From 2acd24f0158b2c86d2fb4e6f1134b67a1150500e Mon Sep 17 00:00:00 2001
From: jmwang66 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期四, 29 二月 2024 17:14:59 +0800
Subject: [PATCH] update whisper lid (#1407)
---
funasr/auto/auto_model.py | 227 ++++++++++++++++++++++++++++++++++----------------------
1 files changed, 138 insertions(+), 89 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 580cca8..64d4dec 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -1,25 +1,33 @@
+#!/usr/bin/env python3
+# -*- encoding: utf-8 -*-
+# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
+# MIT License (https://opensource.org/licenses/MIT)
+
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
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.vad_utils import slice_padding_audio_samples
+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.load_utils import load_audio_text_image_video, extract_fbank
-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):
@@ -87,7 +95,8 @@
class AutoModel:
def __init__(self, **kwargs):
- tables.print()
+ if not kwargs.get("disable_log", True):
+ tables.print()
model, kwargs = self.build_model(**kwargs)
@@ -96,7 +105,7 @@
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 +113,7 @@
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
@@ -112,17 +121,13 @@
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.")
self.spk_mode = spk_mode
- self.preset_spk_num = kwargs.get("preset_spk_num", None)
- if self.preset_spk_num:
- logging.warning("Using preset speaker number: {}".format(self.preset_spk_num))
- logging.warning("Many to print when using speaker model...")
self.kwargs = kwargs
self.model = model
@@ -132,8 +137,7 @@
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):
assert "model" in kwargs
@@ -144,9 +148,9 @@
set_all_random_seed(kwargs.get("seed", 0))
device = kwargs.get("device", "cuda")
- if not torch.cuda.is_available() or kwargs.get("ngpu", 0):
+ if not torch.cuda.is_available() or kwargs.get("ngpu", 1) == 0:
device = "cpu"
- # kwargs["batch_size"] = 1
+ kwargs["batch_size"] = 1
kwargs["device"] = device
if kwargs.get("ncpu", None):
@@ -158,8 +162,10 @@
tokenizer_class = tables.tokenizer_classes.get(tokenizer)
tokenizer = tokenizer_class(**kwargs["tokenizer_conf"])
kwargs["tokenizer"] = tokenizer
- kwargs["token_list"] = tokenizer.token_list
- vocab_size = len(tokenizer.token_list)
+
+ kwargs["token_list"] = tokenizer.token_list if hasattr(tokenizer, "token_list") else None
+ kwargs["token_list"] = tokenizer.get_vocab() if hasattr(tokenizer, "get_vocab") else kwargs["token_list"]
+ vocab_size = len(kwargs["token_list"])
else:
vocab_size = -1
@@ -174,19 +180,24 @@
# 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
init_param = kwargs.get("init_param", None)
if init_param is not None:
- logging.info(f"Loading pretrained params from {init_param}")
- load_pretrained_model(
- model=model,
- init_param=init_param,
- ignore_init_mismatch=kwargs.get("ignore_init_mismatch", False),
- oss_bucket=kwargs.get("oss_bucket", None),
- )
+ if os.path.exists(init_param):
+ logging.info(f"Loading pretrained params from {init_param}")
+ load_pretrained_model(
+ model=model,
+ 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", []),
+ excludes=kwargs.get("excludes", None),
+ )
+ else:
+ print(f"error, init_param does not exist!: {init_param}")
return model, kwargs
@@ -195,8 +206,6 @@
kwargs.update(cfg)
res = self.model(*args, kwargs)
return res
-
-
def generate(self, input, input_len=None, **cfg):
if self.vad_model is None:
@@ -209,6 +218,7 @@
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":
@@ -219,7 +229,8 @@
speed_stats = {}
asr_result_list = []
num_samples = len(data_list)
- pbar = tqdm(colour="blue", total=num_samples+1, dynamic_ncols=True)
+ 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
for beg_idx in range(0, num_samples, batch_size):
@@ -227,18 +238,17 @@
data_batch = data_list[beg_idx:end_idx]
key_batch = key_list[beg_idx:end_idx]
batch = {"data_in": data_batch, "key": key_batch}
- if (end_idx - beg_idx) == 1 and isinstance(data_batch[0], torch.Tensor): # fbank
+ 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)
- 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
@@ -250,39 +260,40 @@
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
-
+
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 = 0.0
+ time_speech_total_all_samples = 1e-6
beg_total = time.time()
- pbar_total = tqdm(colour="red", total=len(res) + 1, 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"]
@@ -293,21 +304,25 @@
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()
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, 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 (
@@ -315,15 +330,14 @@
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])
+ 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:
- 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, \
- 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]
@@ -334,20 +348,19 @@
continue
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}, "
- 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}")
+ # 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):
@@ -364,7 +377,7 @@
result[k] = restored_data[j][k]
else:
result[k] = torch.cat([result[k], restored_data[j][k]], dim=0)
- elif k == 'text':
+ elif 'text' in k:
if k not in result:
result[k] = restored_data[j][k]
else:
@@ -374,43 +387,79 @@
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, **cfg)
- result["text_with_punc"] = punc_res[0]["text"]
-
+ if not len(result["text"]):
+ if return_raw_text:
+ result['raw_text'] = ''
+ else:
+ self.punc_kwargs.update(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
+
# speaker embedding cluster after resorted
- if self.spk_model is not None:
+ 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, oracle_num=self.preset_spk_num)
- del result['spk_embedding']
+ labels = self.cb_model(spk_embedding.cpu(), oracle_num=kwargs.get('preset_spk_num', None))
+ # del result['spk_embedding']
sv_output = postprocess(all_segments, None, labels, spk_embedding.cpu())
- if self.spk_mode == 'vad_segment':
+ 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['text'],
- "timestamp": res['timestamp']})
- else: # punc_segment
- sentence_list = timestamp_sentence(punc_res[0]['punc_array'], \
- result['timestamp'], \
- result['text'])
+ 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':
+ 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):
+ if not len(result['text']):
+ sentence_list = []
+ 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']
+
result["key"] = key
results_ret_list.append(result)
- pbar_total.update(1)
-
- pbar_total.update(1)
- end_total = time.time()
- time_escape_total_all_samples = end_total - beg_total
- pbar_total.set_description(f"rtf_avg_all_samples: {time_escape_total_all_samples / time_speech_total_all_samples:0.3f}, "
- f"time_speech_total_all_samples: {time_speech_total_all_samples: 0.3f}, "
- f"time_escape_total_all_samples: {time_escape_total_all_samples:0.3f}")
+ end_asr_total = time.time()
+ time_escape_total_per_sample = end_asr_total - beg_asr_total
+ 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}")
+
+
+ # end_total = time.time()
+ # time_escape_total_all_samples = end_total - beg_total
+ # print(f"rtf_avg_all: {time_escape_total_all_samples / time_speech_total_all_samples:0.3f}, "
+ # f"time_speech_all: {time_speech_total_all_samples: 0.3f}, "
+ # f"time_escape_all: {time_escape_total_all_samples:0.3f}")
return results_ret_list
--
Gitblit v1.9.1