Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,881 Bytes
f499d3b |
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 |
import numpy as np
from numpy import ndarray
from typing import Dict, Tuple, Union, List
from .spec import TokenizerSpec, TokenizeInput, DetokenzeOutput, TokenizerConfig
from .spec import make_skeleton
from ..data.order import get_order
class TokenizerPart(TokenizerSpec):
def __init__(
self,
config: TokenizerConfig,
):
super().__init__()
self._num_discrete = config.num_discrete
self._continuous_range = config.continuous_range
self.cls_token_id = config.cls_token_id.copy()
self.parts_token_id = config.parts_token_id.copy()
self.order = get_order(config.order_config)
_offset = config.num_discrete
self.token_id_branch = _offset + 0
self.token_id_bos = _offset + 1
self.token_id_eos = _offset + 2
self.token_id_pad = _offset + 3
_offset += 4
self.token_id_spring = _offset + 0
_offset += 1
assert None not in self.parts_token_id
for i in self.parts_token_id:
self.parts_token_id[i] += _offset
_offset += len(self.parts_token_id)
self.token_id_cls_none = _offset + 0
_offset += 1
for i in self.cls_token_id:
self.cls_token_id[i] += _offset
_offset += len(self.cls_token_id)
self._vocab_size = _offset
self.parts_token_id_name = [x for x in self.parts_token_id]
self.part_token_to_name = {v: k for k, v in self.parts_token_id.items()}
assert len(self.part_token_to_name) == len(self.parts_token_id), 'names with same token found in parts_token_id'
self.part_token_to_name[self.token_id_spring] = None
self.cls_token_to_name = {v: k for k, v in self.cls_token_id.items()}
assert len(self.cls_token_to_name) == len(self.cls_token_id), 'names with same token found in cls_token_id'
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 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)
tails = discretize(t=input.tails, continuous_range=self.continuous_range, num_discrete=self.num_discrete)
branch = input.branch
is_leaf = input.is_leaf
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])
use_leaf = False
for i in range(num_bones):
# add parts token id
if i in input.parts_bias:
part = input.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) -> DetokenzeOutput:
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
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:
names = self.order.make_names(cls=cls, parts=parts, num_bones=num_bones)
else:
names = [f"bone_{i}" for i in range(num_bones)]
return DetokenzeOutput(
tokens=ids,
parents=parents,
bones=bones,
tails=tails,
no_skin=None,
cls=cls,
parts=parts,
names=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
@property
def num_discrete(self):
return self._num_discrete
@property
def continuous_range(self) -> Tuple[float, float]:
return self._continuous_range
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
|