Spaces:
Paused
Paused
File size: 14,845 Bytes
9d7cf7f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 | from dataclasses import dataclass, field
from numpy import ndarray
from typing import Dict, Tuple, Union, List, Optional
import numpy as np
from .spec import Tokenizer, TokenizeInput, DetokenizeOutput
from .spec import make_skeleton
from ..data.order import Order
@dataclass
class TokenizerPart(Tokenizer):
# cls token id
cls_token_id: Dict[str, int]
# parts token id
parts_token_id: Dict[str, int]
part_token_to_name: Dict[int, str]
cls_token_to_name: Dict[int, str]
parts_token_id_name: List[str]
# normalization range
continuous_range: Tuple[float, float]
# coordinate discrete
num_discrete: int
token_id_branch: int
token_id_bos: int
token_id_eos: int
token_id_pad: int
token_id_spring: int
token_id_cls_none: int
_vocab_size: int
order: Optional[Order]=None
@classmethod
def parse(
cls,
**kwargs,
):
num_discrete = kwargs.pop('num_discrete')
continuous_range = kwargs.pop('continuous_range')
cls_token_id = kwargs.pop('cls_token_id')
parts_token_id = kwargs.pop('parts_token_id')
order = kwargs.get('order')
if order is not None:
assert isinstance(order, Order)
_offset = num_discrete
token_id_branch = _offset + 0
token_id_bos = _offset + 1
token_id_eos = _offset + 2
token_id_pad = _offset + 3
_offset += 4
token_id_spring = _offset + 0
_offset += 1
assert None not in parts_token_id
for i in parts_token_id:
parts_token_id[i] += _offset
_offset += len(parts_token_id)
token_id_cls_none = _offset + 0
_offset += 1
for i in cls_token_id:
cls_token_id[i] += _offset
_offset += len(cls_token_id)
_vocab_size = _offset
parts_token_id_name = [x for x in parts_token_id]
part_token_to_name = {v: k for k, v in parts_token_id.items()}
assert len(part_token_to_name) == len(parts_token_id), 'names with same token found in parts_token_id'
part_token_to_name[token_id_spring] = None
cls_token_to_name = {v: k for k, v in cls_token_id.items()}
assert len(cls_token_to_name) == len(cls_token_id), 'names with same token found in cls_token_id'
return TokenizerPart(
num_discrete=num_discrete,
continuous_range=continuous_range,
cls_token_id=cls_token_id,
parts_token_id=parts_token_id,
order=order,
token_id_branch=token_id_branch,
token_id_bos=token_id_bos,
token_id_eos=token_id_eos,
token_id_pad=token_id_pad,
token_id_spring=token_id_spring,
token_id_cls_none=token_id_cls_none,
parts_token_id_name=parts_token_id_name,
part_token_to_name=part_token_to_name,
cls_token_to_name=cls_token_to_name,
_vocab_size=_vocab_size,
)
def make_cls_head(self, **kwargs) -> List[int]:
cls = kwargs.get('cls', None)
if cls is not None:
return [self.cls_name_to_token(cls=cls)]
return [self.token_id_cls_none]
def cls_name_to_token(self, cls: str) -> int:
if cls not in self.cls_token_id:
return self.token_id_cls_none
return self.cls_token_id[cls]
def part_name_to_token(self, part: str) -> int:
assert part in self.parts_token_id, f"do not find part name `{part}` in tokenizer"
return self.parts_token_id[part]
def next_posible_token(self, ids: ndarray) -> List[int]:
if ids.shape[0] == 0 or ids.ndim == 0:
return [self.token_id_bos]
assert ids.ndim == 1, "expect an array"
state = 'expect_bos'
for id in ids:
if state == 'expect_bos':
assert id == self.token_id_bos, 'ids do not start with bos'
state = 'expect_cls_or_part_or_joint'
elif state == 'expect_cls_or_part_or_joint':
if id < self.num_discrete:
state = 'expect_joint_2'
elif id == self.token_id_cls_none or id in self.cls_token_id.values():
state = 'expect_part_or_joint'
else: # a part
state = 'expect_joint'
