Spaces:
Running
on
Zero
Running
on
Zero
File size: 9,142 Bytes
52e4f53 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 |
import json
import math
import os
import torch
import natsort
from vita_audio.tokenizer import get_audio_tokenizer
class AudioProcessor:
def __init__(
self,
audio_tokenizer_path=None,
audio_tokenizer_type=None,
text_audio_interval_ratio=None,
):
self.audio_tokenizer = get_audio_tokenizer(
audio_tokenizer_path,
audio_tokenizer_type,
)
self.audio_tokenizer_type = audio_tokenizer_type
self.text_audio_interval_ratio = text_audio_interval_ratio
# self.load_model()
def load_model(self):
if self.audio_tokenizer is not None:
self.audio_tokenizer.load_model()
def process_audios(self, audio_path, is_discrete=False, is_contiguous=False, **kwargs):
assert not (is_discrete and is_contiguous)
assert is_discrete or is_contiguous
if is_discrete:
audio_tokenizer_type = self.audio_tokenizer_type.split("_")[-1]
cache_path = os.path.splitext(audio_path)[0] + f"_{audio_tokenizer_type}.json"
try:
if os.path.isfile(cache_path):
with open(cache_path, "r") as f:
audio_data = json.load(f)
return audio_data
except Exception as e:
pass
audio_data = self.audio_tokenizer.encode(
audio_path, is_discrete=is_discrete, is_contiguous=is_contiguous, **kwargs
)
# print(f"{len(audio_data)=}")
if is_discrete:
try:
if isinstance(audio_data, list):
with open(cache_path, "w") as f:
json.dump(audio_data, f)
except Exception as e:
pass
return audio_data
@property
def is_discrete(self):
return self.audio_tokenizer.is_discrete
@property
def is_contiguous(self):
return self.audio_tokenizer.is_contiguous
def apply_to_role(self, role, **kwargs):
return self.audio_tokenizer.apply_to_role(role, **kwargs)
def text_audio_interval(self, content_input_id, AUD_START_ID, AUD_END_ID):
return text_audio_interval(
content_input_id,
AUD_START_ID,
AUD_END_ID,
self.text_audio_interval_ratio,
)
def add_audio_input_contiguous(input_ids, audio_paths, tokenizer, audio_tokenizer):
from ...constants import (
AUD_START_TOKEN,
AUD_END_TOKEN,
AUD_TAG_TOKEN,
AUD_CONTEXT_TOKEN,
)
AUD_CONTEXT_ID = tokenizer(AUD_CONTEXT_TOKEN, add_special_tokens=False).input_ids
AUD_TAG_ID = tokenizer(AUD_TAG_TOKEN, add_special_tokens=False).input_ids
AUD_START_ID = tokenizer(AUD_START_TOKEN, add_special_tokens=False).input_ids
AUD_END_ID = tokenizer(AUD_END_TOKEN, add_special_tokens=False).input_ids
AUD_CONTEXT_ID = AUD_CONTEXT_ID[0]
AUD_TAG_ID = AUD_TAG_ID[0]
AUD_START_ID = AUD_START_ID[0]
AUD_END_ID = AUD_END_ID[0]
aud_positions = [i for i, x in enumerate(input_ids) if x == AUD_TAG_ID]
audios = []
audio_indices = []
new_input_ids = []
st = 0
for aud_idx, aud_pos in enumerate(aud_positions):
audio = audio_tokenizer.encode(audio_paths[aud_idx], is_contiguous=True)
audios.append(audio)
audio_token_length = audio.size(0) + 4
new_input_ids += input_ids[st:aud_pos]
new_input_ids += [AUD_START_ID]
audio_indice_b = torch.zeros(
1, audio_token_length, dtype=torch.int64
) # This will change in collate_fn
audio_indice_s = (
torch.arange(len(new_input_ids), len(new_input_ids) + audio_token_length)
.unsqueeze(0)
.repeat(1, 1)
)
audio_indice_b_s = torch.stack(
[audio_indice_b, audio_indice_s], dim=0
) # 2, num_image, image_length
audio_indices.append(audio_indice_b_s)
new_input_ids += [AUD_CONTEXT_ID] * audio_token_length
new_input_ids += [AUD_END_ID]
st = aud_pos + 1
new_input_ids += input_ids[st:]
inputs_ids = new_input_ids
return inputs_ids, audios, audio_indices
def text_audio_interval_old(input_ids, AUD_START_ID, AUD_END_ID, text_audio_interval_ratio):
if text_audio_interval_ratio is not None:
text_num, audio_num = text_audio_interval_ratio
else:
text_num = 13
audio_num = 26
text_num = 4
audio_num = 10
# exclude AUD_START and AUD_END
audio_num = audio_num - 2
st = [i for i, x in enumerate(input_ids) if x == AUD_START_ID]
ed = [i for i, x in enumerate(input_ids) if x == AUD_END_ID]
# only text
if len(st) == 0 and len(ed) == 0:
return input_ids
assert len(st) == 1
assert len(ed) == 1
st = st[0]
ed = ed[0]
assert st < ed
# only audio
if st == 0 and ed == len(input_ids) - 1:
return input_ids
audio_tokens = input_ids[st + 1 : ed]
text_tokens = input_ids[:st] + input_ids[ed + 1 :]
if False:
audio_tokens_chunks = [
audio_tokens[i : i + audio_num] for i in range(0, len(audio_tokens), audio_num)
]
text_tokens_chunks = [
text_tokens[i : i + text_num] for i in range(0, len(text_tokens), text_num)
]
if False:
