mirror of
https://github.com/modelscope/FunASR.git
synced 2026-10-12 02:19:18 +00:00
Shi Xian
@@ -79,6 +79,7 @@ FunASR has open-sourced a large number of pre-trained models on industrial data.
|
||||
| fsmn-vad <br> ( [⭐](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary) [🤗](https://huggingface.co/funasr/fsmn-vad) ) | voice activity detection | 5000 hours, Mandarin and English | 0.4M |
|
||||
| fa-zh <br> ( [⭐](https://modelscope.cn/models/damo/speech_timestamp_prediction-v1-16k-offline/summary) [🤗](https://huggingface.co/funasr/fa-zh) ) | timestamp prediction | 5000 hours, Mandarin | 38M |
|
||||
| cam++ <br> ( [⭐](https://modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common/summary) [🤗](https://huggingface.co/funasr/campplus) ) | speaker verification/diarization | 5000 hours | 7.2M |
|
||||
| whisper-large-v2 <br> ([⭐](https://www.modelscope.cn/models/iic/speech_whisper-large_asr_multilingual/summary) [🤗]() ) | speech recognition, with timestamps, non-streaming | multilingual | 1G |
|
||||
|
||||
|
||||
|
||||
|
||||
+11
-10
@@ -71,16 +71,17 @@ FunASR开源了大量在工业数据上预训练模型，您可以在[模型许
|
||||
（注：⭐ 表示ModelScope模型仓库链接，🤗 表示Huggingface模型仓库链接）
|
||||
|
||||
|
||||
| 模型名字 | 任务详情 | 训练数据 | 参数量 |
|
||||
|:------------------------------------------------------------------------------------------------------------------------------------------------------------:|:------------------:|:------------:|:----:|
|
||||
| paraformer-zh <br> ([⭐](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) [🤗](https://huggingface.co/funasr/paraformer-tp) ) | 语音识别，带时间戳输出，非实时 | 60000小时，中文 | 220M |
|
||||
| paraformer-zh-streaming <br> ( [⭐](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online/summary) [🤗](https://huggingface.co/funasr/paraformer-zh-streaming) ) | 语音识别，实时 | 60000小时，中文 | 220M |
|
||||
| paraformer-en <br> ( [⭐](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-en-16k-common-vocab10020/summary) [🤗](https://huggingface.co/funasr/paraformer-en) ) | 语音识别，非实时 | 50000小时，英文 | 220M |
|
||||
| conformer-en <br> ( [⭐](https://modelscope.cn/models/damo/speech_conformer_asr-en-16k-vocab4199-pytorch/summary) [🤗](https://huggingface.co/funasr/conformer-en) ) | 语音识别，非实时 | 50000小时，英文 | 220M |
|
||||
| ct-punc <br> ( [⭐](https://modelscope.cn/models/damo/punc_ct-transformer_cn-en-common-vocab471067-large/summary) [🤗](https://huggingface.co/funasr/ct-punc) ) | 标点恢复 | 100M，中文与英文 | 1.1G |
|
||||
| fsmn-vad <br> ( [⭐](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary) [🤗](https://huggingface.co/funasr/fsmn-vad) ) | 语音端点检测，实时 | 5000小时，中文与英文 | 0.4M |
|
||||
| fa-zh <br> ( [⭐](https://modelscope.cn/models/damo/speech_timestamp_prediction-v1-16k-offline/summary) [🤗](https://huggingface.co/funasr/fa-zh) ) | 字级别时间戳预测 | 50000小时，中文 | 38M |
|
||||
| cam++ <br> ( [⭐](https://modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common/summary) [🤗](https://huggingface.co/funasr/campplus) ) | 说话人确认/分割 | 5000小时 | 7.2M |
|
||||
| 模型名字 | 任务详情 | 训练数据 | 参数量 |
|
||||
|:------------------------------------------------------------------------------------------------------------------------------------------------------------:|:---------------:|:------------:|:----:|
|
||||
| paraformer-zh <br> ([⭐](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) [🤗](https://huggingface.co/funasr/paraformer-tp) ) | 语音识别，带时间戳输出，非实时 | 60000小时，中文 | 220M |
|
||||
| paraformer-zh-streaming <br> ( [⭐](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-online/summary) [🤗](https://huggingface.co/funasr/paraformer-zh-streaming) ) | 语音识别，实时 | 60000小时，中文 | 220M |
|
||||
| paraformer-en <br> ( [⭐](https://www.modelscope.cn/models/damo/speech_paraformer-large-vad-punc_asr_nat-en-16k-common-vocab10020/summary) [🤗](https://huggingface.co/funasr/paraformer-en) ) | 语音识别，非实时 | 50000小时，英文 | 220M |
|
||||
| conformer-en <br> ( [⭐](https://modelscope.cn/models/damo/speech_conformer_asr-en-16k-vocab4199-pytorch/summary) [🤗](https://huggingface.co/funasr/conformer-en) ) | 语音识别，非实时 | 50000小时，英文 | 220M |
|
||||
| ct-punc <br> ( [⭐](https://modelscope.cn/models/damo/punc_ct-transformer_cn-en-common-vocab471067-large/summary) [🤗](https://huggingface.co/funasr/ct-punc) ) | 标点恢复 | 100M，中文与英文 | 1.1G |
|
||||
| fsmn-vad <br> ( [⭐](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/summary) [🤗](https://huggingface.co/funasr/fsmn-vad) ) | 语音端点检测，实时 | 5000小时，中文与英文 | 0.4M |
|
||||
| fa-zh <br> ( [⭐](https://modelscope.cn/models/damo/speech_timestamp_prediction-v1-16k-offline/summary) [🤗](https://huggingface.co/funasr/fa-zh) ) | 字级别时间戳预测 | 50000小时，中文 | 38M |
|
||||
| cam++ <br> ( [⭐](https://modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common/summary) [🤗](https://huggingface.co/funasr/campplus) ) | 说话人确认/分割 | 5000小时 | 7.2M |
|
||||
| whisper-large-v2 <br> ([⭐](https://www.modelscope.cn/models/iic/speech_whisper-large_asr_multilingual/summary) [🤗]() ) | 语音识别，带时间戳输出，非实时 | 多语言 | 1G |
|
||||
|
||||
|
||||
<a name="快速开始"></a>
|
||||
|
||||
@@ -0,0 +1,13 @@
|
||||
#!/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)
|
||||
|
||||
from funasr import AutoModel
|
||||
|
||||
model = AutoModel(model="iic/speech_whisper-large_asr_multilingual",
|
||||
model_revision="v2.0.4",
|
||||
)
|
||||
|
||||
res = model.generate(input="https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav", language=None)
|
||||
print(res)
|
||||
@@ -0,0 +1,21 @@
|
||||
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
|
||||
# MIT License (https://opensource.org/licenses/MIT)
|
||||
|
||||
# method1, inference from model hub
|
||||
|
||||
# for more input type, please ref to readme.md
|
||||
input="https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav"
|
||||
|
||||
output_dir="./outputs/debug"
|
||||
|
||||
model="iic/speech_whisper-large_asr_multilingual"
|
||||
model_revision="v2.0.4"
|
||||
|
||||
device="cuda:0" # "cuda:0" for gpu0, "cuda:1" for gpu1, "cpu"
|
||||
|
||||
python -m funasr.bin.inference \
|
||||
++model=${model} \
|
||||
++model_revision=${model_revision} \
|
||||
++input="${input}" \
|
||||
++output_dir="${output_dir}" \
|
||||
++device="${device}" \
|
||||
@@ -0,0 +1,34 @@
|
||||
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
|
||||
# MIT License (https://opensource.org/licenses/MIT)
|
||||
|
||||
# method2, inference from local model
|
||||
|
||||
# for more input type, please ref to readme.md
|
||||
input="https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav"
|
||||
|
||||
output_dir="./outputs/debug"
|
||||
|
||||
workspace=`pwd`
|
||||
|
||||
# download model
|
||||
local_path_root=${workspace}/modelscope_models
|
||||
mkdir -p ${local_path_root}
|
||||
local_path=${local_path_root}/speech_whisper-large_asr_multilingual
|
||||
git clone https://www.modelscope.cn/iic/speech_whisper-large_asr_multilingual.git ${local_path}
|
||||
|
||||
device="cuda:0" # "cuda:0" for gpu0, "cuda:1" for gpu1, "cpu"
|
||||
|
||||
config="config.yaml"
|
||||
init_param="${local_path}/large-v2.pt"
|
||||
|
||||
python -m funasr.bin.inference \
|
||||
--config-path "${local_path}" \
|
||||
--config-name "${config}" \
|
||||
++init_param="${init_param}" \
|
||||
++input="${input}" \
|
||||
++output_dir="${output_dir}" \
|
||||
++device="${device}" \
|
||||
|
||||
|
||||
|
||||
|
||||
@@ -165,17 +165,18 @@ class AutoModel:
|
||||
|
||||
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"])
|
||||
vocab_size = len(kwargs["token_list"]) if kwargs["token_list"] is not None else -1
|
||||
else:
|
||||
vocab_size = -1
|
||||
|
||||
# build frontend
|
||||
frontend = kwargs.get("frontend", None)
|
||||
kwargs["input_size"] = None
|
||||
if frontend is not None:
|
||||
frontend_class = tables.frontend_classes.get(frontend)
|
||||
frontend = frontend_class(**kwargs["frontend_conf"])
|
||||
kwargs["frontend"] = frontend
|
||||
kwargs["input_size"] = frontend.output_size()
|
||||
kwargs["input_size"] = frontend.output_size() if hasattr(frontend, "output_size") else None
|
||||
|
||||
# build model
|
||||
model_class = tables.model_classes.get(kwargs["model"])
|
||||
|
||||
@@ -129,3 +129,129 @@ class AudioLLMDataset(torch.utils.data.Dataset):
|
||||
|
||||
outputs[key] = torch.nn.utils.rnn.pad_sequence(data_list, batch_first=True, padding_value=pad_value)
|
||||
return outputs
|
||||
|
||||
|
||||
@tables.register("dataset_classes", "AudioLLMARDataset")
|
||||
class AudioLLMARDataset(torch.utils.data.Dataset):
|
||||
"""
|
||||
AudioLLMDataset
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
path,
|
||||
index_ds: str = None,
|
||||
frontend=None,
|
||||
tokenizer=None,
|
||||
int_pad_value: int = -1,
|
||||
float_pad_value: float = 0.0,
|
||||
**kwargs):
|
||||
super().__init__()
|
||||
index_ds_class = tables.index_ds_classes.get(index_ds)
|
||||
self.index_ds = index_ds_class(path, **kwargs)
|
||||
preprocessor_speech = kwargs.get("preprocessor_speech", None)
|
||||
if preprocessor_speech:
|
||||
preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
|
||||
preprocessor_speech = preprocessor_speech_class(**kwargs.get("preprocessor_speech_conf", {}))
|
||||
self.preprocessor_speech = preprocessor_speech
|
||||
preprocessor_text = kwargs.get("preprocessor_text", None)
|
||||
if preprocessor_text:
|
||||
preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
|
||||
preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf", {}))
|
||||
self.preprocessor_text = preprocessor_text
|
||||
|
||||
self.frontend = frontend
|
||||
self.fs = 16000 if frontend is None else frontend.fs
|
||||
self.data_type = "sound"
|
||||
self.tokenizer = tokenizer
|
||||
|
||||
self.float_pad_value = float_pad_value
|
||||
self.prompt = kwargs.get("prompt", "Transcribe speech to text.")
|
||||
self.prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(
|
||||
self.prompt) # "USER: \nINSTRUCTION: {}\nnINPUT: {}\nASSISTANT: "
|
||||
self.prompt_af = ""
|
||||
self.IGNORE_INDEX = kwargs.get("IGNORE_INDEX", -100)
|
||||
self.int_pad_value = self.IGNORE_INDEX
|
||||
|
||||
def get_source_len(self, index):
|
||||
item = self.index_ds[index]
|
||||
return self.index_ds.get_source_len(item)
|
||||
|
||||
def get_target_len(self, index):
|
||||
item = self.index_ds[index]
|
||||
return self.index_ds.get_target_len(item)
|
||||
|
||||
def __len__(self):
|
||||
return len(self.index_ds)
|
||||
|
||||
def __getitem__(self, index):
|
||||
item = self.index_ds[index]
|
||||
# import pdb;
|
||||
# pdb.set_trace()
|
||||
source = item["source"]
|
||||
data_src = load_audio_text_image_video(source, fs=self.fs)
|
||||
if self.preprocessor_speech:
|
||||
data_src = self.preprocessor_speech(data_src, fs=self.fs)
|
||||
speech, speech_lengths = extract_fbank(data_src, data_type=self.data_type, frontend=self.frontend,
|
||||
is_final=True) # speech: [b, T, d]
|
||||
speech = speech.squeeze(0)
|
||||
|
||||
target = item["target"]
|
||||
if self.preprocessor_text:
|
||||
target = self.preprocessor_text(target)
|
||||
|
||||
prompt_ids_pre = self.tokenizer.encode(self.prompt_pre) # [bos,prompt]
|
||||
prompt_pre_length = len(prompt_ids_pre)
|
||||
|
||||
prompt_input = "{}{}".format(self.prompt_pre, target)
|
||||
prompt_input_ids = self.tokenizer.encode(prompt_input)
|
||||
audio_length = len(prompt_input_ids) - prompt_pre_length
|
||||
input_ids = prompt_input_ids + [self.tokenizer.pad_token_id]
|
||||
input_ids = torch.tensor(input_ids, dtype=torch.int64) # [bos, prompt, input, pad]
|
||||
input_ids[prompt_pre_length:] = -1 # [bos, prompt,-1,-1]
|
||||
attention_mask = input_ids.ge(-1) # [true, true, true, true], length mask
|
||||
|
||||
prompt_answer = "{}{}".format(self.prompt_pre, target)
|
||||
prompt_answer_ids = self.tokenizer.encode(prompt_answer)
|
||||
answer_length = len(prompt_answer_ids) - prompt_pre_length
|
||||
labels_ids = copy.deepcopy(prompt_input_ids) + [self.tokenizer.eos_token_id]
|
||||
labels_ids = torch.tensor(labels_ids, dtype=torch.int64) # [bos, prompt, input, eos]
|
||||
labels_ids[:prompt_pre_length] = -1 # [-1, -1, input, eos]
|
||||
label_mask = labels_ids.ge(0) # [False,False,True,True]
|
||||
labels_ids[~label_mask] = self.IGNORE_INDEX # [-100,-100,input,eos]
|
||||
|
||||
audio_mask = [0] * prompt_pre_length + [1] * audio_length + [0]
|
||||
audio_mask = torch.tensor(audio_mask, dtype=torch.float32)
|
||||
|
||||
ids = self.tokenizer.encode(target) # token ids is different from labels_ids
|
||||
text = torch.tensor(ids, dtype=torch.int64)
|
||||
text_lengths = torch.tensor([len(ids)], dtype=torch.int32)
|
||||
|
||||
return {"speech": speech,
|
||||
"speech_lengths": speech_lengths,
|
||||
"text": text,
|
||||
"text_lengths": text_lengths,
|
||||
"input_ids": input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"labels_ids": labels_ids,
|
||||
"label_mask": label_mask,
|
||||
"audio_mask": audio_mask,
|
||||
}
|
||||
|
||||
def collator(self, samples: list = None):
|
||||
outputs = {}
|
||||
for sample in samples:
|
||||
for key in sample.keys():
|
||||
if key not in outputs:
|
||||
outputs[key] = []
|
||||
outputs[key].append(sample[key])
|
||||
|
||||
for key, data_list in outputs.items():
|
||||
if isinstance(data_list[0], torch.Tensor):
|
||||
if data_list[0].dtype == torch.int64:
|
||||
|
||||
pad_value = self.int_pad_value
|
||||
else:
|
||||
pad_value = self.float_pad_value
|
||||