elif state == 'expect_part_or_joint':
if id < self.num_discrete:
state = 'expect_joint_2'
else:
state = 'expect_part_or_joint'
elif state == 'expect_joint_2':
state = 'expect_joint_3'
elif state == 'expect_joint_3':
state = 'expect_branch_or_part_or_joint'
elif state == 'expect_branch_or_part_or_joint':
if id == self.token_id_branch:
state = 'expect_joint'
elif id < self.num_discrete:
state = 'expect_joint_2'
else: # find a part
state = 'expect_joint'
elif state == 'expect_joint':
state = 'expect_joint_2'
else:
assert 0, state
s = []
def add_cls():
s.append(self.token_id_cls_none)
for v in self.cls_token_id.values():
s.append(v)
def add_part():
s.append(self.token_id_spring)
for v in self.parts_token_id.values():
s.append(v)
def add_joint():
for i in range(self.num_discrete):
s.append(i)
def add_branch():
s.append(self.token_id_branch)
def add_eos():
s.append(self.token_id_eos)
def add_bos():
s.append(self.token_id_bos)
if state == 'expect_bos':
add_bos()
elif state == 'expect_cls_or_part_or_joint':
add_cls()
add_part()
add_joint()
elif state == 'expect_cls':
add_cls()
elif state == 'expect_part_or_joint':
add_part()
add_joint()
add_eos()
elif state == 'expect_joint_2':
add_joint()
elif state == 'expect_joint_3':
add_joint()
elif state == 'expect_branch_or_part_or_joint':
add_joint()
add_part()
add_branch()
add_eos()
elif state == 'expect_joint':
add_joint()
else:
assert 0, state
return s
def bones_in_sequence(self, ids: ndarray):
assert ids.ndim == 1, "expect an array"
s = 0
is_branch = False
state = 'expect_bos'
for id in ids:
if state == 'expect_bos':
assert id == self.token_id_bos, 'ids do not start with bos'
state = 'expect_cls_or_part_or_joint'
elif state == 'expect_cls_or_part_or_joint':
if id < self.num_discrete:
state = 'expect_joint_2'
elif id == self.token_id_cls_none or id in self.cls_token_id.values():
state = 'expect_part_or_joint'
else: # a part
state = 'expect_joint'
elif state == 'expect_part_or_joint':
if id < self.num_discrete:
state = 'expect_joint_2'
else:
state = 'expect_part_or_joint'
elif state == 'expect_joint_2':
state = 'expect_joint_3'
elif state == 'expect_joint_3':
if not is_branch:
s += 1
is_branch = False
state = 'expect_branch_or_part_or_joint'
elif state == 'expect_branch_or_part_or_joint':
if id == self.token_id_branch:
state = 'expect_joint'
is_branch = True
elif id < self.num_discrete:
state = 'expect_joint_2'
else: # find a part
state = 'expect_joint'
elif state == 'expect_joint':
state = 'expect_joint_2'
else:
assert 0, state
if id == self.token_id_eos:
break
return s
def tokenize(self, input: TokenizeInput) -> ndarray:
num_bones = input.num_bones
bones = discretize(t=input.bones, continuous_range=self.continuous_range, num_discrete=self.num_discrete)
branch = input.branch
tokens = [self.token_id_bos]
if input.cls is None or input.cls not in self.cls_token_id:
tokens.append(self.token_id_cls_none)
else:
tokens.append(self.cls_token_id[input.cls])
if self.order is not None and input.cls is not None and input.joint_names is not None:
_, parts_bias = self.order.arrange_names(cls=input.cls, names=input.joint_names, parents=input.parents)
else:
parts_bias = []
for i in range(num_bones):
# add parts token id
if i in parts_bias:
part = parts_bias[i]
if part is None:
tokens.append(self.token_id_spring)
else:
assert part in self.parts_token_id, f"do not find part name {part} in tokenizer {self.__class__}"
tokens.append(self.parts_token_id[part])
if branch[i]:
tokens.append(self.token_id_branch)
tokens.append(bones[i, 0])
tokens.append(bones[i, 1])
tokens.append(bones[i, 2])