# [0 1] [2 3 4 5 6 audio_num-1] ...
audio_tokens_chunks = [audio_tokens[:2], audio_tokens[2:audio_num]] + [
audio_tokens[i : i + audio_num] for i in range(audio_num, len(audio_tokens), audio_num)
]
# [0] [1 2 text_num-1] ...
text_tokens_chunks = [text_tokens[:1], text_tokens[1:text_num]] + [
text_tokens[i : i + text_num] for i in range(text_num, len(text_tokens), text_num)
]
if True:
# [0 1 2 3 4 5 6 audio_num] [] ...
audio_tokens_chunks = [audio_tokens[:audio_num]] + [
audio_tokens[i : i + audio_num] for i in range(audio_num, len(audio_tokens), audio_num)
]
# [0] [] ...
text_tokens_chunks = [text_tokens[:1]] + [
text_tokens[i : i + text_num] for i in range(1, len(text_tokens), text_num)
]
chunk_num = min(len(audio_tokens_chunks), len(text_tokens_chunks))
audio_tokens_chunks = audio_tokens_chunks[: chunk_num - 1] + [
sum(audio_tokens_chunks[chunk_num - 1 :], [])
]
text_tokens_chunks = text_tokens_chunks[: chunk_num - 1] + [
sum(text_tokens_chunks[chunk_num - 1 :], [])
]
interval_input_ids = []
for text_tokens, audio_tokens in zip(text_tokens_chunks, audio_tokens_chunks):
interval_input_ids += text_tokens + [AUD_START_ID] + audio_tokens + [AUD_END_ID]
# interval_input_ids += text_tokens + audio_tokens
return interval_input_ids
def text_audio_interval(input_ids, AUD_START_ID, AUD_END_ID, text_audio_interval_ratio):
if text_audio_interval_ratio is None:
# T A
text_audio_interval_ratio = [13, 26]
# T A T A T A
text_audio_interval_ratio = [1, 4, 3, 8, 4, 10]
# T A T A
text_audio_interval_ratio = [1, 10, 4, 10]
text_nums = text_audio_interval_ratio[::2]
audio_nums = text_audio_interval_ratio[1::2]
# exclude AUD_START and AUD_END
audio_nums = [x - 2 for x in audio_nums]
st = [i for i, x in enumerate(input_ids) if x == AUD_START_ID]
ed = [i for i, x in enumerate(input_ids) if x == AUD_END_ID]
# only text
if len(st) == 0 and len(ed) == 0:
return input_ids
assert len(st) == 1
assert len(ed) == 1
st = st[0]
ed = ed[0]
assert st < ed
# only audio
if st == 0 and ed == len(input_ids) - 1:
return input_ids
audio_tokens = input_ids[st + 1 : ed]
text_tokens = input_ids[:st] + input_ids[ed + 1 :]
audio_tokens_chunks = []
while len(audio_tokens) > 0:
if len(audio_nums) > 1:
audio_num = audio_nums.pop(0)
else:
audio_num = audio_nums[0]
audio_tokens_chunks.append(audio_tokens[:audio_num])
audio_tokens = audio_tokens[audio_num:]
text_tokens_chunks = []
while len(text_tokens) > 0:
if len(text_nums) > 1:
text_num = text_nums.pop(0)
else:
text_num = text_nums[0]
text_tokens_chunks.append(text_tokens[:text_num])
text_tokens = text_tokens[text_num:]
chunk_num = min(len(audio_tokens_chunks), len(text_tokens_chunks))
audio_tokens_chunks = audio_tokens_chunks[: chunk_num - 1] + [
sum(audio_tokens_chunks[chunk_num - 1 :], [])
]
text_tokens_chunks = text_tokens_chunks[: chunk_num - 1] + [
sum(text_tokens_chunks[chunk_num - 1 :], [])
]
interval_input_ids = []
for text_tokens, audio_tokens in zip(text_tokens_chunks, audio_tokens_chunks):
interval_input_ids += text_tokens + [AUD_START_ID] + audio_tokens + [AUD_END_ID]
# interval_input_ids += text_tokens + audio_tokens
return interval_input_ids
|