|
||||
outputs[key] = torch.nn.utils.rnn.pad_sequence(data_list, batch_first=True, padding_value=pad_value)
|
||||
return outputs
|
||||
|
||||
@@ -18,14 +18,16 @@ def download_from_ms(**kwargs):
|
||||
model_or_path = name_maps_ms[model_or_path]
|
||||
model_revision = kwargs.get("model_revision")
|
||||
if not os.path.exists(model_or_path):
|
||||
model_or_path = get_or_download_model_dir(model_or_path, model_revision, is_training=kwargs.get("is_training"), check_latest=kwargs.get("kwargs", True))
|
||||
model_or_path = get_or_download_model_dir(model_or_path, model_revision, is_training=kwargs.get("is_training"), check_latest=kwargs.get("check_latest", True))
|
||||
kwargs["model_path"] = model_or_path
|
||||
|
||||
if os.path.exists(os.path.join(model_or_path, "configuration.json")):
|
||||
with open(os.path.join(model_or_path, "configuration.json"), 'r', encoding='utf-8') as f:
|
||||
conf_json = json.load(f)
|
||||
|
||||
cfg = {}
|
||||
add_file_root_path(model_or_path, conf_json["file_path_metas"], cfg)
|
||||
if "file_path_metas" in conf_json:
|
||||
add_file_root_path(model_or_path, conf_json["file_path_metas"], cfg)
|
||||
cfg.update(kwargs)
|
||||
config = OmegaConf.load(cfg["config"])
|
||||
kwargs = OmegaConf.merge(config, cfg)
|
||||
|
||||
@@ -8,6 +8,7 @@ name_maps_ms = {
|
||||
"ct-punc-c": "damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch",
|
||||
"fa-zh": "damo/speech_timestamp_prediction-v1-16k-offline",
|
||||
"cam++": "damo/speech_campplus_sv_zh-cn_16k-common",
|
||||
"whisper-large-v2": "iic/speech_whisper-large_asr_multilingual",
|
||||
}
|
||||
|
||||
name_maps_hf = {
|
||||
|
||||
@@ -17,8 +17,9 @@ class WhisperFrontend(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
fs: int = 16000,
|
||||
whisper_model: str = "large-v3",
|
||||
whisper_model: str = None,
|
||||
do_pad_trim: bool = True,
|
||||
n_mels: int = 80,
|
||||
):
|
||||
super().__init__()
|
||||
assert fs == 16000
|
||||
@@ -30,17 +31,16 @@ class WhisperFrontend(nn.Module):
|
||||
self.pad_samples = N_SAMPLES
|
||||
self.frame_shift = self.hop_length
|
||||
self.lfr_n = 1
|
||||
self.n_mels = n_mels
|
||||
if whisper_model == "large-v3" or whisper_model == "large":
|
||||
self.n_mels = 128
|
||||
else:
|
||||
self.n_mels = 80
|
||||
|
||||
self.mel_filters = whisper.audio.mel_filters
|
||||
self.do_pad_trim = do_pad_trim
|
||||
if do_pad_trim:
|
||||
self.pad_or_trim = whisper.pad_or_trim
|
||||
|
||||
assert whisper_model in whisper.available_models()
|
||||
# assert whisper_model in whisper.available_models()
|
||||
|
||||
def output_size(self) -> int:
|
||||
return self.n_mels
|
||||
|
||||
@@ -0,0 +1,62 @@
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from funasr.register import tables
|
||||
|
||||
@tables.register("adaptor_classes", "Linear")
|
||||
class Linear(nn.Module):
|
||||
def __init__(self, downsample_rate, encoder_dim, llm_dim, ffn_dim: int = 2048, **kwargs):
|
||||
super().__init__()
|
||||
self.k = downsample_rate
|
||||
self.encoder_dim = encoder_dim
|
||||
self.llm_dim = llm_dim
|
||||
self.linear1 = nn.Linear(self.encoder_dim * self.k, ffn_dim)
|
||||
self.relu = nn.ReLU()
|
||||
self.linear2 = nn.Linear(ffn_dim, self.llm_dim)
|
||||
|
||||
def forward(self, x):
|
||||
batch_size, seq_len, dim = x.size()
|
||||
num_frames_to_discard = seq_len % self.k
|
||||
if num_frames_to_discard > 0:
|
||||
x = x[:, :-num_frames_to_discard, :]
|
||||
seq_len = x.size(1)
|
||||
|
||||
x = x.contiguous()
|
||||
x = x.view(batch_size, seq_len // self.k, dim * self.k)
|
||||
x = self.linear1(x)
|
||||
x = self.relu(x)
|
||||
x = self.linear2(x)
|
||||
return x
|
||||
|
||||
@tables.register("adaptor_classes", "QFormer")
|
||||
class EncoderProjectorQFormer(nn.Module):
|
||||
def __init__(self, downsample_rate, encoder_dim, llm_dim, ffn_dim: int = 2048, **kwargs):
|
||||
super().__init__()
|
||||
self.encoder_dim = encoder_dim
|
||||
self.llm_dim = llm_dim
|
||||
from transformers import Blip2QFormerConfig, Blip2QFormerModel
|
||||
configuration = Blip2QFormerConfig()
|
||||
configuration.encoder_hidden_size = self.encoder_dim
|
||||
configuration.num_hidden_layers = 2
|
||||
|
||||
self.query_len = 64
|
||||
self.query = nn.Parameter(torch.zeros(1, self.query_len, configuration.hidden_size))
|
||||
self.query.data.normal_(mean=0.0, std=1.0)
|
||||
self.qformer = Blip2QFormerModel(configuration)
|
||||
|
||||
self.linear = nn.Linear(configuration.hidden_size, self.llm_dim)
|
||||
self.norm = nn.LayerNorm(self.llm_dim, eps=1e-5)
|
||||
|
||||
def forward(self, x, atts):
|
||||
query = self.query.expand(x.shape[0], -1, -1)
|
||||
|
||||
query_output = self.qformer(
|
||||
query_embeds=query,
|
||||
encoder_hidden_states=x,
|
||||
encoder_attention_mask=atts,
|
||||
return_dict=True,
|
||||
)
|
||||
|
||||
query_proj = self.norm(self.linear(query_output.last_hidden_state))
|
||||
|
||||
return query_proj
|
||||
@@ -0,0 +1,341 @@
|
||||
import logging
|
||||
from typing import Union, Dict, List, Tuple, Optional
|
||||
|
||||
import time
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from torch.cuda.amp import autocast
|
||||
|
||||
from funasr.models.scama.utils import sequence_mask
|
||||
from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
|
||||
from funasr.models.ctc.ctc import CTC
|
||||
from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
|
||||
from funasr.metrics.compute_acc import th_accuracy, compute_accuracy
|
||||
# from funasr.models.e2e_asr_common import ErrorCalculator
|
||||
from funasr.train_utils.device_funcs import force_gatherable
|
||||
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
|
||||
from funasr.utils import postprocess_utils
|
||||
from funasr.utils.datadir_writer import DatadirWriter
|
||||
from funasr.register import tables
|
||||
|
||||
|
||||
@tables.register("model_classes", "LLMASR")
|
||||
class LLMASR(nn.Module):
|
||||
""" """
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
specaug: str = None,
|
||||
specaug_conf: dict = None,
|
||||
normalize: str = None,
|
||||
normalize_conf: dict = None,
|
||||
encoder: str = None,
|
||||
encoder_conf: dict = None,
|
||||
decoder: str = None,
|
||||
decoder_conf: dict = None,
|
||||
ctc: str = None,
|
||||
ctc_conf: dict = None,
|
||||
ctc_weight: float = 0.5,
|
||||
llm: str = None,
|
||||
llm_conf: dict = None,
|
||||
adaptor: str = None,
|
||||
adaptor_conf: dict = None,
|
||||
input_size: int = 80,
|
||||
vocab_size: int = -1,
|
||||
ignore_id: int = -1,
|
||||
blank_id: int = 0,
|
||||
sos: int = 1,
|
||||
eos: int = 2,
|
||||
lsm_weight: float = 0.0,
|
||||
length_normalized_loss: bool = False,
|
||||
report_cer: bool = True,
|
||||
report_wer: bool = True,
|
||||
sym_space: str = "<space>",
|
||||
sym_blank: str = "<blank>",
|
||||
# extract_feats_in_collect_stats: bool = True,
|
||||
share_embedding: bool = False,
|
||||
# preencoder: Optional[AbsPreEncoder] = None,
|
||||
# postencoder: Optional[AbsPostEncoder] = None,
|
||||
**kwargs,
|
||||
):
|
||||
|
||||
super().__init__()
|
||||
|
||||
if specaug is not None:
|
||||
specaug_class = tables.specaug_classes.get(specaug)
|
||||
specaug = specaug_class(**specaug_conf)
|
||||
if normalize is not None:
|
||||
normalize_class = tables.normalize_classes.get(normalize)
|
||||
normalize = normalize_class(**normalize_conf)
|
||||
|
||||
# audio encoder
|
||||
hub = encoder_conf.get("hub", None)
|
||||
if hub == "funasr":
|
||||
from funasr import AutoModel
|
||||
init_param_path = encoder_conf.get("init_param_path", "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch")
|
||||
model = AutoModel(model=init_param_path, model_revision="v2.0.4")
|
||||
# frontend = model.kwargs.get("frontend")
|
||||
model.model.decoder = None
|
||||
|
||||
self.audio_encoder = model.model
|
||||
# self.frontend = frontend
|
||||
|
||||
elif hub == "hf":
|
||||
pass
|
||||
else:
|
||||
encoder_class = tables.encoder_classes.get(encoder)
|
||||
encoder = encoder_class(input_size=input_size, **encoder_conf)
|
||||
encoder_output_size = encoder.output_size()
|
||||
|
||||
# llm
|
||||
hub = llm_conf.get("hub", "hf")
|
||||
self.llm = None
|
||||
if hub == "hf":
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
|
||||
|
||||
init_param_path = llm_conf.get("init_param_path", "vicuna-7b-v1.5")
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
init_param_path,
|
||||
load_in_8bit=None,
|
||||
device_map=None,
|
||||
use_cache=None,
|
||||
)
|
||||
freeze = llm_conf.get("freeze", True)
|
||||
if freeze:
|
||||
for name, param in model.named_parameters():
|
||||
param.requires_grad = False
|
||||
model.eval()
|
||||
self.llm = model
|
||||
|
||||
# adaptor
|
||||
adaptor_class = tables.adaptor_classes.get(adaptor)
|
||||
adaptor = adaptor_class(**adaptor_conf)
|
||||
|
||||
self.adaptor = adaptor
|
||||
|
||||
|
||||
self.blank_id = blank_id
|
||||
self.sos = sos if sos is not None else vocab_size - 1
|
||||
self.eos = eos if eos is not None else vocab_size - 1
|
||||
self.vocab_size = vocab_size
|
||||
self.ignore_id = ignore_id
|
||||
self.specaug = specaug
|
||||
self.normalize = normalize
|
||||
self.encoder = encoder
|
||||
|
||||
|
||||
self.criterion_att = LabelSmoothingLoss(
|
||||
size=vocab_size,
|
||||
padding_idx=ignore_id,
|
||||
smoothing=lsm_weight,
|
||||
normalize_length=length_normalized_loss,
|
||||
)
|
||||
#
|
||||
# if report_cer or report_wer:
|
||||
# self.error_calculator = ErrorCalculator(
|
||||
# token_list, sym_space, sym_blank, report_cer, report_wer
|
||||
# )
|
||||
#
|
||||
self.error_calculator = None
|
||||
|
||||
self.length_normalized_loss = length_normalized_loss
|
||||
self.beam_search = None
|
||||
|
||||
def forward(
|
||||
self,
|
||||
speech: torch.Tensor,
|
||||
speech_lengths: torch.Tensor,
|
||||
text: torch.Tensor,
|
||||
text_lengths: torch.Tensor,
|
||||
input_ids: torch.Tensor,
|
||||
attention_mask:torch.Tensor,
|
||||
labels_ids: torch.Tensor,
|
||||
label_mask: torch.Tensor,
|
||||
audio_mask: torch.Tensor,
|
||||
**kwargs,
|
||||
) -> Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]:
|
||||
"""Encoder + Decoder + Calc loss
|
||||
Args:
|
||||
speech: (Batch, Length, ...)
|
||||
speech_lengths: (Batch, )
|
||||
text: (Batch, Length)
|
||||
text_lengths: (Batch,)
|
||||
"""
|
||||
# import pdb;
|
||||
# pdb.set_trace()
|
||||
if len(text_lengths.size()) > 1:
|
||||
text_lengths = text_lengths[:, 0]
|
||||
if len(speech_lengths.size()) > 1:
|
||||
speech_lengths = speech_lengths[:, 0]
|
||||
|
||||
batch_size = speech.shape[0]
|
||||
|
||||
# audio encoder
|
||||
encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, audio_mask=audio_mask)
|
||||
|
||||
# adaptor
|
||||
encoder_out = self.adaptor(encoder_out)
|
||||
|
||||
if input_ids is not None:
|
||||
input_ids[input_ids == -1] = 0
|
||||
input_ids[input_ids == -100] = 0
|
||||
if hasattr(self.llm.model, "embed_tokens"):
|
||||
inputs_embeds = self.llm.model.embed_tokens(input_ids)
|
||||
elif hasattr(self.llm.model.model, "embed_tokens"):
|
||||
inputs_embeds = self.llm.model.model.embed_tokens(input_ids)
|
||||
else:
|
||||
inputs_embeds = self.llm.model.model.model.embed_tokens(input_ids)
|
||||
|
||||
if audio_mask is not None:
|
||||
batch_size, token_num, dims = inputs_embeds.shape
|
||||
_, l, _ = encoder_out.shape
|
||||
encoder_outs_pad = F.pad(encoder_out, (0, 0, token_num-l-1, 1, 0, 0), value=0.0)
|
||||
inputs_embeds = encoder_outs_pad * audio_mask[:, :, None] + inputs_embeds * (1.0-audio_mask[:, :, None])
|
||||
inputs_embeds = F.pad(inputs_embeds[:, 1:, :], (0, 0, 0, 1, 0, 0), value=0.0)
|
||||
|
||||
model_outputs = self.llm(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids)
|
||||
loss = model_outputs.loss
|
||||
|
||||
|
||||
stats = {}
|
||||
with torch.no_grad():
|
||||
preds = torch.argmax(model_outputs.logits, -1)
|
||||
acc_att = compute_accuracy(preds[:, :-1], labels_ids[:, 1:], ignore_label=-100)
|
||||
stats["acc"] = acc_att
|
||||
|
||||
stats["loss"] = torch.clone(loss.detach())
|
||||
|
||||
# force_gatherable: to-device and to-tensor if scalar for DataParallel
|
||||
if self.length_normalized_loss:
|
||||
batch_size = int((text_lengths + 1).sum())
|
||||
loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
|
||||
return loss, stats, weight
|
||||
|
||||
def encode(
|
||||
self, speech: torch.Tensor, speech_lengths: torch.Tensor, **kwargs,
|
||||
) -> Tuple[torch.Tensor, torch.Tensor]:
|
||||
|
||||
audio_mask = kwargs.get("audio_mask", None)
|
||||
audio_token_lengths = audio_mask.sum(-1) if audio_mask is not None else None
|
||||
|
||||
batch = {"speech": speech, "speech_lengths": speech_lengths}
|
||||
enc, enc_lens = self.audio_encoder.encode(**batch)
|
||||
with autocast(False):
|
||||
enc_mask = sequence_mask(enc_lens, enc.size(1), device=enc.device)[:, None, :]
|
||||
pre_acoustic_embeds, pre_token_length, _, _ = self.audio_encoder.predictor(enc,
|
||||
mask=enc_mask,
|
||||
target_label_length=audio_token_lengths,
|
||||
)
|
||||
|
||||
return pre_acoustic_embeds, pre_token_length
|
||||
|
||||
|
||||
def inference(self,
|
||||
data_in,
|
||||
data_lengths=None,
|
||||
key: list = None,
|
||||
tokenizer=None,
|
||||
frontend=None,
|
||||
**kwargs,
|
||||
):
|
||||
|
||||
prompt = kwargs.get("prompt", "Transcribe speech to text.")
