From b4affb21ef34ac11cb0252b386080e6494fbfb30 Mon Sep 17 00:00:00 2001
From: xuan <admin@exuan.org>
Date: 星期一, 26 二月 2024 09:41:22 +0800
Subject: [PATCH] fix: TabError (#1388)
---
funasr/auto/auto_model.py | 98 +++++++++++++++++++++++++++++--------------------
1 files changed, 58 insertions(+), 40 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 8e00703..66c0750 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -1,14 +1,13 @@
import json
import time
+import copy
import torch
-import hydra
import random
import string
import logging
import os.path
import numpy as np
from tqdm import tqdm
-from omegaconf import DictConfig, OmegaConf, ListConfig
from funasr.register import tables
from funasr.utils.load_utils import load_bytes
@@ -17,10 +16,13 @@
from funasr.utils.vad_utils import slice_padding_audio_samples
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.load_utils import load_audio_text_image_video, extract_fbank
+from funasr.utils.load_utils import load_audio_text_image_video
from funasr.utils.timestamp_tools import timestamp_sentence
from funasr.models.campplus.utils import sv_chunk, postprocess, distribute_spk
-from funasr.models.campplus.cluster_backend import ClusterBackend
+try:
+ from funasr.models.campplus.cluster_backend import ClusterBackend
+except:
+ print("If you want to use the speaker diarization, please `pip install hdbscan`")
def prepare_data_iterator(data_in, input_len=None, data_type=None, key=None):
@@ -171,7 +173,7 @@
# build model
model_class = tables.model_classes.get(kwargs["model"])
model = model_class(**kwargs, **kwargs["model_conf"], vocab_size=vocab_size)
- model.eval()
+
model.to(device)
# init_param
@@ -206,6 +208,7 @@
kwargs = self.kwargs if kwargs is None else kwargs
kwargs.update(cfg)
model = self.model if model is None else model
+ model.eval()
batch_size = kwargs.get("batch_size", 1)
# if kwargs.get("device", "cpu") == "cpu":
@@ -216,7 +219,7 @@
speed_stats = {}
asr_result_list = []
num_samples = len(data_list)
- disable_pbar = kwargs.get("disable_pbar", False)
+ disable_pbar = self.kwargs.get("disable_pbar", False)
pbar = tqdm(colour="blue", total=num_samples, dynamic_ncols=True) if not disable_pbar else None
time_speech_total = 0.0
time_escape_total = 0.0
@@ -228,12 +231,12 @@
if (end_idx - beg_idx) == 1 and kwargs.get("data_type", None) == "fbank": # fbank
batch["data_in"] = data_batch[0]
batch["data_lengths"] = input_len
-
+
time1 = time.perf_counter()
with torch.no_grad():
results, meta_data = model.inference(**batch, **kwargs)
time2 = time.perf_counter()
-
+
asr_result_list.extend(results)
# batch_data_time = time_per_frame_s * data_batch_i["speech_lengths"].sum().item()
@@ -258,31 +261,29 @@
pbar.set_description(f"rtf_avg: {time_escape_total/time_speech_total:0.3f}")
torch.cuda.empty_cache()
return asr_result_list
-
+
def inference_with_vad(self, input, input_len=None, **cfg):
-
+ kwargs = self.kwargs
# step.1: compute the vad model
self.vad_kwargs.update(cfg)
beg_vad = time.time()
res = self.inference(input, input_len=input_len, model=self.vad_model, kwargs=self.vad_kwargs, **cfg)
end_vad = time.time()
- print(f"time cost vad: {end_vad - beg_vad:0.3f}")
# step.2 compute asr model
model = self.model
- kwargs = self.kwargs
kwargs.update(cfg)
batch_size = int(kwargs.get("batch_size_s", 300))*1000
batch_size_threshold_ms = int(kwargs.get("batch_size_threshold_s", 60))*1000
kwargs["batch_size"] = batch_size
-
+
key_list, data_list = prepare_data_iterator(input, input_len=input_len, data_type=kwargs.get("data_type", None))
results_ret_list = []
time_speech_total_all_samples = 1e-6
beg_total = time.time()
- pbar_total = tqdm(colour="red", total=len(res), dynamic_ncols=True)
+ pbar_total = tqdm(colour="red", total=len(res), dynamic_ncols=True) if not kwargs.get("disable_pbar", False) else None
for i in range(len(res)):
key = res[i]["key"]
vadsegments = res[i]["value"]
@@ -293,14 +294,14 @@
data_with_index = [(vadsegments[i], i) for i in range(n)]
sorted_data = sorted(data_with_index, key=lambda x: x[0][1] - x[0][0])
results_sorted = []
-
+
if not len(sorted_data):
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()
@@ -319,8 +320,8 @@
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])
- results = self.inference(speech_j, input_len=None, model=model, kwargs=kwargs, disable_pbar=True, **cfg)
+ 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)
if self.spk_model is not None:
