From 6d66287c2d352d77d4022c5e7da5743be58b7387 Mon Sep 17 00:00:00 2001
From: shixian.shi <shixian.shi@alibaba-inc.com>
Date: 星期一, 22 一月 2024 18:23:30 +0800
Subject: [PATCH] update device
---
funasr/auto/auto_model.py | 50 ++++++++++++++++++++++++++++++--------------------
1 files changed, 30 insertions(+), 20 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 740614c..107c78e 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -6,6 +6,7 @@
import string
import logging
import os.path
+import numpy as np
from tqdm import tqdm
from omegaconf import DictConfig, OmegaConf, ListConfig
@@ -96,7 +97,7 @@
vad_kwargs = kwargs.get("vad_model_revision", None)
if vad_model is not None:
logging.info("Building VAD model.")
- vad_kwargs = {"model": vad_model, "model_revision": vad_kwargs}
+ vad_kwargs = {"model": vad_model, "model_revision": vad_kwargs, "device": kwargs["device"]}
vad_model, vad_kwargs = self.build_model(**vad_kwargs)
# if punc_model is not None, build punc model else None
@@ -104,7 +105,7 @@
punc_kwargs = kwargs.get("punc_model_revision", None)
if punc_model is not None:
logging.info("Building punc model.")
- punc_kwargs = {"model": punc_model, "model_revision": punc_kwargs}
+ punc_kwargs = {"model": punc_model, "model_revision": punc_kwargs, "device": kwargs["device"]}
punc_model, punc_kwargs = self.build_model(**punc_kwargs)
# if spk_model is not None, build spk model else None
@@ -112,9 +113,9 @@
spk_kwargs = kwargs.get("spk_model_revision", None)
if spk_model is not None:
logging.info("Building SPK model.")
- spk_kwargs = {"model": spk_model, "model_revision": spk_kwargs}
+ spk_kwargs = {"model": spk_model, "model_revision": spk_kwargs, "device": kwargs["device"]}
spk_model, spk_kwargs = self.build_model(**spk_kwargs)
- self.cb_model = ClusterBackend()
+ self.cb_model = ClusterBackend().to(kwargs["device"])
spk_mode = kwargs.get("spk_mode", 'punc_segment')
if spk_mode not in ["default", "vad_segment", "punc_segment"]:
logging.error("spk_mode should be one of default, vad_segment and punc_segment.")
@@ -132,7 +133,8 @@
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):
@@ -221,7 +223,8 @@
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):
@@ -239,8 +242,7 @@
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
@@ -252,12 +254,15 @@
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
@@ -309,8 +314,11 @@
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 (
@@ -319,14 +327,15 @@
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:
-
+
+
# 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, \
- speech_j[_b]]]
+ vad_segments = [[sorted_data[beg_idx:end_idx][_b][0][0]/1000.0,
+ sorted_data[beg_idx:end_idx][_b][0][1]/1000.0,
+ np.array(speech_j[_b])]]
segments = sv_chunk(vad_segments)
all_segments.extend(segments)
speech_b = [i[2] for i in segments]
@@ -338,12 +347,13 @@
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):
@@ -382,7 +392,7 @@
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