|
||||
|
||||
if kwargs.get("batch_size", 1) > 1:
|
||||
raise NotImplementedError("batch decoding is not implemented")
|
||||
|
||||
|
||||
|
||||
meta_data = {}
|
||||
if isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank": # fbank
|
||||
speech, speech_lengths = data_in, data_lengths
|
||||
if len(speech.shape) < 3:
|
||||
speech = speech[None, :, :]
|
||||
if speech_lengths is None:
|
||||
speech_lengths = speech.shape[1]
|
||||
else:
|
||||
# extract fbank feats
|
||||
time1 = time.perf_counter()
|
||||
audio_sample_list = load_audio_text_image_video(data_in, fs=frontend.fs, audio_fs=kwargs.get("fs", 16000),
|
||||
data_type=kwargs.get("data_type", "sound"),
|
||||
tokenizer=tokenizer)
|
||||
time2 = time.perf_counter()
|
||||
meta_data["load_data"] = f"{time2 - time1:0.3f}"
|
||||
speech, speech_lengths = extract_fbank(audio_sample_list, data_type=kwargs.get("data_type", "sound"),
|
||||
frontend=frontend)
|
||||
time3 = time.perf_counter()
|
||||
meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
|
||||
meta_data["batch_data_time"] = speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
|
||||
|
||||
speech = speech.to(device=kwargs["device"])
|
||||
speech_lengths = speech_lengths.to(device=kwargs["device"])
|
||||
|
||||
# Encoder
|
||||
encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
|
||||
|
||||
# adaptor
|
||||
encoder_out = self.adaptor(encoder_out)
|
||||
|
||||
|
||||
prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(prompt)
|
||||
prompt_ids = tokenizer.encode(prompt_pre)
|
||||
prompt_length = len(prompt_ids)
|
||||
prompt_ids = torch.tensor(prompt_ids, dtype=torch.int64).to(kwargs["device"])
|
||||
|
||||
|
||||
if hasattr(self.llm.model, "embed_tokens"):
|
||||
inputs_embeds = self.llm.model.embed_tokens(prompt_ids)
|
||||
elif hasattr(self.llm.model.model, "embed_tokens"):
|
||||
inputs_embeds = self.llm.model.model.embed_tokens(prompt_ids)
|
||||
else:
|
||||
inputs_embeds = self.llm.model.model.model.embed_tokens(prompt_ids)
|
||||
|
||||
inputs_embeds = torch.cat((inputs_embeds[None, :, :], encoder_out), dim=1) # [prompt, audio]
|
||||
attention_mask = torch.ones(inputs_embeds.size()[:-1], dtype=torch.long).to(kwargs["device"])
|
||||
|
||||
# model_outputs = self.llm.generate(
|
||||
# inputs_embeds=inputs_embeds,
|
||||
# max_length=kwargs.get("max_length", 200),
|
||||
# max_new_tokens=kwargs.get("max_new_tokens", 200),
|
||||
# num_beams=kwargs.get("num_beams", 4),
|
||||
# do_sample=kwargs.get("do_sample", False),
|
||||
# min_length=kwargs.get("min_length", 1),
|
||||
# top_p=kwargs.get("top_p", 1.0),
|
||||
# repetition_penalty=kwargs.get("repetition_penalty", 1.0),
|
||||
# length_penalty=kwargs.get("length_penalty", 1.0),
|
||||
# temperature=kwargs.get("temperature", 1.0),
|
||||
# attention_mask=attention_mask,
|
||||
# bos_token_id=tokenizer.bos_token_id,
|
||||
# eos_token_id=tokenizer.eos_token_id,
|
||||
# pad_token_id=tokenizer.pad_token_id
|
||||
# )
|
||||
|
||||
|
||||
model_outputs = self.llm(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=None)
|
||||
preds = torch.argmax(model_outputs.logits, -1)
|
||||
text = tokenizer.batch_decode(preds, add_special_tokens=False, skip_special_tokens=True)
|
||||
|
||||
text = text[0].split(': ')[-1]
|
||||
text = text.strip()
|
||||
|
||||
# preds = torch.argmax(model_outputs.logits, -1)
|
||||
|
||||
ibest_writer = None
|
||||
if kwargs.get("output_dir") is not None:
|
||||
if not hasattr(self, "writer"):
|
||||
self.writer = DatadirWriter(kwargs.get("output_dir"))
|
||||
ibest_writer = self.writer[f"{0 + 1}best_recog"]
|
||||
|
||||
results = []
|
||||
result_i = {"key": key[0], "text": text}
|
||||
results.append(result_i)
|
||||
|
||||
if ibest_writer is not None:
|
||||
ibest_writer["text"][key[0]] = text
|
||||
|
||||
|
||||
|
||||
|
||||
return results, meta_data
|
||||
|
||||
+64
-252
@@ -1,273 +1,85 @@
|
||||
from dataclasses import dataclass
|
||||
from typing import Dict
|
||||
from typing import Iterable, Optional
|
||||
|
||||
import time
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import Tensor
|
||||
from torch import nn
|
||||
import whisper
|
||||
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
|
||||
|
||||
|
||||
from funasr.models.whisper.utils.decoding import detect_language as detect_language_function, decode as decode_function
|
||||
from funasr.register import tables
|
||||
|
||||
|
||||
@dataclass
|
||||
class ModelDimensions:
|
||||
n_mels: int
|
||||
n_audio_ctx: int
|
||||
n_audio_state: int
|
||||
n_audio_head: int
|
||||
n_audio_layer: int
|
||||
n_vocab: int
|
||||
n_text_ctx: int
|
||||
n_text_state: int
|
||||
n_text_head: int
|
||||
n_text_layer: int
|
||||
|
||||
|
||||
class LayerNorm(nn.LayerNorm):
|
||||
def forward(self, x: Tensor) -> Tensor:
|
||||
return super().forward(x.float()).type(x.dtype)
|
||||
|
||||
|
||||
class Linear(nn.Linear):
|
||||
def forward(self, x: Tensor) -> Tensor:
|
||||
return F.linear(
|
||||
x, self.weight.to(x.dtype), None if self.bias is None else self.bias.to(x.dtype)
|
||||
)
|
||||
|
||||
|
||||
class Conv1d(nn.Conv1d):
|
||||
def _conv_forward(self, x: Tensor, weight: Tensor, bias: Optional[Tensor]) -> Tensor:
|
||||
return super()._conv_forward(
|
||||
x, weight.to(x.dtype), None if bias is None else bias.to(x.dtype)
|
||||
)
|
||||
|
||||
|
||||
def sinusoids(length, channels, max_timescale=10000):
|
||||
"""Returns sinusoids for positional embedding"""
|
||||
assert channels % 2 == 0
|
||||
log_timescale_increment = np.log(max_timescale) / (channels // 2 - 1)
|
||||
inv_timescales = torch.exp(-log_timescale_increment * torch.arange(channels // 2))
|
||||
scaled_time = torch.arange(length)[:, np.newaxis] * inv_timescales[np.newaxis, :]
|
||||
return torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1)
|
||||
|
||||
|
||||
class MultiHeadAttention(nn.Module):
|
||||
def __init__(self, n_state: int, n_head: int):
|
||||
@tables.register("model_classes", "WhisperWarp")
|
||||
class WhisperWarp(nn.Module):
|
||||
def __init__(self, whisper_dims: dict, **kwargs):
|
||||
super().__init__()
|
||||
self.n_head = n_head
|
||||
self.query = Linear(n_state, n_state)
|
||||
self.key = Linear(n_state, n_state, bias=False)
|
||||
self.value = Linear(n_state, n_state)
|
||||
self.out = Linear(n_state, n_state)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
x: Tensor,
|
||||
xa: Optional[Tensor] = None,
|
||||
mask: Optional[Tensor] = None,
|
||||
kv_cache: Optional[dict] = None,
|
||||
):
|
||||
q = self.query(x)
|
||||
|
||||
if kv_cache is None or xa is None or self.key not in kv_cache:
|
||||
# hooks, if installed (i.e. kv_cache is not None), will prepend the cached kv tensors;
|
||||
# otherwise, perform key/value projections for self- or cross-attention as usual.
|
||||
k = self.key(x if xa is None else xa)
|
||||
v = self.value(x if xa is None else xa)
|
||||
hub = kwargs.get("hub", "funasr")
|
||||
if hub == "openai":
|
||||
init_param_path = kwargs.get("init_param_path", "large-v3")
|
||||
model = whisper.load_model(init_param_path)
|
||||
else:
|
||||
# for cross-attention, calculate keys and values once and reuse in subsequent calls.
|
||||
k = kv_cache[self.key]
|
||||
v = kv_cache[self.value]
|
||||
dims = whisper.model.ModelDimensions(**whisper_dims)
|
||||
model = whisper.model.Whisper(dims=dims)
|
||||
|
||||
self.model = model
|
||||
|
||||
def forward(self, ):
|
||||
pass
|
||||
|
||||
def inference(self,
|
||||
data_in,
|
||||
data_lengths=None,
|
||||
key: list = None,
|
||||
tokenizer=None,
|
||||
frontend=None,
|
||||
**kwargs,
|
||||
):
|
||||
if kwargs.get("batch_size", 1) > 1:
|
||||
raise NotImplementedError("batch decoding is not implemented")
|
||||
|
||||
wv, qk = self.qkv_attention(q, k, v, mask)
|
||||
return self.out(wv), qk
|
||||
meta_data = {}
|
||||
if isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank": # fbank
|
||||
speech, speech_lengths = data_in, data_lengths
|
||||
if len(speech.shape) < 3:
|
||||
speech = speech[None, :, :]
|
||||
if speech_lengths is None:
|
||||
speech_lengths = speech.shape[1]
|
||||
else:
|
||||
# extract fbank feats
|
||||
time1 = time.perf_counter()
|
||||
audio_sample_list = load_audio_text_image_video(data_in, fs=frontend.fs, audio_fs=kwargs.get("fs", 16000),
|
||||
data_type=kwargs.get("data_type", "sound"),
|
||||
tokenizer=tokenizer)
|
||||
time2 = time.perf_counter()
|
||||
meta_data["load_data"] = f"{time2 - time1:0.3f}"
|
||||
speech, speech_lengths = extract_fbank(audio_sample_list, data_type=kwargs.get("data_type", "sound"),
|
||||
frontend=frontend)
|
||||
time3 = time.perf_counter()
|
||||
meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
|
||||
frame_shift = frontend.frame_shift if hasattr(frontend, "frame_shift") else 10
|
||||
lfr_n = frontend.lfr_n if hasattr(frontend, "lfr_n") else 1
|
||||
meta_data["batch_data_time"] = speech_lengths.sum().item() * frame_shift * lfr_n / 1000
|
||||
|
||||
def qkv_attention(self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None):
|
||||
n_batch, n_ctx, n_state = q.shape
|
||||
scale = (n_state // self.n_head) ** -0.25
|
||||
q = q.view(*q.shape[:2], self.n_head, -1).permute(0, 2, 1, 3) * scale
|
||||
k = k.view(*k.shape[:2], self.n_head, -1).permute(0, 2, 3, 1) * scale
|
||||
v = v.view(*v.shape[:2], self.n_head, -1).permute(0, 2, 1, 3)
|
||||
speech = speech.to(device=kwargs["device"])[0, :, :]
|
||||
speech_lengths = speech_lengths.to(device=kwargs["device"])
|
||||
|
||||
qk = q @ k
|
||||
if mask is not None:
|
||||
qk = qk + mask[:n_ctx, :n_ctx]
|
||||
qk = qk.float()
|
||||
# detect the spoken language
|
||||
_, probs = self.model.detect_language(speech)
|
||||
print(f"Detected language: {max(probs, key=probs.get)}")
|
||||
|
||||
w = F.softmax(qk, dim=-1).to(q.dtype)
|
||||
return (w @ v).permute(0, 2, 1, 3).flatten(start_dim=2), qk.detach()
|
||||
# decode the audio
|
||||
options = whisper.DecodingOptions(language=kwargs.get("language", None), fp16=False)
|
||||
result = whisper.decode(self.model, speech, options)
|
||||
|
||||
results = []
|
||||
result_i = {"key": key[0], "text": result.text}
|
||||
|
||||
class ResidualAttentionBlock(nn.Module):
|
||||
def __init__(self, n_state: int, n_head: int, cross_attention: bool = False):
|
||||
super().__init__()
|
||||
|
||||
self.attn = MultiHeadAttention(n_state, n_head)
|
||||
self.attn_ln = LayerNorm(n_state)
|
||||
|
||||
self.cross_attn = MultiHeadAttention(n_state, n_head) if cross_attention else None
|
||||
self.cross_attn_ln = LayerNorm(n_state) if cross_attention else None
|
||||
|
||||
n_mlp = n_state * 4
|
||||
self.mlp = nn.Sequential(Linear(n_state, n_mlp), nn.GELU(), Linear(n_mlp, n_state))
|
||||
self.mlp_ln = LayerNorm(n_state)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
x: Tensor,
|
||||
xa: Optional[Tensor] = None,
|
||||
mask: Optional[Tensor] = None,
|
||||
kv_cache: Optional[dict] = None,
|
||||
):
|
||||
x = x + self.attn(self.attn_ln(x), mask=mask, kv_cache=kv_cache)[0]
|
||||
if self.cross_attn:
|
||||
x = x + self.cross_attn(self.cross_attn_ln(x), xa, kv_cache=kv_cache)[0]
|
||||
x = x + self.mlp(self.mlp_ln(x))
|
||||
return x
|
||||
|
||||
|
||||
|
||||
@tables.register("encoder_classes", "WhisperEncoder")
|
||||
class AudioEncoder(nn.Module):
|
||||
def __init__(self, n_mels: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
|
||||
super().__init__()
|
||||
self.conv1 = Conv1d(n_mels, n_state, kernel_size=3, padding=1)
|
||||
self.conv2 = Conv1d(n_state, n_state, kernel_size=3, stride=2, padding=1)
|
||||
self.register_buffer("positional_embedding", sinusoids(n_ctx, n_state))
|
||||
|
||||
self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList(
|
||||
[ResidualAttentionBlock(n_state, n_head) for _ in range(n_layer)]
|
||||
)
|
||||
self.ln_post = LayerNorm(n_state)
|
||||
|
||||
def forward(self, x: Tensor):
|
||||
"""
|
||||
x : torch.Tensor, shape = (batch_size, n_mels, n_ctx)
|
||||
the mel spectrogram of the audio
|
||||
"""
|
||||
x = F.gelu(self.conv1(x))
|
||||
x = F.gelu(self.conv2(x))
|
||||
x = x.permute(0, 2, 1)
|
||||
|
||||
assert x.shape[1:] == self.positional_embedding.shape, "incorrect audio shape"
|
||||
x = (x + self.positional_embedding).to(x.dtype)
|
||||
|
||||
for block in self.blocks:
|
||||
x = block(x)
|
||||
|
||||
x = self.ln_post(x)
|
||||
return x
|
||||
|
||||
@tables.register("decoder_classes", "WhisperDecoder")
|
||||
class TextDecoder(nn.Module):
|
||||
def __init__(self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
|
||||
super().__init__()
|
||||
|
||||
self.token_embedding = nn.Embedding(n_vocab, n_state)
|
||||
self.positional_embedding = nn.Parameter(torch.empty(n_ctx, n_state))
|
||||
|
||||
self.blocks: Iterable[ResidualAttentionBlock] = nn.ModuleList(
|
||||
[ResidualAttentionBlock(n_state, n_head, cross_attention=True) for _ in range(n_layer)]
|
||||
)
|
||||
self.ln = LayerNorm(n_state)
|
||||
|
||||
mask = torch.empty(n_ctx, n_ctx).fill_(-np.inf).triu_(1)
|
||||
self.register_buffer("mask", mask, persistent=False)
|
||||
|
||||
def forward(self, x: Tensor, xa: Tensor, kv_cache: Optional[dict] = None):
|
||||
"""
|
||||
x : torch.LongTensor, shape = (batch_size, <= n_ctx)
|
||||
the text tokens
|
||||
xa : torch.Tensor, shape = (batch_size, n_mels, n_audio_ctx)
|
||||
the encoded audio features to be attended on
|
||||
"""
|
||||
offset = next(iter(kv_cache.values())).shape[1] if kv_cache else 0
|
||||
x = self.token_embedding(x) + self.positional_embedding[offset : offset + x.shape[-1]]
|
||||
x = x.to(xa.dtype)
|
||||
|
||||
for block in self.blocks:
|
||||
x = block(x, xa, mask=self.mask, kv_cache=kv_cache)
|
||||
|
||||
x = self.ln(x)
|
||||
logits = (x @ torch.transpose(self.token_embedding.weight.to(x.dtype), 0, 1)).float()
|
||||
|
||||
return logits
|
||||
|
||||
@tables.register("model_classes", "Whisper")
|
||||
class Whisper(nn.Module):
|
||||
def __init__(self, dims: dict):
|
||||
super().__init__()
|
||||
dims = ModelDimensions(**dims)
|
||||
self.dims = dims
|
||||
self.sos = 1
|
||||
self.eos = 1
|
||||
self.encoder = AudioEncoder(
|
||||
self.dims.n_mels,
|
||||
self.dims.n_audio_ctx,
|
||||
self.dims.n_audio_state,
|
||||
self.dims.n_audio_head,
|
||||
self.dims.n_audio_layer,
|
||||
)
|
||||
self.decoder = TextDecoder(
|
||||
self.dims.n_vocab,
|
||||
self.dims.n_text_ctx,
|
||||
self.dims.n_text_state,
|
||||
self.dims.n_text_head,
|
||||
self.dims.n_text_layer,
|
||||
)
|
||||
|
||||
def embed_audio(self, mel: torch.Tensor):
|
||||
return self.encoder(mel)
|
||||
|
||||
def logits(self, tokens: torch.Tensor, audio_features: torch.Tensor):
|
||||
return self.decoder(tokens, audio_features)
|
||||
|
||||
def forward(self, mel: torch.Tensor, tokens: torch.Tensor) -> Dict[str, torch.Tensor]:
|
||||
return self.decoder(tokens, self.encoder(mel))
|
||||
|
||||
@property
|
||||
def device(self):
|
||||
return next(self.parameters()).device
|
||||
|
||||
@property
|
||||
def is_multilingual(self):
|
||||
return self.dims.n_vocab == 51865
|
||||
|
||||
def install_kv_cache_hooks(self, cache: Optional[dict] = None):
|
||||
"""
|
||||
The `MultiHeadAttention` module optionally accepts `kv_cache` which stores the key and value
|
||||
tensors calculated for the previous positions. This method returns a dictionary that stores
|
||||
all caches, and the necessary hooks for the key and value projection modules that save the
|
||||
intermediate tensors to be reused during later calculations.