# compose vad segments: [[start_time_sec, end_time_sec, speech], [...]]
for _b in range(len(speech_j)):
@@ -330,26 +331,26 @@
segments = sv_chunk(vad_segments)
all_segments.extend(segments)
speech_b = [i[2] for i in segments]
- spk_res = self.inference(speech_b, input_len=None, model=self.spk_model, kwargs=kwargs, disable_pbar=True, **cfg)
+ spk_res = self.inference(speech_b, input_len=None, model=self.spk_model, kwargs=kwargs, **cfg)
results[_b]['spk_embedding'] = spk_res[0]['spk_embedding']
beg_idx = end_idx
if len(results) < 1:
continue
results_sorted.extend(results)
-
+
# end_asr_total = time.time()
# time_escape_total_per_sample = end_asr_total - beg_asr_total
# pbar_sample.update(1)
# 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):
index = sorted_data[j][1]
restored_data[index] = results_sorted[j]
result = {}
-
+
# results combine for texts, timestamps, speaker embeddings and others
# TODO: rewrite for clean code
for j in range(n):
@@ -376,16 +377,22 @@
result[k] = restored_data[j][k]
else:
result[k] += restored_data[j][k]
-
+
+ return_raw_text = kwargs.get('return_raw_text', False)
# step.3 compute punc model
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, disable_pbar=True, **cfg)
- import copy; raw_text = copy.copy(result["text"])
+ 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
+
# speaker embedding cluster after resorted
if self.spk_model is not None and kwargs.get('return_spk_res', True):
+ if raw_text is None:
+ logging.error("Missing punc_model, which is required by spk_model.")
all_segments = sorted(all_segments, key=lambda x: x[0])
spk_embedding = result['spk_embedding']
labels = self.cb_model(spk_embedding.cpu(), oracle_num=kwargs.get('preset_spk_num', None))
@@ -394,29 +401,40 @@
if self.spk_mode == 'vad_segment': # recover sentence_list
sentence_list = []
for res, vadsegment in zip(restored_data, vadsegments):
- sentence_list.append({"start": vadsegment[0],\
- "end": vadsegment[1],
- "sentence": res['raw_text'],
- "timestamp": res['timestamp']})
+ if 'timestamp' not in res:
+ logging.error("Only 'iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch' \
+ 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.append({"start": vadsegment[0],
+ "end": vadsegment[1],
+ "sentence": res['text'],
+ "timestamp": res['timestamp']})
elif self.spk_mode == 'punc_segment':
- sentence_list = timestamp_sentence(punc_res[0]['punc_array'], \
- result['timestamp'], \
- result['raw_text'])
+ if 'timestamp' not in result:
+ logging.error("Only 'iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch' \
+ 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)
distribute_spk(sentence_list, sv_output)
result['sentence_info'] = sentence_list
elif kwargs.get("sentence_timestamp", False):
- sentence_list = timestamp_sentence(punc_res[0]['punc_array'], \
- result['timestamp'], \
- result['raw_text'])
+ sentence_list = timestamp_sentence(punc_res[0]['punc_array'],
+ result['timestamp'],
+ raw_text,
+ return_raw_text=return_raw_text)
result['sentence_info'] = sentence_list
- del result['spk_embedding']
-
+ if "spk_embedding" in result: del result['spk_embedding']
+
result["key"] = key
results_ret_list.append(result)
end_asr_total = time.time()
time_escape_total_per_sample = end_asr_total - beg_asr_total
- pbar_total.update(1)
- pbar_total.set_description(f"rtf_avg: {time_escape_total_per_sample / time_speech_total_per_sample:0.3f}, "
+ if pbar_total:
+ pbar_total.update(1)
+ pbar_total.set_description(f"rtf_avg: {time_escape_total_per_sample / time_speech_total_per_sample:0.3f}, "
f"time_speech: {time_speech_total_per_sample: 0.3f}, "
f"time_escape: {time_escape_total_per_sample:0.3f}")
--
Gitblit v1.9.1