From 3919d7454c070702e94b149e4032e9db08d28fa3 Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期一, 22 一月 2024 15:42:45 +0800
Subject: [PATCH] Funasr1.0 (#1279)
---
funasr/auto/auto_model.py | 46 +++++++++++++++++++++++++++++-----------------
1 files changed, 29 insertions(+), 17 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 580cca8..ca6189d 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -132,7 +132,8 @@
self.punc_kwargs = punc_kwargs
self.spk_model = spk_model
self.spk_kwargs = spk_kwargs
- self.model_path = kwargs["model_path"]
+ self.model_path = kwargs.get("model_path")
+
def build_model(self, **kwargs):
@@ -146,7 +147,7 @@
device = kwargs.get("device", "cuda")
if not torch.cuda.is_available() or kwargs.get("ngpu", 0):
device = "cpu"
- # kwargs["batch_size"] = 1
+ kwargs["batch_size"] = 1
kwargs["device"] = device
if kwargs.get("ncpu", None):
@@ -183,9 +184,11 @@
logging.info(f"Loading pretrained params from {init_param}")
load_pretrained_model(
model=model,
- init_param=init_param,
+ path=init_param,
ignore_init_mismatch=kwargs.get("ignore_init_mismatch", False),
oss_bucket=kwargs.get("oss_bucket", None),
+ scope_map=kwargs.get("scope_map", None),
+ excludes=kwargs.get("excludes", None),
)
return model, kwargs
@@ -219,7 +222,8 @@
speed_stats = {}
asr_result_list = []
num_samples = len(data_list)
- pbar = tqdm(colour="blue", total=num_samples+1, dynamic_ncols=True)
+ disable_pbar = kwargs.get("disable_pbar", False)
+ pbar = tqdm(colour="blue", total=num_samples+1, dynamic_ncols=True) if not disable_pbar else None
time_speech_total = 0.0
time_escape_total = 0.0
for beg_idx in range(0, num_samples, batch_size):
@@ -237,8 +241,7 @@
time2 = time.perf_counter()
asr_result_list.extend(results)
- pbar.update(1)
-
+
# batch_data_time = time_per_frame_s * data_batch_i["speech_lengths"].sum().item()
batch_data_time = meta_data.get("batch_data_time", -1)
time_escape = time2 - time1
@@ -250,12 +253,15 @@
description = (
f"{speed_stats}, "
)
- pbar.set_description(description)
+ if pbar:
+ pbar.update(1)
+ pbar.set_description(description)
time_speech_total += batch_data_time
time_escape_total += time_escape
-
- pbar.update(1)
- pbar.set_description(f"rtf_avg: {time_escape_total/time_speech_total:0.3f}")
+
+ if pbar:
+ pbar.update(1)
+ pbar.set_description(f"rtf_avg: {time_escape_total/time_speech_total:0.3f}")
torch.cuda.empty_cache()
return asr_result_list
@@ -307,7 +313,11 @@
time_speech_total_per_sample = speech_lengths/16000
time_speech_total_all_samples += time_speech_total_per_sample
+ pbar_sample = tqdm(colour="blue", total=n + 1, dynamic_ncols=True)
+
+ all_segments = []
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])
if j < n - 1 and (
batch_size_ms_cum + sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0]) < batch_size and (
@@ -316,13 +326,14 @@
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])
- results = self.inference(speech_j, input_len=None, model=model, kwargs=kwargs, **cfg)
+ results = self.inference(speech_j, input_len=None, model=model, kwargs=kwargs, disable_pbar=True, **cfg)
if self.spk_model is not None:
- all_segments = []
+
+
# compose vad segments: [[start_time_sec, end_time_sec, speech], [...]]
for _b in range(len(speech_j)):
- vad_segments = [[sorted_data[beg_idx:end_idx][_b][0][0]/1000.0, \
- sorted_data[beg_idx:end_idx][_b][0][1]/1000.0, \
+ vad_segments = [[sorted_data[beg_idx:end_idx][_b][0][0]/1000.0,
+ sorted_data[beg_idx:end_idx][_b][0][1]/1000.0,
speech_j[_b]]]
segments = sv_chunk(vad_segments)
all_segments.extend(segments)
@@ -335,12 +346,13 @@
results_sorted.extend(results)
- pbar_total.update(1)
+
end_asr_total = time.time()
time_escape_total_per_sample = end_asr_total - beg_asr_total
- pbar_total.set_description(f"rtf_avg_per_sample: {time_escape_total_per_sample / time_speech_total_per_sample:0.3f}, "
+ pbar_sample.set_description(f"rtf_avg_per_sample: {time_escape_total_per_sample / time_speech_total_per_sample:0.3f}, "
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}")
+
restored_data = [0] * n
for j in range(n):
@@ -379,7 +391,7 @@
if self.punc_model is not None:
self.punc_kwargs.update(cfg)
punc_res = self.inference(result["text"], model=self.punc_model, kwargs=self.punc_kwargs, **cfg)
- result["text_with_punc"] = punc_res[0]["text"]
+ result["text"] = punc_res[0]["text"]
# speaker embedding cluster after resorted
if self.spk_model is not None:
--
Gitblit v1.9.1