|
||||
|
||||
Returns
|
||||
-------
|
||||
cache : Dict[nn.Module, torch.Tensor]
|
||||
A dictionary object mapping the key/value projection modules to its cache
|
||||
hooks : List[RemovableHandle]
|
||||
List of PyTorch RemovableHandle objects to stop the hooks to be called
|
||||
"""
|
||||
cache = {**cache} if cache is not None else {}
|
||||
hooks = []
|
||||
|
||||
def save_to_cache(module, _, output):
|
||||
if module not in cache or output.shape[1] > self.decoder.positional_embedding.shape[0]:
|
||||
cache[module] = output # save as-is, for the first token or cross attention
|
||||
else:
|
||||
cache[module] = torch.cat([cache[module], output], dim=1).detach()
|
||||
return cache[module]
|
||||
|
||||
def install_hooks(layer: nn.Module):
|
||||
if isinstance(layer, MultiHeadAttention):
|
||||
hooks.append(layer.key.register_forward_hook(save_to_cache))
|
||||
hooks.append(layer.value.register_forward_hook(save_to_cache))
|
||||
|
||||
self.decoder.apply(install_hooks)
|
||||
return cache, hooks
|
||||
|
||||
detect_language = detect_language_function
|
||||
decode = decode_function
|
||||
results.append(result_i)
|
||||
|
||||
return results, meta_data
|
||||
|
||||
@@ -0,0 +1,46 @@
|
||||
# This is an example that demonstrates how to configure a model file.
|
||||
# You can modify the configuration according to your own requirements.
|
||||
|
||||
# to print the register_table:
|
||||
# from funasr.register import tables
|
||||
# tables.print()
|
||||
|
||||
# network architecture
|
||||
model: WhisperWarp
|
||||
model_conf:
|
||||
lsm_weight: 0.1
|
||||
length_normalized_loss: true
|
||||
hub: funasr # openai
|
||||
init_param_path: null # large-v2 or large-v3 if hub == "openai"
|
||||
|
||||
|
||||
|
||||
# only use for hub == funasr,
|
||||
# if hub == openai, whisper_dims is automaticall download
|
||||
whisper_dims:
|
||||
'n_mels': 80
|
||||
'n_vocab': 51865
|
||||
'n_audio_ctx': 1500
|
||||
'n_audio_state': 1280
|
||||
'n_audio_head': 20
|
||||
'n_audio_layer': 32
|
||||
'n_text_ctx': 448
|
||||
'n_text_state': 1280
|
||||
'n_text_head': 20
|
||||
'n_text_layer': 32
|
||||
|
||||
# frontend related
|
||||
frontend: WhisperFrontend
|
||||
frontend_conf:
|
||||
fs: 16000
|
||||
n_mels: 80
|
||||
do_pad_trim: true
|
||||
|
||||
tokenizer: WhisperTokenizer
|
||||
tokenizer_conf:
|
||||
language: null
|
||||
task: transcribe
|
||||
is_multilingual: true
|
||||
num_languages: 99
|
||||
|
||||
scope_map: ['none', "model."]
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1 +0,0 @@
|
||||
{"bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "unk_token": "<|endoftext|>"}
|
||||
@@ -1 +0,0 @@
|
||||
{"unk_token": "<|endoftext|>", "bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "add_prefix_space": false, "model_max_length": 1024, "special_tokens_map_file": null, "name_or_path": "gpt2", "tokenizer_class": "GPT2Tokenizer"}
|
||||
File diff suppressed because one or more lines are too long
Binary file not shown.
@@ -1 +0,0 @@
|
||||
{"<|endoftext|>": 50257}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1 +0,0 @@
|
||||
{"bos_token": "<|endoftext|>", "eos_token": "<|endoftext|>", "unk_token": "<|endoftext|>"}
|
||||
@@ -1 +0,0 @@
|
||||
{"unk_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "bos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "eos_token": {"content": "<|endoftext|>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "add_prefix_space": false, "model_max_length": 1024, "special_tokens_map_file": null, "name_or_path": "multilingual", "errors": "replace", "tokenizer_class": "GPT2Tokenizer"}
|
||||
File diff suppressed because one or more lines are too long
@@ -1,127 +0,0 @@
|
||||
import os
|
||||
from functools import lru_cache
|
||||
from typing import Union
|
||||
try:
|
||||
import ffmpeg
|
||||
except:
|
||||
print("Please Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.")
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
|
||||
from funasr.models.whisper.utils.utils import exact_div
|
||||
|
||||
# hard-coded audio hyperparameters
|
||||
SAMPLE_RATE = 16000
|
||||
N_FFT = 400
|
||||
N_MELS = 80
|
||||
HOP_LENGTH = 160
|
||||
CHUNK_LENGTH = 30
|
||||
N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE # 480000: number of samples in a chunk
|
||||
N_FRAMES = exact_div(N_SAMPLES, HOP_LENGTH) # 3000: number of frames in a mel spectrogram input
|
||||
|
||||
|
||||
def load_audio(file: str, sr: int = SAMPLE_RATE):
|
||||
"""
|
||||
Open an audio file and read as mono waveform, resampling as necessary
|
||||
|
||||
Parameters
|
||||
----------
|
||||
file: str
|
||||
The audio file to open
|
||||
|
||||
sr: int
|
||||
The sample rate to resample the audio if necessary
|
||||
|
||||
Returns
|
||||
-------
|
||||
A NumPy array containing the audio waveform, in float32 dtype.
|
||||
"""
|
||||
try:
|
||||
# This launches a subprocess to decode audio while down-mixing and resampling as necessary.
|
||||
# Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.
|
||||
out, _ = (
|
||||
ffmpeg.input(file, threads=0)
|
||||
.output("-", format="s16le", acodec="pcm_s16le", ac=1, ar=sr)
|
||||
.run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True)
|
||||
)
|
||||
except ffmpeg.Error as e:
|
||||
raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e
|
||||
|
||||
return np.frombuffer(out, np.int16).flatten().astype(np.float32) / 32768.0
|
||||
|
||||
|
||||
def pad_or_trim(array, length: int = N_SAMPLES, *, axis: int = -1):
|
||||
"""
|
||||
Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
|
||||
"""
|
||||
if torch.is_tensor(array):
|
||||
if array.shape[axis] > length:
|
||||
array = array.index_select(dim=axis, index=torch.arange(length, device=array.device))
|
||||
|
||||
if array.shape[axis] < length:
|
||||
pad_widths = [(0, 0)] * array.ndim
|
||||
pad_widths[axis] = (0, length - array.shape[axis])
|
||||
array = F.pad(array, [pad for sizes in pad_widths[::-1] for pad in sizes])
|
||||
else:
|
||||
if array.shape[axis] > length:
|
||||
array = array.take(indices=range(length), axis=axis)
|
||||
|
||||
if array.shape[axis] < length:
|
||||
pad_widths = [(0, 0)] * array.ndim
|
||||
pad_widths[axis] = (0, length - array.shape[axis])
|
||||
array = np.pad(array, pad_widths)
|
||||
|
||||
return array
|
||||
|
||||
|
||||
@lru_cache(maxsize=None)
|
||||
def mel_filters(device, n_mels: int = N_MELS) -> torch.Tensor:
|
||||
"""
|
||||
load the mel filterbank matrix for projecting STFT into a Mel spectrogram.
|
||||
Allows decoupling librosa dependency; saved using:
|
||||
|
||||
np.savez_compressed(
|
||||
"mel_filters.npz",
|
||||
mel_80=librosa.filters.mel(sr=16000, n_fft=400, n_mels=80),
|
||||
)
|
||||
"""
|
||||
assert n_mels == 80, f"Unsupported n_mels: {n_mels}"
|
||||
with np.load(os.path.join(os.path.dirname(__file__), "assets", "mel_filters.npz")) as f:
|
||||
return torch.from_numpy(f[f"mel_{n_mels}"]).to(device)
|
||||
|
||||
|
||||
def log_mel_spectrogram(audio: Union[str, np.ndarray, torch.Tensor], n_mels: int = N_MELS):
|
||||
"""
|
||||
Compute the log-Mel spectrogram of
|
||||
|
||||
Parameters
|
||||
----------
|
||||
audio: Union[str, np.ndarray, torch.Tensor], shape = (*)
|
||||
The path to audio or either a NumPy array or Tensor containing the audio waveform in 16 kHz
|
||||
|
||||
n_mels: int
|
||||
The number of Mel-frequency filters, only 80 is supported
|
||||
|
||||
Returns
|
||||
-------
|
||||
torch.Tensor, shape = (80, n_frames)
|
||||
A Tensor that contains the Mel spectrogram
|
||||
"""
|
||||
if not torch.is_tensor(audio):
|
||||
if isinstance(audio, str):
|
||||
audio = load_audio(audio)
|
||||
audio = torch.from_numpy(audio)
|
||||
|
||||
window = torch.hann_window(N_FFT).to(audio.device)
|
||||
stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=True)
|
||||
magnitudes = stft[..., :-1].abs() ** 2
|
||||
|
||||
filters = mel_filters(audio.device, n_mels)
|
||||
mel_spec = filters @ magnitudes
|
||||
|
||||
log_spec = torch.clamp(mel_spec, min=1e-10).log10()
|
||||
log_spec = torch.maximum(log_spec, log_spec.max() - 8.0)
|
||||
log_spec = (log_spec + 4.0) / 4.0
|
||||
return log_spec
|
||||
@@ -1,710 +0,0 @@
|
||||
from dataclasses import dataclass, field
|
||||
from typing import Dict, List, Tuple, Iterable, Optional, Sequence, Union, TYPE_CHECKING
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch import Tensor
|
||||
from torch.distributions import Categorical
|
||||
|
||||
from funasr.models.whisper.utils.audio import CHUNK_LENGTH
|
||||
from funasr.models.whisper.utils.tokenizer import Tokenizer, get_tokenizer
|
||||
from funasr.models.whisper.utils.utils import compression_ratio
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from funasr.models.whisper_models.model import Whisper
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def detect_language(model: "Whisper", mel: Tensor, tokenizer: Tokenizer = None) -> Tuple[Tensor, List[dict]]:
|
||||
"""
|
||||
Detect the spoken language in the audio, and return them as list of strings, along with the ids
|
||||
of the most probable language tokens and the probability distribution over all language tokens.
|
||||
This is performed outside the main decode loop in order to not interfere with kv-caching.
|
||||
|
||||
Returns
|
||||
-------
|
||||
language_tokens : Tensor, shape = (n_audio,)
|
||||
ids of the most probable language tokens, which appears after the startoftranscript token.
|
||||
language_probs : List[Dict[str, float]], length = n_audio
|
||||
list of dictionaries containing the probability distribution over all languages.
