Spaces:
Running
on
Zero
Running
on
Zero
File size: 7,300 Bytes
bc752b1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 |
# Copyright (c) 2022 Binbin Zhang ([email protected])
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Dict, List, Optional, Tuple
import torch
from torch import nn
import torchaudio
import torchaudio.compliance.kaldi as kaldi
from .adapter import CNNAdapter, CNNSubsampling, LinearAdapter
from .cmvn import GlobalCMVN, load_cmvn
from .module.encoder.encoder import whaleEncoder
class audioEncoderProcessor:
def __init__(
self,
dataset_conf: dict = None,
):
self.dataset_conf = dataset_conf
def process(self, wav_path):
try:
waveform, sample_rate = torchaudio.load(wav_path)
except Exception as e:
print(f"cannot open {wav_path}!!!!!!!!!!!!!!!!")
if sample_rate != self.dataset_conf["resample_conf"]["resample_rate"]:
waveform = torchaudio.transforms.Resample(
orig_freq=sample_rate, new_freq=self.dataset_conf["resample_conf"]["resample_rate"]
)(waveform)
sample_rate = self.dataset_conf['resample_conf']['resample_rate']
waveform = waveform * (1 << 15)
# Only keep key, feat, label
mat = kaldi.fbank(
waveform,
num_mel_bins=self.dataset_conf["fbank_conf"]["num_mel_bins"],
frame_length=self.dataset_conf["fbank_conf"]["frame_length"],
frame_shift=self.dataset_conf["fbank_conf"]["frame_shift"],
dither=self.dataset_conf["fbank_conf"]["dither"],
energy_floor=0.0,
sample_frequency=sample_rate,
)
attn_mask = torch.ones(mat.shape[0])
attn_mask = attn_mask[2::2][2::2][0::2]
return mat, attn_mask.shape[0]
class audioEncoder(torch.nn.Module):
def __init__(
self,
encoder: torch.nn.Module,
llm_path: str,
freeze_llm: bool = True,
enc_out_dim: int = 512,
llm_embed_dim: int = 4096,
kernel_size: int = 3,
IGNORE_ID: int = -100,
adpter_type: str = "cnn",
add_audio_bos_eos: bool = False,
task_num: int = 10,
task_before_audio: bool = False,
task_type: str = "prompt",
freeze_encoder: bool = False,
freeze_adpter: bool = False,
audio_prompt_finetune: bool = False,
audio_prompt_num: int = 25,
activation_func: str = "relu",
norm: str = "batch",
chat_template=None,
):
super().__init__()
self.encoder = encoder
self.enc_out_dim = enc_out_dim
self.llm_embed_dim = llm_embed_dim
self.IGNORE_ID = IGNORE_ID
self.add_audio_bos_eos = add_audio_bos_eos
self.task_before_audio = task_before_audio
self.task_type = task_type
self.freeze_encoder = freeze_encoder
self.freeze_adpter = freeze_adpter
self.audio_prompt_finetune = audio_prompt_finetune
self.audio_prompt_num = audio_prompt_num
if adpter_type == "cnn":
self.adpter = CNNAdapter(enc_out_dim, llm_embed_dim, kernel_size)
elif adpter_type == "linear":
self.adpter = LinearAdapter(enc_out_dim, llm_embed_dim)
elif adpter_type == "subsampling":
self.adpter = CNNSubsampling(
enc_out_dim, llm_embed_dim, kernel_size, activation_func, norm
)
if self.freeze_encoder:
self.encoder.eval()
for (name, param) in self.encoder.named_parameters():
param.requires_grad = False
if self.freeze_adpter:
self.adpter.eval()
for (name, param) in self.adpter.named_parameters():
param.requires_grad = False
if self.audio_prompt_finetune:
self.prompt_embeddings = nn.Embedding(audio_prompt_num, llm_embed_dim)
self.prompt_ids = torch.tensor([i for i in range(audio_prompt_num)]).long()
def forward(
self,
speech: torch.Tensor,
speech_lengths: torch.Tensor,
) -> Dict[str, Optional[torch.Tensor]]:
speech = speech.to(next(self.parameters()).dtype)
# 1. Encoder
encoder_out, encoder_mask = self.encoder(speech, speech_lengths)
inputs_embeds, encoder_mask = self.adpter(encoder_out, encoder_mask) # B, T, D
attention_mask = encoder_mask.squeeze(1) # B, T
assert inputs_embeds.size(1) == attention_mask.size(1)
# audio bos/eos
if self.add_audio_bos_eos:
inputs_embeds, attention_mask, target = self._add_bos_eos(
"audio", "/audio", inputs_embeds, attention_mask, target
)
B, _, _ = inputs_embeds.shape
if self.audio_prompt_finetune:
prompt_ids = self.prompt_ids.repeat(B, 1).to(inputs_embeds.device)
prompt_embeds = self.prompt_embeddings(
prompt_ids.to(inputs_embeds.device)) # B, 5, D
inputs_embeds = torch.cat((prompt_embeds, inputs_embeds), 1) # B, (T+5), D
outputs = {
"inputs_embeds": inputs_embeds,
"attention_mask": attention_mask,
}
return outputs
def _add_bos_eos(self, bos, eos, inputs_embeds, attention_mask, target=None):
B = len(inputs_embeds)
bos_embed = self.task_embeddings(
torch.full([B, 1], self.task_ids[bos]).to(inputs_embeds.device)
) # B, 1, D
eos_embed = self.task_embeddings(
torch.full([B, 1], self.task_ids[eos]).to(inputs_embeds.device)
) # B, 1, D
bos_eos_target = torch.full([B, 2], self.IGNORE_ID).to(inputs_embeds.device) # B, 2
bos_eos_mask = torch.full([B, 1], True).to(inputs_embeds.device) # B, 1
inputs_embeds = torch.cat((bos_embed, inputs_embeds), 1) # B, (1+T), D
inputs_embeds = torch.cat((inputs_embeds, eos_embed), 1) # B, (1+T+1), D
attention_mask = torch.cat((bos_eos_mask, attention_mask), 1) # B, (1+T)
attention_mask = torch.cat((attention_mask, bos_eos_mask), 1) # B, (1+T+1)
if target is not None:
target = torch.cat((target, bos_eos_target), 1) # B, (T+2), D
return inputs_embeds, attention_mask, target
def init_model(configs):
if configs["cmvn_file"] is not None:
mean, istd = load_cmvn(configs["cmvn_file"], configs["is_json_cmvn"])
global_cmvn = GlobalCMVN(torch.from_numpy(mean).float(), torch.from_numpy(istd).float())
else:
global_cmvn = None
input_dim = configs["input_dim"]
encoder = whaleEncoder(input_dim, global_cmvn=global_cmvn, **configs["encoder_conf"])
model = audioEncoder(encoder=encoder, **configs["model_conf"])
processor = audioEncoderProcessor(dataset_conf=configs["dataset_conf"])
model.audio_processor = processor
return model
|