tokens.append(bones[i, 3])
tokens.append(bones[i, 4])
tokens.append(bones[i, 5])
else:
tokens.append(bones[i, 3])
tokens.append(bones[i, 4])
tokens.append(bones[i, 5])
tokens.append(self.token_id_eos)
return np.array(tokens, dtype=np.int64)
def detokenize(self, ids: ndarray, **kwargs) -> DetokenizeOutput:
assert isinstance(ids, ndarray), 'expect ids to be ndarray'
if ids[0] != self.token_id_bos:
raise ValueError(f"first token is not bos")
trailing_pad = 0
while trailing_pad < ids.shape[0] and ids[-trailing_pad-1] == self.token_id_pad:
trailing_pad += 1
if ids[-1-trailing_pad] != self.token_id_eos:
raise ValueError(f"last token is not eos")
ids = ids[1:-1-trailing_pad]
joints = []
p_joints = []
tails_dict = {}
parts = []
i = 0
is_branch = False
last_joint = None
num_bones = 0
cls = None
while i < len(ids):
if ids[i] < self.num_discrete:
if is_branch:
p_joint = undiscretize(t=ids[i:i+3], continuous_range=self.continuous_range, num_discrete=self.num_discrete)
current_joint = undiscretize(t=ids[i+3:i+6], continuous_range=self.continuous_range, num_discrete=self.num_discrete)
joints.append(current_joint)
p_joints.append(p_joint)
i += 6
else:
current_joint = undiscretize(t=ids[i:i+3], continuous_range=self.continuous_range, num_discrete=self.num_discrete)
joints.append(current_joint)
if len(p_joints) == 0: # root
p_joints.append(current_joint)
p_joint = current_joint
else:
assert last_joint is not None
p_joints.append(last_joint)
p_joint = last_joint
i += 3
if last_joint is not None:
tails_dict[num_bones-1] = current_joint
last_joint = current_joint
num_bones += 1
is_branch = False
elif ids[i]==self.token_id_branch:
is_branch = True
last_joint = None
i += 1
elif ids[i]==self.token_id_spring or ids[i] in self.parts_token_id.values():
parts.append(self.part_token_to_name[ids[i]])
i += 1
elif ids[i] in self.cls_token_id.values():
cls = ids[i]
i += 1
elif ids[i] == self.token_id_cls_none:
cls = None
i += 1
else:
raise ValueError(f"unexpected token found: {ids[i]}")
joints = np.stack(joints)
p_joints = np.stack(p_joints)
# leaf is ignored in this tokenizer so need to extrude tails for leaf and branch
bones, tails, available_bones_id, parents = make_skeleton(
joints=joints,
p_joints=p_joints,
tails_dict=tails_dict,
convert_leaf_bones_to_tails=False,
extrude_tail_for_leaf=True,
extrude_tail_for_branch=True,
)
bones = bones[available_bones_id]
tails = tails[available_bones_id]
if cls in self.cls_token_to_name:
cls = self.cls_token_to_name[cls]
else:
cls = None
if self.order is not None:
joint_names = self.order.make_names(cls=cls, parts=parts, num_bones=num_bones)
else:
joint_names = [f"bone_{i}" for i in range(num_bones)]
return DetokenizeOutput(
tokens=ids,
bones=bones,
parents=parents,
cls=cls,
joint_names=joint_names,
continuous_range=self.continuous_range,
)
def get_require_parts(self) -> List[str]:
return self.parts_token_id_name
@property
def vocab_size(self):
return self._vocab_size
@property
def pad(self):
return self.token_id_pad
@property
def bos(self):
return self.token_id_bos
@property
def eos(self):
return self.token_id_eos
def discretize(
t: ndarray,
continuous_range: Tuple[float, float],
num_discrete: int,
) -> ndarray:
lo, hi = continuous_range
assert hi >= lo
t = (t - lo) / (hi - lo)
t *= num_discrete
return np.clip(t.round(), 0, num_discrete - 1).astype(np.int64)
def undiscretize(
t: ndarray,
continuous_range: Tuple[float, float],
num_discrete: int,
) -> ndarray:
lo, hi = continuous_range
assert hi >= lo
t = t.astype(np.float32) + 0.5
t /= num_discrete
return t * (hi - lo) + lo
|