|
||||
"""
|
||||
if tokenizer is None:
|
||||
tokenizer = get_tokenizer(model.is_multilingual)
|
||||
if tokenizer.language is None or tokenizer.language_token not in tokenizer.sot_sequence:
|
||||
raise ValueError(f"This model doesn't have language tokens so it can't perform lang id")
|
||||
|
||||
single = mel.ndim == 2
|
||||
if single:
|
||||
mel = mel.unsqueeze(0)
|
||||
|
||||
# skip encoder forward pass if already-encoded audio features were given
|
||||
if mel.shape[-2:] != (model.dims.n_audio_ctx, model.dims.n_audio_state):
|
||||
mel = model.encoder(mel)
|
||||
|
||||
# forward pass using a single token, startoftranscript
|
||||
n_audio = mel.shape[0]
|
||||
x = torch.tensor([[tokenizer.sot]] * n_audio).to(mel.device) # [n_audio, 1]
|
||||
logits = model.logits(x, mel)[:, 0]
|
||||
|
||||
# collect detected languages; suppress all non-language tokens
|
||||
mask = torch.ones(logits.shape[-1], dtype=torch.bool)
|
||||
mask[list(tokenizer.all_language_tokens)] = False
|
||||
logits[:, mask] = -np.inf
|
||||
language_tokens = logits.argmax(dim=-1)
|
||||
language_token_probs = logits.softmax(dim=-1).cpu()
|
||||
language_probs = [
|
||||
{
|
||||
c: language_token_probs[i, j].item()
|
||||
for j, c in zip(tokenizer.all_language_tokens, tokenizer.all_language_codes)
|
||||
}
|
||||
for i in range(n_audio)
|
||||
]
|
||||
|
||||
if single:
|
||||
language_tokens = language_tokens[0]
|
||||
language_probs = language_probs[0]
|
||||
|
||||
return language_tokens, language_probs
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class DecodingOptions:
|
||||
task: str = "transcribe" # whether to perform X->X "transcribe" or X->English "translate"
|
||||
language: Optional[str] = None # language that the audio is in; uses detected language if None
|
||||
|
||||
# sampling-related options
|
||||
temperature: float = 0.0
|
||||
sample_len: Optional[int] = None # maximum number of tokens to sample
|
||||
best_of: Optional[int] = None # number of independent samples to collect, when t > 0
|
||||
beam_size: Optional[int] = None # number of beams in beam search, when t == 0
|
||||
patience: Optional[float] = None # patience in beam search (https://arxiv.org/abs/2204.05424)
|
||||
|
||||
# options for ranking generations (either beams or best-of-N samples)
|
||||
length_penalty: Optional[float] = None # "alpha" in Google NMT, None defaults to length norm
|
||||
|
||||
# prompt, prefix, and token suppression
|
||||
prompt: Optional[Union[str, List[int]]] = None # text or tokens for the previous context
|
||||
prefix: Optional[Union[str, List[int]]] = None # text or tokens to prefix the current context
|
||||
suppress_blank: bool = True # this will suppress blank outputs
|
||||
|
||||
# list of tokens ids (or comma-separated token ids) to suppress
|
||||
# "-1" will suppress a set of symbols as defined in `tokenizer.non_speech_tokens()`
|
||||
suppress_tokens: Optional[Union[str, Iterable[int]]] = "-1"
|
||||
|
||||
# timestamp sampling options
|
||||
without_timestamps: bool = False # use <|notimestamps|> to sample text tokens only
|
||||
max_initial_timestamp: Optional[float] = 1.0 # the initial timestamp cannot be later than this
|
||||
|
||||
# implementation details
|
||||
fp16: bool = True # use fp16 for most of the calculation
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class DecodingResult:
|
||||
audio_features: Tensor
|
||||
language: str
|
||||
language_probs: Optional[Dict[str, float]] = None
|
||||
tokens: List[int] = field(default_factory=list)
|
||||
text: str = ""
|
||||
avg_logprob: float = np.nan
|
||||
no_speech_prob: float = np.nan
|
||||
temperature: float = np.nan
|
||||
compression_ratio: float = np.nan
|
||||
|
||||
|
||||
class Inference:
|
||||
def logits(self, tokens: Tensor, audio_features: Tensor) -> Tensor:
|
||||
"""Perform a forward pass on the decoder and return per-token logits"""
|
||||
raise NotImplementedError
|
||||
|
||||
def rearrange_kv_cache(self, source_indices) -> None:
|
||||
"""Update the key-value cache according to the updated beams"""
|
||||
raise NotImplementedError
|
||||
|
||||
def cleanup_caching(self) -> None:
|
||||
"""Clean up any resources or hooks after decoding is finished"""
|
||||
pass
|
||||
|
||||
|
||||
class PyTorchInference(Inference):
|
||||
def __init__(self, model: "Whisper", initial_token_length: int):
|
||||
self.model: "Whisper" = model
|
||||
self.initial_token_length = initial_token_length
|
||||
self.kv_cache = {}
|
||||
self.hooks = []
|
||||
|
||||
def logits(self, tokens: Tensor, audio_features: Tensor) -> Tensor:
|
||||
if not self.kv_cache:
|
||||
self.kv_cache, self.hooks = self.model.install_kv_cache_hooks()
|
||||
|
||||
if tokens.shape[-1] > self.initial_token_length:
|
||||
# only need to use the last token except in the first forward pass
|
||||
tokens = tokens[:, -1:]
|
||||
|
||||
return self.model.decoder(tokens, audio_features, kv_cache=self.kv_cache)
|
||||
|
||||
def cleanup_caching(self):
|
||||
for hook in self.hooks:
|
||||
hook.remove()
|
||||
|
||||
self.kv_cache = {}
|
||||
self.hooks = []
|
||||
|
||||
def rearrange_kv_cache(self, source_indices):
|
||||
for module, tensor in self.kv_cache.items():
|
||||
# update the key/value cache to contain the selected sequences
|
||||
self.kv_cache[module] = tensor[source_indices].detach()
|
||||
|
||||
|
||||
class SequenceRanker:
|
||||
def rank(self, tokens: List[List[Tensor]], sum_logprobs: List[List[float]]) -> List[int]:
|
||||
"""
|
||||
Given a list of groups of samples and their cumulative log probabilities,
|
||||
return the indices of the samples in each group to select as the final result
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
|
||||
class MaximumLikelihoodRanker(SequenceRanker):
|
||||
"""
|
||||
Select the sample with the highest log probabilities, penalized using either
|
||||
a simple length normalization or Google NMT paper's length penalty
|
||||
"""
|
||||
|
||||
def __init__(self, length_penalty: Optional[float]):
|
||||
self.length_penalty = length_penalty
|
||||
|
||||
def rank(self, tokens: List[List[Tensor]], sum_logprobs: List[List[float]]):
|
||||
def scores(logprobs, lengths):
|
||||
result = []
|
||||
for logprob, length in zip(logprobs, lengths):
|
||||
if self.length_penalty is None:
|
||||
penalty = length
|
||||
else:
|
||||
# from the Google NMT paper
|
||||
penalty = ((5 + length) / 6) ** self.length_penalty
|
||||
result.append(logprob / penalty)
|
||||
return result
|
||||
|
||||
# get the sequence with the highest score
|
||||
lengths = [[len(t) for t in s] for s in tokens]
|
||||
return [np.argmax(scores(p, l)) for p, l in zip(sum_logprobs, lengths)]
|
||||
|
||||
|
||||
class TokenDecoder:
|
||||
def reset(self):
|
||||
"""Initialize any stateful variables for decoding a new sequence"""
|
||||
|
||||
def update(self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor) -> Tuple[Tensor, bool]:
|
||||
"""Specify how to select the next token, based on the current trace and logits
|
||||
|
||||
Parameters
|
||||
----------
|
||||
tokens : Tensor, shape = (n_batch, current_sequence_length)
|
||||
all tokens in the context so far, including the prefix and sot_sequence tokens
|
||||
|
||||
logits : Tensor, shape = (n_batch, vocab_size)
|
||||
per-token logits of the probability distribution at the current step
|
||||
|
||||
sum_logprobs : Tensor, shape = (n_batch)
|
||||
cumulative log probabilities for each sequence
|
||||
|
||||
Returns
|
||||
-------
|
||||
tokens : Tensor, shape = (n_batch, current_sequence_length + 1)
|
||||
the tokens, appended with the selected next token
|
||||
|
||||
completed : bool
|
||||
True if all sequences has reached the end of text
|
||||
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
def finalize(
|
||||
self, tokens: Tensor, sum_logprobs: Tensor
|
||||
) -> Tuple[Sequence[Sequence[Tensor]], List[List[float]]]:
|
||||
"""Finalize search and return the final candidate sequences
|
||||
|
||||
Parameters
|
||||
----------
|
||||
tokens : Tensor, shape = (n_audio, n_group, current_sequence_length)
|
||||
all tokens in the context so far, including the prefix and sot_sequence
|
||||
|
||||
sum_logprobs : Tensor, shape = (n_audio, n_group)
|
||||
cumulative log probabilities for each sequence
|
||||
|
||||
Returns
|
||||
-------
|
||||
tokens : Sequence[Sequence[Tensor]], length = n_audio
|
||||
sequence of Tensors containing candidate token sequences, for each audio input
|
||||
|
||||
sum_logprobs : List[List[float]], length = n_audio
|
||||
sequence of cumulative log probabilities corresponding to the above
|
||||
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
|
||||
class GreedyDecoder(TokenDecoder):
|
||||
def __init__(self, temperature: float, eot: int):
|
||||
self.temperature = temperature
|
||||
self.eot = eot
|
||||
|
||||
def update(self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor) -> Tuple[Tensor, bool]:
|
||||
temperature = self.temperature
|
||||
if temperature == 0:
|
||||
next_tokens = logits.argmax(dim=-1)
|
||||
else:
|
||||
next_tokens = Categorical(logits=logits / temperature).sample()
|
||||
|
||||
logprobs = F.log_softmax(logits.float(), dim=-1)
|
||||
current_logprobs = logprobs[torch.arange(logprobs.shape[0]), next_tokens]
|
||||
sum_logprobs += current_logprobs * (tokens[:, -1] != self.eot)
|
||||
|
||||
next_tokens[tokens[:, -1] == self.eot] = self.eot
|
||||
tokens = torch.cat([tokens, next_tokens[:, None]], dim=-1)
|
||||
|
||||
completed = (tokens[:, -1] == self.eot).all()
|
||||
return tokens, completed
|
||||
|
||||
def finalize(self, tokens: Tensor, sum_logprobs: Tensor):
|
||||
# make sure each sequence has at least one EOT token at the end
|
||||
tokens = F.pad(tokens, (0, 1), value=self.eot)
|
||||
return tokens, sum_logprobs.tolist()
|
||||
|
||||
|
||||
class BeamSearchDecoder(TokenDecoder):
|
||||
def __init__(self, beam_size: int, eot: int, inference: Inference, patience: Optional[float] = None):
|
||||
self.beam_size = beam_size
|
||||
self.eot = eot
|
||||
self.inference = inference
|
||||
self.patience = patience or 1.0
|
||||
self.max_candidates: int = round(beam_size * self.patience)
|
||||
self.finished_sequences = None
|
||||
|
||||
assert self.max_candidates > 0, f"Invalid beam size ({beam_size}) or patience ({patience})"
|
||||
|
||||
def reset(self):
|
||||
self.finished_sequences = None
|
||||
|
||||
def update(self, tokens: Tensor, logits: Tensor, sum_logprobs: Tensor) -> Tuple[Tensor, bool]:
|
||||
if tokens.shape[0] % self.beam_size != 0:
|
||||
raise ValueError(f"{tokens.shape}[0] % {self.beam_size} != 0")
|
||||
|
||||
n_audio = tokens.shape[0] // self.beam_size
|
||||
if self.finished_sequences is None: # for the first update
|
||||
self.finished_sequences = [{} for _ in range(n_audio)]
|
||||
|
||||
logprobs = F.log_softmax(logits.float(), dim=-1)
|
||||
next_tokens, source_indices, finished_sequences = [], [], []
|
||||
for i in range(n_audio):
|
||||
scores, sources, finished = {}, {}, {}
|
||||
|
||||
# STEP 1: calculate the cumulative log probabilities for possible candidates
|
||||
for j in range(self.beam_size):
|
||||
idx = i * self.beam_size + j
|
||||
prefix = tokens[idx].tolist()
|
||||
for logprob, token in zip(*logprobs[idx].topk(self.beam_size + 1)):
|
||||
new_logprob = (sum_logprobs[idx] + logprob).item()
|
||||
sequence = tuple(prefix + [token.item()])
|
||||
scores[sequence] = new_logprob
|
||||
sources[sequence] = idx
|
||||
|
||||
# STEP 2: rank the candidates and keep the top beam_size sequences for each audio
|
||||
saved = 0
|
||||
for sequence in sorted(scores, key=scores.get, reverse=True):
|
||||
if sequence[-1] == self.eot:
|
||||
finished[sequence] = scores[sequence]
|
||||
else:
|
||||
sum_logprobs[len(next_tokens)] = scores[sequence]
|
||||
next_tokens.append(sequence)
|
||||
source_indices.append(sources[sequence])
|
||||
|
||||
saved += 1
|
||||
if saved == self.beam_size:
|
||||
break
|
||||
|
||||
finished_sequences.append(finished)
|
||||
|
||||
tokens = torch.tensor(next_tokens, device=tokens.device)
|
||||
self.inference.rearrange_kv_cache(source_indices)
|
||||
|
||||
# add newly finished sequences to self.finished_sequences
|
||||
assert len(self.finished_sequences) == len(finished_sequences)
|
||||
for previously_finished, newly_finished in zip(self.finished_sequences, finished_sequences):
|
||||
for seq in sorted(newly_finished, key=newly_finished.get, reverse=True):
|
||||
if len(previously_finished) >= self.max_candidates:
|
||||
break # the candidate list is full
|
||||
previously_finished[seq] = newly_finished[seq]
|
||||
|
||||
# mark as completed if all audio has enough number of samples
|
||||
completed = all(
|
||||
len(sequences) >= self.max_candidates for sequences in self.finished_sequences
|
||||
)
|
||||
return tokens, completed
|
||||
|
||||
def finalize(self, preceding_tokens: Tensor, sum_logprobs: Tensor):
|
||||
# collect all finished sequences, including patience, and add unfinished ones if not enough
|
||||
sum_logprobs = sum_logprobs.cpu()
|
||||
for i, sequences in enumerate(self.finished_sequences):
|
||||
if len(sequences) < self.beam_size: # when not enough sequences are finished
|
||||
for j in list(np.argsort(sum_logprobs[i]))[::-1]:
|
||||
sequence = preceding_tokens[i, j].tolist() + [self.eot]
|
||||
sequences[tuple(sequence)] = sum_logprobs[i][j].item()
|
||||
if len(sequences) >= self.beam_size:
|
||||
break
|
||||
|
||||
tokens: List[List[Tensor]] = [
|
||||
[torch.tensor(seq) for seq in sequences.keys()] for sequences in self.finished_sequences
|
||||
]
|
||||
sum_logprobs: List[List[float]] = [
|
||||
list(sequences.values()) for sequences in self.finished_sequences
|
||||
]
|
||||
return tokens, sum_logprobs
|
||||
|
||||
|
||||
class LogitFilter:
|
||||
def apply(self, logits: Tensor, tokens: Tensor) -> None:
|
||||
"""Apply any filtering or masking to logits in-place
|
||||
|
||||
Parameters
|
||||
----------
|
||||
logits : Tensor, shape = (n_batch, vocab_size)
|
||||
per-token logits of the probability distribution at the current step
|
||||
|
||||
tokens : Tensor, shape = (n_batch, current_sequence_length)
|
||||
all tokens in the context so far, including the prefix and sot_sequence tokens
|
||||
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
|
||||
class SuppressBlank(LogitFilter):
|
||||
def __init__(self, tokenizer: Tokenizer, sample_begin: int):
|
||||
self.tokenizer = tokenizer
|
||||
self.sample_begin = sample_begin
|
||||
|
||||
def apply(self, logits: Tensor, tokens: Tensor):
|
||||
if tokens.shape[1] == self.sample_begin:
|
||||
logits[:, self.tokenizer.encode(" ") + [self.tokenizer.eot]] = -np.inf
|
||||
|
||||
|
||||
class SuppressTokens(LogitFilter):
|
||||
def __init__(self, suppress_tokens: Sequence[int]):
|
||||
self.suppress_tokens = list(suppress_tokens)
|
||||
|
||||
def apply(self, logits: Tensor, tokens: Tensor):
|
||||
logits[:, self.suppress_tokens] = -np.inf
|
||||
|
||||
|
||||
class ApplyTimestampRules(LogitFilter):
|
||||
def __init__(
|
||||
self, tokenizer: Tokenizer, sample_begin: int, max_initial_timestamp_index: Optional[int]
|
||||
):
|
||||
self.tokenizer = tokenizer
|
||||
self.sample_begin = sample_begin
|
||||
self.max_initial_timestamp_index = max_initial_timestamp_index
|
||||
|
||||
def apply(self, logits: Tensor, tokens: Tensor):
|
||||
# suppress <|notimestamps|> which is handled by without_timestamps
|
||||
if self.tokenizer.no_timestamps is not None:
|
||||
logits[:, self.tokenizer.no_timestamps] = -np.inf
|
||||
|
||||
# timestamps have to appear in pairs, except directly before EOT; mask logits accordingly
|
||||
for k in range(tokens.shape[0]):
|
||||
seq = [t for t in tokens[k, self.sample_begin :].tolist()]
|
||||
last_was_timestamp = len(seq) >= 1 and seq[-1] >= self.tokenizer.timestamp_begin
|
||||
penultimate_was_timestamp = len(seq) < 2 or seq[-2] >= self.tokenizer.timestamp_begin
|
||||
|
||||
if last_was_timestamp:
|
||||
if penultimate_was_timestamp: # has to be non-timestamp
|
||||
logits[k, self.tokenizer.timestamp_begin :] = -np.inf
|
||||
else: # cannot be normal text tokens
|
||||
logits[k, : self.tokenizer.eot] = -np.inf
|
||||
|
||||
if tokens.shape[1] == self.sample_begin:
|
||||
# suppress generating non-timestamp tokens at the beginning
|
||||
