "ms")))
@@ -204,11 +215,11 @@
kwargs["frontend"] = frontend
# build model
model_class = tables.model_classes.get(kwargs["model"])
+ assert model_class is not None, f'{kwargs["model"]} is not registered'
model_conf = {}
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)
@@ -229,6 +240,13 @@
# fp16
if kwargs.get("fp16", False):
model.to(torch.float16)
+ elif kwargs.get("bf16", False):
+ model.to(torch.bfloat16)
+ model.to(device)
+
+ if not kwargs.get("disable_log", True):
+ tables.print()
+
return model, kwargs
def __call__(self, *args, **cfg):
@@ -281,7 +299,7 @@
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()
@@ -297,7 +315,7 @@
speed_stats["rtf"] = f"{(time_escape) / batch_data_time:0.3f}"
description = f"{speed_stats}, "
if pbar:
- pbar.update(1)
+ pbar.update(end_idx - beg_idx)
pbar.set_description(description)
time_speech_total += batch_data_time
time_escape_total += time_escape
@@ -355,13 +373,13 @@
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
@@ -370,19 +388,22 @@
# 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]
)
@@ -407,6 +428,8 @@
)
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)
@@ -418,6 +441,10 @@
# 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]
@@ -451,23 +478,20 @@
else:
result[k] += restored_data[j][k]
+ if not len(result["text"].strip()):
+ continue
return_raw_text = kwargs.get("return_raw_text", False)
# step.3 compute punc model
+ raw_text = None
if self.punc_model is not None:
- if not len(result["text"].strip()):
- if return_raw_text:
- result["raw_text"] = ""
- else:
- deep_update(self.punc_kwargs, cfg)
- punc_res = self.inference(
- result["text"], model=self.punc_model, kwargs=self.punc_kwargs, **cfg
- )
- raw_text = copy.copy(result["text"])
- if return_raw_text:
- result["raw_text"] = raw_text
- result["text"] = punc_res[0]["text"]
- else:
- raw_text = None
+ deep_update(self.punc_kwargs, cfg)
+ punc_res = self.inference(
+ result["text"], model=self.punc_model, kwargs=self.punc_kwargs, **cfg
+ )
+ raw_text = copy.copy(result["text"])
+ if return_raw_text:
+ result["raw_text"] = raw_text
+ result["text"] = punc_res[0]["text"]
# speaker embedding cluster after resorted
if self.spk_model is not None and kwargs.get("return_spk_res", True):
@@ -504,24 +528,40 @@
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"]
@@ -573,12 +613,6 @@
)
with torch.no_grad():
-
- if type == "onnx":
- export_dir = export_utils.export_onnx(model=model, data_in=data_list, **kwargs)
- else:
- export_dir = export_utils.export_torchscripts(
- model=model, data_in=data_list, **kwargs
- )
+ export_dir = export_utils.export(model=model, data_in=data_list, **kwargs)
return export_dir
--
Gitblit v1.9.1