logits[:, : self.tokenizer.timestamp_begin] = -np.inf
|
||||
|
||||
# apply the `max_initial_timestamp` option
|
||||
if self.max_initial_timestamp_index is not None:
|
||||
last_allowed = self.tokenizer.timestamp_begin + self.max_initial_timestamp_index
|
||||
logits[:, last_allowed + 1 :] = -np.inf
|
||||
|
||||
# if sum of probability over timestamps is above any other token, sample timestamp
|
||||
logprobs = F.log_softmax(logits.float(), dim=-1)
|
||||
for k in range(tokens.shape[0]):
|
||||
timestamp_logprob = logprobs[k, self.tokenizer.timestamp_begin :].logsumexp(dim=-1)
|
||||
max_text_token_logprob = logprobs[k, : self.tokenizer.timestamp_begin].max()
|
||||
if timestamp_logprob > max_text_token_logprob:
|
||||
logits[k, : self.tokenizer.timestamp_begin] = -np.inf
|
||||
|
||||
|
||||
class DecodingTask:
|
||||
inference: Inference
|
||||
sequence_ranker: SequenceRanker
|
||||
decoder: TokenDecoder
|
||||
logit_filters: List[LogitFilter]
|
||||
|
||||
def __init__(self, model: "Whisper", options: DecodingOptions):
|
||||
self.model = model
|
||||
|
||||
language = options.language or "en"
|
||||
tokenizer = get_tokenizer(model.is_multilingual, language=language, task=options.task)
|
||||
self.tokenizer: Tokenizer = tokenizer
|
||||
self.options: DecodingOptions = self._verify_options(options)
|
||||
|
||||
self.n_group: int = options.beam_size or options.best_of or 1
|
||||
self.n_ctx: int = model.dims.n_text_ctx
|
||||
self.sample_len: int = options.sample_len or model.dims.n_text_ctx // 2
|
||||
|
||||
self.sot_sequence: Tuple[int] = tokenizer.sot_sequence
|
||||
if self.options.without_timestamps:
|
||||
self.sot_sequence = tokenizer.sot_sequence_including_notimestamps
|
||||
|
||||
self.initial_tokens: Tuple[int] = self._get_initial_tokens()
|
||||
self.sample_begin: int = len(self.initial_tokens)
|
||||
self.sot_index: int = self.initial_tokens.index(tokenizer.sot)
|
||||
|
||||
# inference: implements the forward pass through the decoder, including kv caching
|
||||
self.inference = PyTorchInference(model, len(self.initial_tokens))
|
||||
|
||||
# sequence ranker: implements how to rank a group of sampled sequences
|
||||
self.sequence_ranker = MaximumLikelihoodRanker(options.length_penalty)
|
||||
|
||||
# decoder: implements how to select the next tokens, given the autoregressive distribution
|
||||
if options.beam_size is not None:
|
||||
self.decoder = BeamSearchDecoder(
|
||||
options.beam_size, tokenizer.eot, self.inference, options.patience
|
||||
)
|
||||
else:
|
||||
self.decoder = GreedyDecoder(options.temperature, tokenizer.eot)
|
||||
|
||||
# logit filters: applies various rules to suppress or penalize certain tokens
|
||||
self.logit_filters = []
|
||||
if self.options.suppress_blank:
|
||||
self.logit_filters.append(SuppressBlank(self.tokenizer, self.sample_begin))
|
||||
if self.options.suppress_tokens:
|
||||
self.logit_filters.append(SuppressTokens(self._get_suppress_tokens()))
|
||||
if not options.without_timestamps:
|
||||
precision = CHUNK_LENGTH / model.dims.n_audio_ctx # usually 0.02 seconds
|
||||
max_initial_timestamp_index = None
|
||||
if options.max_initial_timestamp:
|
||||
max_initial_timestamp_index = round(self.options.max_initial_timestamp / precision)
|
||||
self.logit_filters.append(
|
||||
ApplyTimestampRules(tokenizer, self.sample_begin, max_initial_timestamp_index)
|
||||
)
|
||||
|
||||
def _verify_options(self, options: DecodingOptions) -> DecodingOptions:
|
||||
if options.beam_size is not None and options.best_of is not None:
|
||||
raise ValueError("beam_size and best_of can't be given together")
|
||||
if options.temperature == 0:
|
||||
if options.best_of is not None:
|
||||
raise ValueError("best_of with greedy sampling (T=0) is not compatible")
|
||||
if options.patience is not None and options.beam_size is None:
|
||||
raise ValueError("patience requires beam_size to be given")
|
||||
if options.length_penalty is not None and not (0 <= options.length_penalty <= 1):
|
||||
raise ValueError("length_penalty (alpha) should be a value between 0 and 1")
|
||||
|
||||
return options
|
||||
|
||||
def _get_initial_tokens(self) -> Tuple[int]:
|
||||
tokens = list(self.sot_sequence)
|
||||
prefix = self.options.prefix
|
||||
prompt = self.options.prompt
|
||||
|
||||
if prefix:
|
||||
prefix_tokens = (
|
||||
self.tokenizer.encode(" " + prefix.strip()) if isinstance(prefix, str) else prefix
|
||||
)
|
||||
if self.sample_len is not None:
|
||||
max_prefix_len = self.n_ctx // 2 - self.sample_len
|
||||
prefix_tokens = prefix_tokens[-max_prefix_len:]
|
||||
tokens = tokens + prefix_tokens
|
||||
|
||||
if prompt:
|
||||
prompt_tokens = (
|
||||
self.tokenizer.encode(" " + prompt.strip()) if isinstance(prompt, str) else prompt
|
||||
)
|
||||
tokens = [self.tokenizer.sot_prev] + prompt_tokens[-(self.n_ctx // 2 - 1) :] + tokens
|
||||
|
||||
return tuple(tokens)
|
||||
|
||||
def _get_suppress_tokens(self) -> Tuple[int]:
|
||||
suppress_tokens = self.options.suppress_tokens
|
||||
|
||||
if isinstance(suppress_tokens, str):
|
||||
suppress_tokens = [int(t) for t in suppress_tokens.split(",")]
|
||||
|
||||
if -1 in suppress_tokens:
|
||||
suppress_tokens = [t for t in suppress_tokens if t >= 0]
|
||||
suppress_tokens.extend(self.tokenizer.non_speech_tokens)
|
||||
elif suppress_tokens is None or len(suppress_tokens) == 0:
|
||||
suppress_tokens = [] # interpret empty string as an empty list
|
||||
else:
|
||||
assert isinstance(suppress_tokens, list), "suppress_tokens must be a list"
|
||||
|
||||
suppress_tokens.extend(
|
||||
[self.tokenizer.sot, self.tokenizer.sot_prev, self.tokenizer.sot_lm]
|
||||
)
|
||||
if self.tokenizer.no_speech is not None:
|
||||
# no-speech probability is collected separately
|
||||
suppress_tokens.append(self.tokenizer.no_speech)
|
||||
|
||||
return tuple(sorted(set(suppress_tokens)))
|
||||
|
||||
def _get_audio_features(self, mel: Tensor):
|
||||
if self.options.fp16:
|
||||
mel = mel.half()
|
||||
|
||||
if mel.shape[-2:] == (self.model.dims.n_audio_ctx, self.model.dims.n_audio_state):
|
||||
# encoded audio features are given; skip audio encoding
|
||||
audio_features = mel
|
||||
else:
|
||||
audio_features = self.model.encoder(mel)
|
||||
|
||||
if audio_features.dtype != (torch.float16 if self.options.fp16 else torch.float32):
|
||||
return TypeError(f"audio_features has an incorrect dtype: {audio_features.dtype}")
|
||||
|
||||
return audio_features
|
||||
|
||||
def _detect_language(self, audio_features: Tensor, tokens: Tensor):
|
||||
languages = [self.options.language] * audio_features.shape[0]
|
||||
lang_probs = None
|
||||
|
||||
if self.options.language is None or self.options.task == "lang_id":
|
||||
lang_tokens, lang_probs = self.model.detect_language(audio_features, self.tokenizer)
|
||||
languages = [max(probs, key=probs.get) for probs in lang_probs]
|
||||
if self.options.language is None:
|
||||
tokens[:, self.sot_index + 1] = lang_tokens # write language tokens
|
||||
|
||||
return languages, lang_probs
|
||||
|
||||
def _main_loop(self, audio_features: Tensor, tokens: Tensor):
|
||||
assert audio_features.shape[0] == tokens.shape[0]
|
||||
n_batch = tokens.shape[0]
|
||||
sum_logprobs: Tensor = torch.zeros(n_batch, device=audio_features.device)
|
||||
no_speech_probs = [np.nan] * n_batch
|
||||
|
||||
try:
|
||||
for i in range(self.sample_len):
|
||||
logits = self.inference.logits(tokens, audio_features)
|
||||
|
||||
if i == 0 and self.tokenizer.no_speech is not None: # save no_speech_probs
|
||||
probs_at_sot = logits[:, self.sot_index].float().softmax(dim=-1)
|
||||
no_speech_probs = probs_at_sot[:, self.tokenizer.no_speech].tolist()
|
||||
|
||||
# now we need to consider the logits at the last token only
|
||||
logits = logits[:, -1]
|
||||
|
||||
# apply the logit filters, e.g. for suppressing or applying penalty to
|
||||
for logit_filter in self.logit_filters:
|
||||
logit_filter.apply(logits, tokens)
|
||||
|
||||
# expand the tokens tensor with the selected next tokens
|
||||
tokens, completed = self.decoder.update(tokens, logits, sum_logprobs)
|
||||
|
||||
if completed or tokens.shape[-1] > self.n_ctx:
|
||||
break
|
||||
finally:
|
||||
self.inference.cleanup_caching()
|
||||
|
||||
return tokens, sum_logprobs, no_speech_probs
|
||||
|
||||
@torch.no_grad()
|
||||
def run(self, mel: Tensor) -> List[DecodingResult]:
|
||||
self.decoder.reset()
|
||||
tokenizer: Tokenizer = self.tokenizer
|
||||
n_audio: int = mel.shape[0]
|
||||
|
||||
audio_features: Tensor = self._get_audio_features(mel) # encoder forward pass
|
||||
tokens: Tensor = torch.tensor([self.initial_tokens]).repeat(n_audio, 1)
|
||||
|
||||
# detect language if requested, overwriting the language token
|
||||
languages, language_probs = self._detect_language(audio_features, tokens)
|
||||
if self.options.task == "lang_id":
|
||||
return [
|
||||
DecodingResult(audio_features=features, language=language, language_probs=probs)
|
||||
for features, language, probs in zip(audio_features, languages, language_probs)
|
||||
]
|
||||
|
||||
# repeat the audio & text tensors by the group size, for beam search or best-of-n sampling
|
||||
audio_features = audio_features.repeat_interleave(self.n_group, dim=0)
|
||||
tokens = tokens.repeat_interleave(self.n_group, dim=0).to(audio_features.device)
|
||||
|
||||
# call the main sampling loop
|
||||
tokens, sum_logprobs, no_speech_probs = self._main_loop(audio_features, tokens)
|
||||
|
||||
# reshape the tensors to have (n_audio, n_group) as the first two dimensions
|
||||
audio_features = audio_features[:: self.n_group]
|
||||
no_speech_probs = no_speech_probs[:: self.n_group]
|
||||
assert audio_features.shape[0] == len(no_speech_probs) == n_audio
|
||||
|
||||
tokens = tokens.reshape(n_audio, self.n_group, -1)
|
||||
sum_logprobs = sum_logprobs.reshape(n_audio, self.n_group)
|
||||
|
||||
# get the final candidates for each group, and slice between the first sampled token and EOT
|
||||
tokens, sum_logprobs = self.decoder.finalize(tokens, sum_logprobs)
|
||||
tokens: List[List[Tensor]] = [
|
||||
[t[self.sample_begin : (t == tokenizer.eot).nonzero()[0, 0]] for t in s] for s in tokens
|
||||
]
|
||||
|
||||
# select the top-ranked sample in each group
|
||||
selected = self.sequence_ranker.rank(tokens, sum_logprobs)
|
||||
tokens: List[List[int]] = [t[i].tolist() for i, t in zip(selected, tokens)]
|
||||
texts: List[str] = [tokenizer.decode(t).strip() for t in tokens]
|
||||
|
||||
sum_logprobs: List[float] = [lp[i] for i, lp in zip(selected, sum_logprobs)]
|
||||
avg_logprobs: List[float] = [lp / (len(t) + 1) for t, lp in zip(tokens, sum_logprobs)]
|
||||
|
||||
fields = (texts, languages, tokens, audio_features, avg_logprobs, no_speech_probs)
|
||||
if len(set(map(len, fields))) != 1:
|
||||
raise RuntimeError(f"inconsistent result lengths: {list(map(len, fields))}")
|
||||
|
||||
return [
|
||||
DecodingResult(
|
||||
audio_features=features,
|
||||
language=language,
|
||||
tokens=tokens,
|
||||
text=text,
|
||||
avg_logprob=avg_logprob,
|
||||
no_speech_prob=no_speech_prob,
|
||||
temperature=self.options.temperature,
|
||||
compression_ratio=compression_ratio(text),
|
||||
)
|
||||
for text, language, tokens, features, avg_logprob, no_speech_prob in zip(*fields)
|
||||
]
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def decode(model: "Whisper", mel: Tensor, options: DecodingOptions = DecodingOptions()) -> Union[DecodingResult, List[DecodingResult]]:
|
||||
"""
|
||||
Performs decoding of 30-second audio segment(s), provided as Mel spectrogram(s).
|
||||
|
||||
Parameters
|
||||
----------
|
||||
model: Whisper
|
||||
the Whisper model instance
|
||||
|
||||
mel: torch.Tensor, shape = (80, 3000) or (*, 80, 3000)
|
||||
A tensor containing the Mel spectrogram(s)
|
||||
|
||||
options: DecodingOptions
|
||||
A dataclass that contains all necessary options for decoding 30-second segments
|
||||
|
||||
Returns
|
||||
-------
|
||||
result: Union[DecodingResult, List[DecodingResult]]
|
||||
The result(s) of decoding contained in `DecodingResult` dataclass instance(s)
|
||||
"""
|
||||
single = mel.ndim == 2
|
||||
if single:
|
||||
mel = mel.unsqueeze(0)
|
||||
|
||||
result = DecodingTask(model, options).run(mel)
|
||||
|
||||
if single:
|
||||
result = result[0]
|
||||
|
||||
return result
|
||||
@@ -1,339 +0,0 @@
|
||||
import os
|
||||
from dataclasses import dataclass
|
||||
from functools import lru_cache
|
||||
from typing import List, Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
try:
|
||||
from transformers import GPT2TokenizerFast
|
||||
except ImportError:
|
||||
raise ImportError(
|
||||
"transformers was not installed. Please install transformers first."
|
||||
)
|
||||
|
||||
LANGUAGES = {
|
||||
"en": "english",
|
||||
"zh": "chinese",
|
||||
"de": "german",
|
||||
"es": "spanish",
|
||||
"ru": "russian",
|
||||
"ko": "korean",
|
||||
"fr": "french",
|
||||
"ja": "japanese",
|
||||
"pt": "portuguese",
|
||||
"tr": "turkish",
|
||||
"pl": "polish",
|
||||
"ca": "catalan",
|
||||
"nl": "dutch",
|
||||
"ar": "arabic",
|
||||
"sv": "swedish",
|
||||
"it": "italian",
|
||||
"id": "indonesian",
|
||||
"hi": "hindi",
|
||||
"fi": "finnish",
|
||||
"vi": "vietnamese",
|
||||
"he": "hebrew",
|
||||
"uk": "ukrainian",
|
||||
"el": "greek",
|
||||
"ms": "malay",
|
||||
"cs": "czech",
|
||||
"ro": "romanian",
|
||||
"da": "danish",
|
||||
"hu": "hungarian",
|
||||
"ta": "tamil",
|
||||
"no": "norwegian",
|
||||
"th": "thai",
|
||||
"ur": "urdu",
|
||||
"hr": "croatian",
|
||||
"bg": "bulgarian",
|
||||
"lt": "lithuanian",
|
||||
"la": "latin",
|
||||
"mi": "maori",
|
||||
"ml": "malayalam",
|
||||
"cy": "welsh",
|
||||
"sk": "slovak",
|
||||
"te": "telugu",
|
||||
"fa": "persian",
|
||||
"lv": "latvian",
|
||||
"bn": "bengali",
|
||||
"sr": "serbian",
|
||||
"az": "azerbaijani",
|
||||
"sl": "slovenian",
|
||||
"kn": "kannada",
|
||||
"et": "estonian",
|
||||
"mk": "macedonian",
|
||||
"br": "breton",
|
||||
"eu": "basque",
|
||||
"is": "icelandic",
|
||||
"hy": "armenian",
|
||||
"ne": "nepali",
|
||||
"mn": "mongolian",
|
||||
"bs": "bosnian",
|
||||
"kk": "kazakh",
|
||||
"sq": "albanian",
|
||||
"sw": "swahili",
|
||||
"gl": "galician",
|
||||
"mr": "marathi",
|
||||
"pa": "punjabi",
|
||||
"si": "sinhala",
|
||||
"km": "khmer",
|
||||
"sn": "shona",
|
||||
"yo": "yoruba",
|
||||
"so": "somali",
|
||||
"af": "afrikaans",
|
||||
"oc": "occitan",
|
||||
"ka": "georgian",
|
||||
"be": "belarusian",
|
||||
"tg": "tajik",
|
||||
"sd": "sindhi",
|
||||
"gu": "gujarati",
|
||||
"am": "amharic",
|
||||
"yi": "yiddish",
|
||||
"lo": "lao",
|
||||
"uz": "uzbek",
|
||||
"fo": "faroese",
|
||||
"ht": "haitian creole",
|
||||
"ps": "pashto",
|
||||
"tk": "turkmen",
|
||||
"nn": "nynorsk",
|
||||
"mt": "maltese",
|
||||
"sa": "sanskrit",
|
||||
"lb": "luxembourgish",
|
||||
"my": "myanmar",
|
||||
"bo": "tibetan",
|
||||
"tl": "tagalog",
|
||||
"mg": "malagasy",
|
||||
"as": "assamese",
|
||||
"tt": "tatar",
|
||||
"haw": "hawaiian",
|
||||
"ln": "lingala",
|
||||
"ha": "hausa",
|
||||
"ba": "bashkir",
|
||||
"jw": "javanese",
|
||||
"su": "sundanese",
|
||||
}
|
||||
|
||||
# language code lookup by name, with a few language aliases
|
||||
TO_LANGUAGE_CODE = {
|
||||
**{language: code for code, language in LANGUAGES.items()},
|
||||
"burmese": "my",
|
||||
"valencian": "ca",
|
||||
"flemish": "nl",
|
||||
"haitian": "ht",
|
||||
"letzeburgesch": "lb",
|
||||
"pushto": "ps",
|
||||
"panjabi": "pa",
|
||||
"moldavian": "ro",
|
||||
"moldovan": "ro",
|
||||
"sinhalese": "si",
|
||||
"castilian": "es",
|
||||
}
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Tokenizer:
|
||||
"""A thin wrapper around `GPT2TokenizerFast` providing quick access to special tokens"""
|
||||
|
||||
tokenizer: "GPT2TokenizerFast"
|
||||
language: Optional[str]
|
||||
sot_sequence: Tuple[int]
|
||||
|
||||
def encode(self, text, **kwargs):
|
||||
return self.tokenizer.encode(text, **kwargs)
|
||||
|
||||
def decode(self, token_ids: Union[int, List[int], np.ndarray, torch.Tensor], **kwargs):
|
||||
return self.tokenizer.decode(token_ids, **kwargs)
|
||||
|
||||
def decode_with_timestamps(self, tokens) -> str:
|
||||
"""
|
||||
Timestamp tokens are above the special tokens' id range and are ignored by `decode()`.
|
||||
This method decodes given tokens with timestamps tokens annotated, e.g. "<|1.08|>".
|
||||
"""
|
||||
outputs = [[]]
|
||||
for token in tokens:
|
||||
if token >= self.timestamp_begin:
|
||||
timestamp = f"<|{(token - self.timestamp_begin) * 0.02:.2f}|>"
|
||||
outputs.append(timestamp)
|
||||
outputs.append([])
|
||||
else:
|
||||
outputs[-1].append(token)
|
||||
outputs = [s if isinstance(s, str) else self.tokenizer.decode(s) for s in outputs]
|
||||
return "".join(outputs)
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def eot(self) -> int:
|
||||
return self.tokenizer.eos_token_id
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def sot(self) -> int:
|
||||
return self._get_single_token_id("<|startoftranscript|>")
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def sot_lm(self) -> int:
|
||||
return self._get_single_token_id("<|startoflm|>")
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def sot_prev(self) -> int:
|
||||
return self._get_single_token_id("<|startofprev|>")
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def no_speech(self) -> int:
|
||||
return self._get_single_token_id("<|nospeech|>")
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def no_timestamps(self) -> int:
|
||||
return self._get_single_token_id("<|notimestamps|>")
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def timestamp_begin(self) -> int:
|
||||
return self.tokenizer.all_special_ids[-1] + 1
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def language_token(self) -> int:
|
||||
"""Returns the token id corresponding to the value of the `language` field"""
|
||||
if self.language is None:
|
||||
raise ValueError(f"This tokenizer does not have language token configured")
|
||||
|
||||
additional_tokens = dict(
|
||||
zip(
|
||||
self.tokenizer.additional_special_tokens,
|
||||
self.tokenizer.additional_special_tokens_ids,
|
||||
)
|
||||
)
|
||||
candidate = f"<|{self.language}|>"
|
||||
if candidate in additional_tokens:
|
||||
return additional_tokens[candidate]
|
||||
|
||||
raise KeyError(f"Language {self.language} not found in tokenizer.")
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def all_language_tokens(self) -> Tuple[int]:
|
||||
result = []
|
||||
for token, token_id in zip(
|
||||
self.tokenizer.additional_special_tokens,
|
||||
self.tokenizer.additional_special_tokens_ids,
|
||||
):
|
||||
if token.strip("<|>") in LANGUAGES:
|
||||
result.append(token_id)
|
||||
return tuple(result)
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def all_language_codes(self) -> Tuple[str]:
|
||||
return tuple(self.decode([l]).strip("<|>") for l in self.all_language_tokens)
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def sot_sequence_including_notimestamps(self) -> Tuple[int]:
|
||||
return tuple(list(self.sot_sequence) + [self.no_timestamps])
|
||||
|
||||
@property
|
||||
@lru_cache()
|
||||
def non_speech_tokens(self) -> Tuple[int]:
|
||||
"""
|
||||
Returns the list of tokens to suppress in order to avoid any speaker tags or non-speech
|
||||
annotations, to prevent sampling texts that are not actually spoken in the audio, e.g.
|
||||
|
||||
- ♪♪♪
|
||||
- ( SPEAKING FOREIGN LANGUAGE )
|
||||
- [DAVID] Hey there,
|
||||
|
||||
keeping basic punctuations like commas, periods, question marks, exclamation points, etc.
|
||||
"""
|
||||
symbols = list("\"#()*+/:;<=>@[\\]^_`{|}~「」『』")
|
||||
symbols += "<< >> <<< >>> -- --- -( -[ (' (\" (( )) ((( ))) [[ ]] {{ }} ♪♪ ♪♪♪".split()
|
||||
|
||||
# symbols that may be a single token or multiple tokens depending on the tokenizer.
|
||||
# In case they're multiple tokens, suppress the first token, which is safe because:
|
||||
# These are between U+2640 and U+267F miscellaneous symbols that are okay to suppress
|
||||
# in generations, and in the 3-byte UTF-8 representation they share the first two bytes.
|
||||
miscellaneous = set("♩♪♫♬♭♮♯")
|
||||
assert all(0x2640 <= ord(c) <= 0x267F for c in miscellaneous)
|
||||
|
||||
# allow hyphens "-" and single quotes "'" between words, but not at the beginning of a word
|
||||
result = {self.tokenizer.encode(" -")[0], self.tokenizer.encode(" '")[0]}
|
||||
for symbol in symbols + list(miscellaneous):
|
||||
for tokens in [self.tokenizer.encode(symbol), self.tokenizer.encode(" " + symbol)]:
|
||||
if len(tokens) == 1 or symbol in miscellaneous:
|
||||
result.add(tokens[0])
|
||||
|
||||
return tuple(sorted(result))
|
||||
|
||||
def _get_single_token_id(self, text) -> int:
|
||||
tokens = self.tokenizer.encode(text)
|
||||
assert len(tokens) == 1, f"{text} is not encoded as a single token"
|
||||
return tokens[0]
|
||||
|
||||
|
||||
@lru_cache(maxsize=None)
|
||||
def build_tokenizer(name: str = "gpt2", resource_path: str = None):
|
||||
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
||||
if resource_path is not None:
|
||||
path = os.path.join(resource_path, name)
|
||||
else:
|
||||
path = os.path.join(os.path.dirname(__file__), "assets", name)
|
||||
tokenizer = GPT2TokenizerFast.from_pretrained(path)
|
||||
|
||||
specials = [
|
||||
"<|startoftranscript|>",
|
||||
*[f"<|{lang}|>" for lang in LANGUAGES.keys()],
|
||||
"<|translate|>",
|
||||
"<|transcribe|>",
|
||||
"<|startoflm|>",
|
||||
"<|startofprev|>",
|
||||
"<|nospeech|>",
|
||||
"<|notimestamps|>",
|
||||
]
|
||||
|
||||
tokenizer.add_special_tokens(dict(additional_special_tokens=specials))
|
||||
return tokenizer
|
||||
|
||||
|
||||
@lru_cache(maxsize=None)
|
||||
def get_tokenizer(
|
||||
multilingual: bool,
|
||||
*,
|
||||
task: Optional[str] = None, # Literal["transcribe", "translate", None]
|
||||
language: Optional[str] = None,
|
||||
) -> Tokenizer:
|
||||
if language is not None:
|
||||
language = language.lower()
|
||||
if language not in LANGUAGES:
|
||||
if language in TO_LANGUAGE_CODE:
|
||||
language = TO_LANGUAGE_CODE[language]
|
||||
else:
|
||||
raise ValueError(f"Unsupported language: {language}")
|
||||
|
||||
if multilingual:
|
||||
tokenizer_name = "multilingual"
|
||||
task = task or "transcribe"
|
||||
language = language or "en"
|
||||
else:
|
||||
tokenizer_name = "gpt2"
|
||||
task = None
|
||||
language = None
|
||||
|
||||
tokenizer = build_tokenizer(name=tokenizer_name)
|
||||
all_special_ids: List[int] = tokenizer.all_special_ids
|
||||
sot: int = all_special_ids[1]
|
||||
translate: int = all_special_ids[-6]
|
||||
transcribe: int = all_special_ids[-5]
|
||||
|
||||
langs = tuple(LANGUAGES.keys())
|
||||
sot_sequence = [sot]
|
||||
if language is not None:
|
||||
sot_sequence.append(sot + 1 + langs.index(language))
|
||||
if task is not None:
|
||||
sot_sequence.append(transcribe if task == "transcribe" else translate)
|
||||
|
||||
return Tokenizer(tokenizer=tokenizer, language=language, sot_sequence=tuple(sot_sequence))
|
||||
@@ -1,316 +0,0 @@
|
||||
import argparse
|
||||
import os
|
||||
import warnings
|
||||
from typing import Optional, Tuple, Union, TYPE_CHECKING
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import tqdm
|
||||
|
||||
from funasr.models.whisper.utils.audio import SAMPLE_RATE, N_FRAMES, HOP_LENGTH, pad_or_trim, log_mel_spectrogram
|
||||
from funasr.models.whisper.utils.decoding import DecodingOptions, DecodingResult
|
||||
from funasr.models.whisper.utils.tokenizer import LANGUAGES, TO_LANGUAGE_CODE, get_tokenizer
|
||||
from funasr.models.whisper.utils.utils import exact_div, format_timestamp, make_safe, optional_int, optional_float, str2bool, get_writer
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from .model import Whisper
|
||||
|
||||
|
||||
def transcribe(
|
||||
model: "Whisper",
|
||||
audio: Union[str, np.ndarray, torch.Tensor],
|
||||
*,
|
||||
verbose: Optional[bool] = None,
|
||||
temperature: Union[float, Tuple[float, ...]] = (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
|
||||
compression_ratio_threshold: Optional[float] = 2.4,
|
||||
logprob_threshold: Optional[float] = -1.0,
|
||||
no_speech_threshold: Optional[float] = 0.6,
|
||||
condition_on_previous_text: bool = True,
|
||||
**decode_options,
|
||||
):
|
||||
"""
|
||||
Transcribe an audio file using Whisper
|
||||
|
||||
Parameters
|
||||
----------
|
||||
model: Whisper
|
||||
The Whisper model instance
|
||||
|
||||
audio: Union[str, np.ndarray, torch.Tensor]
|
||||
The path to the audio file to open, or the audio waveform
|
||||
|
||||
verbose: bool
|
||||
Whether to display the text being decoded to the console. If True, displays all the details,
|
||||
If False, displays minimal details. If None, does not display anything
|
||||
|
||||
temperature: Union[float, Tuple[float, ...]]
|
||||
Temperature for sampling. It can be a tuple of temperatures, which will be successively used
|
||||
upon failures according to either `compression_ratio_threshold` or `logprob_threshold`.
|
||||
|
||||
compression_ratio_threshold: float
|
||||
If the gzip compression ratio is above this value, treat as failed
|
||||
|
||||
logprob_threshold: float
|
||||
If the average log probability over sampled tokens is below this value, treat as failed
|
||||
|
||||
no_speech_threshold: float
|
||||
If the no_speech probability is higher than this value AND the average log probability
|
||||
over sampled tokens is below `logprob_threshold`, consider the segment as silent
|
||||
|
||||
condition_on_previous_text: bool
|
||||
if True, the previous output of the model is provided as a prompt for the next window;
|
||||
disabling may make the text inconsistent across windows, but the model becomes less prone to
|
||||
getting stuck in a failure loop, such as repetition looping or timestamps going out of sync.
|
||||
|
||||
decode_options: dict
|
||||
Keyword arguments to construct `DecodingOptions` instances
|
||||
|
||||
Returns
|
||||
-------
|
||||
A dictionary containing the resulting text ("text") and segment-level details ("segments"), and
|
||||
the spoken language ("language"), which is detected when `decode_options["language"]` is None.
|
||||
"""
|
||||
dtype = torch.float16 if decode_options.get("fp16", True) else torch.float32
|
||||
if model.device == torch.device("cpu"):
|
||||
if torch.cuda.is_available():
|
||||
warnings.warn("Performing inference on CPU when CUDA is available")
|
||||
if dtype == torch.float16:
|
||||
warnings.warn("FP16 is not supported on CPU; using FP32 instead")
|
||||
dtype = torch.float32
|
||||
|
||||
if dtype == torch.float32:
|
||||
decode_options["fp16"] = False
|
||||
|
||||
mel = log_mel_spectrogram(audio)
|
||||
|
||||
if decode_options.get("language", None) is None:
|
||||
if not model.is_multilingual:
|
||||
decode_options["language"] = "en"
|
||||
else:
|
||||
if verbose:
|
||||
print("Detecting language using up to the first 30 seconds. Use `--language` to specify the language")
|
||||
segment = pad_or_trim(mel, N_FRAMES).to(model.device).to(dtype)
|
||||
_, probs = model.detect_language(segment)
|
||||
decode_options["language"] = max(probs, key=probs.get)
|
||||
if verbose is not None:
|
||||
print(f"Detected language: {LANGUAGES[decode_options['language']].title()}")
|
||||
|
||||
language = decode_options["language"]
|
||||
task = decode_options.get("task", "transcribe")
|
||||
tokenizer = get_tokenizer(model.is_multilingual, language=language, task=task)
|
||||
|
||||
def decode_with_fallback(segment: torch.Tensor) -> DecodingResult:
|
||||
temperatures = [temperature] if isinstance(temperature, (int, float)) else temperature
|
||||
decode_result = None
|
||||
|
||||
for t in temperatures:
|
||||
kwargs = {**decode_options}
|
||||
if t > 0:
|
||||
# disable beam_size and patience when t > 0
|
||||
kwargs.pop("beam_size", None)
|
||||
kwargs.pop("patience", None)
|
||||
else:
|
||||
# disable best_of when t == 0
|
||||
kwargs.pop("best_of", None)
|
||||
|
||||
options = DecodingOptions(**kwargs, temperature=t)
|
||||
decode_result = model.decode(segment, options)
|
||||
|
||||
needs_fallback = False
|
||||
if compression_ratio_threshold is not None and decode_result.compression_ratio > compression_ratio_threshold:
|
||||
needs_fallback = True # too repetitive
|
||||
if logprob_threshold is not None and decode_result.avg_logprob < logprob_threshold:
|
||||
needs_fallback = True # average log probability is too low
|
||||
|
||||
if not needs_fallback:
|
||||
break
|
||||
|
||||
return decode_result
|
||||
|
||||
seek = 0
|
||||
input_stride = exact_div(
|
||||
N_FRAMES, model.dims.n_audio_ctx
|
||||
) # mel frames per output token: 2
|
||||
time_precision = (
|
||||
input_stride * HOP_LENGTH / SAMPLE_RATE
|
||||
) # time per output token: 0.02 (seconds)
|
||||
all_tokens = []
|
||||
all_segments = []
|
||||
prompt_reset_since = 0
|
||||
|
||||
initial_prompt = decode_options.pop("initial_prompt", None) or []
|
||||
if initial_prompt:
|
||||
initial_prompt = tokenizer.encode(" " + initial_prompt.strip())
|
||||
all_tokens.extend(initial_prompt)
|
||||
|
||||
def add_segment(
|
||||
*, start: float, end: float, text_tokens: torch.Tensor, result: DecodingResult
|
||||
):
|
||||
text = tokenizer.decode([token for token in text_tokens if token < tokenizer.eot])
|
||||
if len(text.strip()) == 0: # skip empty text output
|
||||
return
|
||||
|
||||
all_segments.append(
|
||||
{
|
||||
"id": len(all_segments),
|
||||
"seek": seek,
|
||||
"start": start,
|
||||
"end": end,
|
||||
"text": text,
|
||||
"tokens": text_tokens.tolist(),
|
||||
"temperature": result.temperature,
|
||||
"avg_logprob": result.avg_logprob,
|
||||
"compression_ratio": result.compression_ratio,
|
||||
"no_speech_prob": result.no_speech_prob,
|
||||
}
|
||||
)
|
||||
if verbose:
|
||||
print(make_safe(f"[{format_timestamp(start)} --> {format_timestamp(end)}] {text}"))
|
||||
|
||||
# show the progress bar when verbose is False (otherwise the transcribed text will be printed)
|
||||
num_frames = mel.shape[-1]
|
||||
previous_seek_value = seek
|
||||
|
||||
with tqdm.tqdm(total=num_frames, unit='frames', disable=verbose is not False) as pbar:
|
||||
while seek < num_frames:
|
||||
timestamp_offset = float(seek * HOP_LENGTH / SAMPLE_RATE)
|
||||
segment = pad_or_trim(mel[:, seek:], N_FRAMES).to(model.device).to(dtype)
|
||||
segment_duration = segment.shape[-1] * HOP_LENGTH / SAMPLE_RATE
|
||||
|
||||
decode_options["prompt"] = all_tokens[prompt_reset_since:]
|
||||
result: DecodingResult = decode_with_fallback(segment)
|
||||
tokens = torch.tensor(result.tokens)
|
||||
|
||||
if no_speech_threshold is not None:
|
||||
# no voice activity check
|
||||
should_skip = result.no_speech_prob > no_speech_threshold
|
||||
if logprob_threshold is not None and result.avg_logprob > logprob_threshold:
|
||||
# don't skip if the logprob is high enough, despite the no_speech_prob
|
||||
should_skip = False
|
||||
|
||||
if should_skip:
|
||||
seek += segment.shape[-1] # fast-forward to the next segment boundary
|
||||
continue
|
||||
|
||||
timestamp_tokens: torch.Tensor = tokens.ge(tokenizer.timestamp_begin)
|
||||
consecutive = torch.where(timestamp_tokens[:-1] & timestamp_tokens[1:])[0].add_(1)
|
||||
if len(consecutive) > 0: # if the output contains two consecutive timestamp tokens
|
||||
last_slice = 0
|
||||
for current_slice in consecutive:
|
||||
sliced_tokens = tokens[last_slice:current_slice]
|
||||
start_timestamp_position = (
|
||||
sliced_tokens[0].item() - tokenizer.timestamp_begin
|
||||
)
|
||||
end_timestamp_position = (
|
||||
sliced_tokens[-1].item() - tokenizer.timestamp_begin
|
||||
)
|
||||
add_segment(
|
||||
start=timestamp_offset + start_timestamp_position * time_precision,
|
||||
end=timestamp_offset + end_timestamp_position * time_precision,
|
||||
text_tokens=sliced_tokens[1:-1],
|
||||
result=result,
|
||||
)
|
||||
last_slice = current_slice
|
||||
last_timestamp_position = (
|
||||
tokens[last_slice - 1].item() - tokenizer.timestamp_begin
|
||||
)
|
||||
seek += last_timestamp_position * input_stride
|
||||
all_tokens.extend(tokens[: last_slice + 1].tolist())
|
||||
else:
|
||||
duration = segment_duration
|
||||
timestamps = tokens[timestamp_tokens.nonzero().flatten()]
|
||||
if len(timestamps) > 0 and timestamps[-1].item() != tokenizer.timestamp_begin:
|
||||
# no consecutive timestamps but it has a timestamp; use the last one.
|
||||
# single timestamp at the end means no speech after the last timestamp.
|
||||
last_timestamp_position = timestamps[-1].item() - tokenizer.timestamp_begin
|
||||
duration = last_timestamp_position * time_precision
|
||||
|
||||
add_segment(
|
||||
start=timestamp_offset,
|
||||
end=timestamp_offset + duration,
|
||||
text_tokens=tokens,
|
||||
result=result,
|
||||
)
|
||||
|
||||
seek += segment.shape[-1]
|
||||
all_tokens.extend(tokens.tolist())
|
||||
|
||||
if not condition_on_previous_text or result.temperature > 0.5:
|
||||
# do not feed the prompt tokens if a high temperature was used
|
||||
prompt_reset_since = len(all_tokens)
|
||||
|
||||
# update progress bar
|
||||
pbar.update(min(num_frames, seek) - previous_seek_value)
|
||||
previous_seek_value = seek
|
||||
|
||||
return dict(text=tokenizer.decode(all_tokens[len(initial_prompt):]), segments=all_segments, language=language)
|
||||
|
||||
|
||||
def cli():
|
||||
from . import available_models
|
||||
|
||||
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
|
||||
parser.add_argument("audio", nargs="+", type=str, help="audio file(s) to transcribe")
|
||||
parser.add_argument("--model", default="small", choices=available_models(), help="name of the Whisper model to use")
|
||||
parser.add_argument("--model_dir", type=str, default=None, help="the path to save model files; uses ~/.cache/whisper by default")
|
||||
parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu", help="device to use for PyTorch inference")
|
||||
parser.add_argument("--output_dir", "-o", type=str, default=".", help="directory to save the outputs")
|
||||
parser.add_argument("--output_format", "-f", type=str, default="all", choices=["txt", "vtt", "srt", "tsv", "json", "all"], help="format of the output file; if not specified, all available formats will be produced")
|
||||
parser.add_argument("--verbose", type=str2bool, default=True, help="whether to print out the progress and debug messages")
|
||||
|
||||
parser.add_argument("--task", type=str, default="transcribe", choices=["transcribe", "translate"], help="whether to perform X->X speech recognition ('transcribe') or X->English translation ('translate')")
|
||||
parser.add_argument("--language", type=str, default=None, choices=sorted(LANGUAGES.keys()) + sorted([k.title() for k in TO_LANGUAGE_CODE.keys()]), help="language spoken in the audio, specify None to perform language detection")
|
||||
|
||||
parser.add_argument("--temperature", type=float, default=0, help="temperature to use for sampling")
|
||||
parser.add_argument("--best_of", type=optional_int, default=5, help="number of candidates when sampling with non-zero temperature")
|
||||
parser.add_argument("--beam_size", type=optional_int, default=5, help="number of beams in beam search, only applicable when temperature is zero")
|
||||
parser.add_argument("--patience", type=float, default=None, help="optional patience value to use in beam decoding, as in https://arxiv.org/abs/2204.05424, the default (1.0) is equivalent to conventional beam search")
|
||||
parser.add_argument("--length_penalty", type=float, default=None, help="optional token length penalty coefficient (alpha) as in https://arxiv.org/abs/1609.08144, uses simple length normalization by default")
|
||||
|
||||
parser.add_argument("--suppress_tokens", type=str, default="-1", help="comma-separated list of token ids to suppress during sampling; '-1' will suppress most special characters except common punctuations")
|
||||
parser.add_argument("--initial_prompt", type=str, default=None, help="optional text to provide as a prompt for the first window.")
|
||||
parser.add_argument("--condition_on_previous_text", type=str2bool, default=True, help="if True, provide the previous output of the model as a prompt for the next window; disabling may make the text inconsistent across windows, but the model becomes less prone to getting stuck in a failure loop")
|
||||
parser.add_argument("--fp16", type=str2bool, default=True, help="whether to perform inference in fp16; True by default")
|
||||
|
||||
parser.add_argument("--temperature_increment_on_fallback", type=optional_float, default=0.2, help="temperature to increase when falling back when the decoding fails to meet either of the thresholds below")
|
||||
parser.add_argument("--compression_ratio_threshold", type=optional_float, default=2.4, help="if the gzip compression ratio is higher than this value, treat the decoding as failed")
|
||||
parser.add_argument("--logprob_threshold", type=optional_float, default=-1.0, help="if the average log probability is lower than this value, treat the decoding as failed")
|
||||
parser.add_argument("--no_speech_threshold", type=optional_float, default=0.6, help="if the probability of the <|nospeech|> token is higher than this value AND the decoding has failed due to `logprob_threshold`, consider the segment as silence")
|
||||
parser.add_argument("--threads", type=optional_int, default=0, help="number of threads used by torch for CPU inference; supercedes MKL_NUM_THREADS/OMP_NUM_THREADS")
|
||||
|
||||
args = parser.parse_args().__dict__
|
||||
model_name: str = args.pop("model")
|
||||
model_dir: str = args.pop("model_dir")
|
||||
output_dir: str = args.pop("output_dir")
|
||||
output_format: str = args.pop("output_format")
|
||||
device: str = args.pop("device")
|
||||
os.makedirs(output_dir, exist_ok=True)
|
||||
|
||||
if model_name.endswith(".en") and args["language"] not in {"en", "English"}:
|
||||
if args["language"] is not None:
|
||||
warnings.warn(f"{model_name} is an English-only model but receipted '{args['language']}'; using English instead.")
|
||||
args["language"] = "en"
|
||||
|
||||
temperature = args.pop("temperature")
|
||||
temperature_increment_on_fallback = args.pop("temperature_increment_on_fallback")
|
||||
if temperature_increment_on_fallback is not None:
|
||||
temperature = tuple(np.arange(temperature, 1.0 + 1e-6, temperature_increment_on_fallback))
|
||||
else:
|
||||
temperature = [temperature]
|
||||
|
||||
threads = args.pop("threads")
|
||||
if threads > 0:
|
||||
torch.set_num_threads(threads)
|
||||
|
||||
from . import load_model
|
||||
model = load_model(model_name, device=device, download_root=model_dir)
|
||||
|
||||
writer = get_writer(output_format, output_dir)
|
||||
|
||||
for audio_path in args.pop("audio"):
|
||||
result = transcribe(model, audio_path, temperature=temperature, **args)
|
||||
writer(result, audio_path)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
cli()
|
||||
@@ -1,163 +0,0 @@
|
||||
import json
|
||||
import os
|
||||
import sys
|
||||
import zlib
|
||||
from typing import Callable, TextIO
|
||||
|
||||
system_encoding = sys.getdefaultencoding()
|
||||
|
||||
if system_encoding != "utf-8":
|
||||
def make_safe(string):
|
||||
# replaces any character not representable using the system default encoding with an '?',
|
||||
# avoiding UnicodeEncodeError (https://github.com/openai/whisper/discussions/729).
|
||||
return string.encode(system_encoding, errors="replace").decode(system_encoding)
|
||||
else:
|
||||
def make_safe(string):
|
||||
# utf-8 can encode any Unicode code point, so no need to do the round-trip encoding
|
||||
return string
|
||||
|
||||
|
||||
def exact_div(x, y):
|
||||
assert x % y == 0
|
||||
return x // y
|
||||
|
||||
|
||||
def str2bool(string):
|
||||
str2val = {"True": True, "False": False}
|
||||
if string in str2val:
|
||||
return str2val[string]
|
||||
else:
|
||||
raise ValueError(f"Expected one of {set(str2val.keys())}, got {string}")
|
||||
|
||||
|
||||
def optional_int(string):
|
||||
return None if string == "None" else int(string)
|
||||
|
||||
|
||||
def optional_float(string):
|
||||
return None if string == "None" else float(string)
|
||||
|
||||
|
||||
def compression_ratio(text) -> float:
|
||||
text_bytes = text.encode("utf-8")
|
||||
return len(text_bytes) / len(zlib.compress(text_bytes))
|
||||
|
||||
|
||||
def format_timestamp(seconds: float, always_include_hours: bool = False, decimal_marker: str = '.'):
|
||||
assert seconds >= 0, "non-negative timestamp expected"
|
||||
milliseconds = round(seconds * 1000.0)
|
||||
|
||||
hours = milliseconds // 3_600_000
|
||||
milliseconds -= hours * 3_600_000
|
||||
|
||||
minutes = milliseconds // 60_000
|
||||
milliseconds -= minutes * 60_000
|
||||
|
||||
seconds = milliseconds // 1_000
|
||||
milliseconds -= seconds * 1_000
|
||||
|
||||
hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
|
||||
return f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
|
||||
|
||||
|
||||
class ResultWriter:
|
||||
extension: str
|
||||
|
||||
def __init__(self, output_dir: str):
|
||||
self.output_dir = output_dir
|
||||
|
||||
def __call__(self, result: dict, audio_path: str):
|
||||
audio_basename = os.path.basename(audio_path)
|
||||
output_path = os.path.join(self.output_dir, audio_basename + "." + self.extension)
|
||||
|
||||
with open(output_path, "w", encoding="utf-8") as f:
|
||||
self.write_result(result, file=f)
|
||||
|
||||
def write_result(self, result: dict, file: TextIO):
|
||||
raise NotImplementedError
|
||||
|
||||
|
||||
class WriteTXT(ResultWriter):
|
||||
extension: str = "txt"
|
||||
|
||||
def write_result(self, result: dict, file: TextIO):
|
||||
for segment in result["segments"]:
|
||||
print(segment['text'].strip(), file=file, flush=True)
|
||||
|
||||
|
||||
class WriteVTT(ResultWriter):
|
||||
extension: str = "vtt"
|
||||
|
||||
def write_result(self, result: dict, file: TextIO):
|
||||
print("WEBVTT\n", file=file)
|
||||
for segment in result["segments"]:
|
||||
print(
|
||||
f"{format_timestamp(segment['start'])} --> {format_timestamp(segment['end'])}\n"
|
||||
f"{segment['text'].strip().replace('-->', '->')}\n",
|
||||
file=file,
|
||||
flush=True,
|
||||
)
|
||||
|
||||
|
||||
class WriteSRT(ResultWriter):
|
||||
extension: str = "srt"
|
||||
|
||||
def write_result(self, result: dict, file: TextIO):
|
||||
for i, segment in enumerate(result["segments"], start=1):
|
||||
# write srt lines
|
||||
print(
|
||||
f"{i}\n"
|
||||
f"{format_timestamp(segment['start'], always_include_hours=True, decimal_marker=',')} --> "
|
||||
f"{format_timestamp(segment['end'], always_include_hours=True, decimal_marker=',')}\n"
|
||||
f"{segment['text'].strip().replace('-->', '->')}\n",
|
||||
file=file,
|
||||
flush=True,
|
||||
)
|
||||
|
||||
|
||||
class WriteTSV(ResultWriter):
|
||||
"""
|
||||
Write a transcript to a file in TSV (tab-separated values) format containing lines like:
|
||||
<start time in integer milliseconds>\t<end time in integer milliseconds>\t<transcript text>
|
||||
|
||||
Using integer milliseconds as start and end times means there's no chance of interference from
|
||||
an environment setting a language encoding that causes the decimal in a floating point number
|
||||
to appear as a comma; also is faster and more efficient to parse & store, e.g., in C++.
|
||||
"""
|
||||
extension: str = "tsv"
|
||||
|
||||
def write_result(self, result: dict, file: TextIO):
|
||||
print("start", "end", "text", sep="\t", file=file)
|
||||
for segment in result["segments"]:
|
||||
print(round(1000 * segment['start']), file=file, end="\t")
|
||||
print(round(1000 * segment['end']), file=file, end="\t")
|
||||
print(segment['text'].strip().replace("\t", " "), file=file, flush=True)
|
||||
|
||||
|
||||
class WriteJSON(ResultWriter):
|
||||
extension: str = "json"
|
||||
|
||||
def write_result(self, result: dict, file: TextIO):
|
||||
json.dump(result, file)
|
||||
|
||||
|
||||
def get_writer(output_format: str, output_dir: str) -> Callable[[dict, TextIO], None]:
|
||||
writers = {
|
||||
"txt": WriteTXT,
|
||||
"vtt": WriteVTT,
|
||||
"srt": WriteSRT,
|
||||
"tsv": WriteTSV,
|
||||
"json": WriteJSON,
|
||||
}
|
||||
|
||||
if output_format == "all":
|
||||
all_writers = [writer(output_dir) for writer in writers.values()]
|
||||
|
||||
def write_all(result: dict, file: TextIO):
|
||||
for writer in all_writers:
|
||||
writer(result, file)
|
||||
|
||||
return write_all
|
||||
|
||||
return writers[output_format](output_dir)
|
||||
|
||||
@@ -0,0 +1,24 @@
|
||||
|
||||
try:
|
||||
from whisper.tokenizer import get_tokenizer
|
||||
except:
|
||||
print("If you want to use hugging, please `pip install -U transformers`")
|
||||
|
||||
from funasr.register import tables
|
||||
|
||||
@tables.register("tokenizer_classes", "WhisperTokenizer")
|
||||
def WhisperTokenizer(**kwargs):
|
||||
|
||||
language = kwargs.get("language", None)
|
||||
task = kwargs.get("task", "transcribe")
|
||||
is_multilingual = kwargs.get("is_multilingual", True)
|
||||
num_languages = kwargs.get("num_languages", 99)
|
||||
tokenizer = get_tokenizer(
|
||||
multilingual=is_multilingual,
|
||||
num_languages=num_languages,
|
||||
language=language,
|
||||
task=task,
|
||||
)
|
||||
|
||||
return tokenizer
|
||||
|
||||
@@ -68,9 +68,9 @@ def load_pretrained_model(
|
||||
else:
|
||||
buffer = BytesIO(oss_bucket.get_object(path).read())
|
||||
src_state = torch.load(buffer, map_location=map_location)
|
||||
if "state_dict" in src_state:
|
||||
src_state = src_state["state_dict"]
|
||||
|
||||
|
||||
src_state = src_state["state_dict"] if "state_dict" in src_state else src_state
|
||||
src_state = src_state["model_state_dict"] if "model_state_dict" in src_state else src_state
|
||||
src_state = src_state["model"] if "model" in src_state else src_state
|
||||
|
||||
if isinstance(scope_map, str):
|
||||
|
||||
Reference in New Issue
Block a user