File size: 9,006 Bytes
e308d04 |
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 |
import json
import os
from typing import List, Optional, Union
import numpy as np
import torch
from transformers import PreTrainedTokenizer
class BinnedOmicTokenizer(PreTrainedTokenizer):
def __init__(
self,
n_expressions_bins: int = 64,
min_omic_value: float = 0.0,
max_omic_value: float = 1.0,
use_max_normalization: bool = True,
normalization_factor: float = 1.0,
prepend_cls_token: bool = False,
fixed_sequence_length: Optional[int] = None,
unpadded_length: Optional[int] = None,
**kwargs,
):
bin_tokens = [str(i) for i in range(n_expressions_bins)]
special_tokens = ["<pad>", "<mask>", "<cls>"]
vocab = {tok: i for i, tok in enumerate(bin_tokens)}
offset = len(vocab)
for i, tok in enumerate(special_tokens):
vocab[tok] = offset + i
ids_to_tokens = {i: tok for tok, i in vocab.items()}
self.vocab = vocab
self.ids_to_tokens = ids_to_tokens
self.n_expressions_bins = n_expressions_bins
self.min_omic_value = min_omic_value
self.max_omic_value = max_omic_value
self.use_max_normalization = use_max_normalization
self.normalization_factor = normalization_factor
self.prepend_cls_token = prepend_cls_token
self.fixed_sequence_length = fixed_sequence_length
self.unpadded_length = unpadded_length
self.bin_edges = np.linspace(min_omic_value, max_omic_value, n_expressions_bins)
self.pad_token = "<pad>"
self.mask_token = "<mask>"
self.cls_token = "<cls>"
super().__init__(**kwargs)
self.add_special_tokens(
{
"pad_token": "<pad>",
"mask_token": "<mask>",
"cls_token": "<cls>",
"unk_token": "<pad>",
}
)
def _convert_token_to_id(self, token: str) -> int:
return self.vocab.get(token, self.vocab[self.unk_token])
def _convert_id_to_token(self, index: int) -> str:
return self.ids_to_tokens.get(index, self.unk_token)
def get_vocab(self) -> dict:
return self.vocab
def _tokenize(self, text, **kwargs):
raise NotImplementedError("Use `encode` or `batch_encode_plus` methods.")
def decode(self, token_ids, **kwargs):
return [self._convert_id_to_token(i) for i in token_ids]
def encode(
self,
gene_expr: Union[np.ndarray, List[float]],
pad_to_fixed_length: bool = False,
max_length: Optional[int] = None,
return_tensors: Optional[str] = None,
**kwargs,
) -> Union[List[int], torch.Tensor]:
gene_expr = np.array(gene_expr)
if self.use_max_normalization:
gene_expr = gene_expr / self.normalization_factor
token_ids = np.digitize(gene_expr, self.bin_edges).astype(int)
token_ids[gene_expr == 0.0] = 0
if self.prepend_cls_token:
token_ids = np.concatenate([[self.cls_token_id], token_ids])
if pad_to_fixed_length:
current_max_length = self.fixed_sequence_length or max_length
if current_max_length is None:
raise ValueError("fixed_sequence_length or max_length must be set.")
pad_len = current_max_length - len(token_ids)
if pad_len > 0:
token_ids = np.concatenate([token_ids, [self.pad_token_id] * pad_len])
else:
token_ids = token_ids[:current_max_length]
if return_tensors == "pt":
return torch.tensor(token_ids).unsqueeze(0)
return token_ids.tolist() # type: ignore
def batch_encode_plus(
self,
batch_gene_expr: Union[np.ndarray, List[np.ndarray]],
pad_to_fixed_length: bool = False,
max_length: Optional[int] = None,
return_tensors: Optional[str] = None,
**kwargs,
):
if isinstance(batch_gene_expr, list):
batch_gene_expr = np.array(batch_gene_expr)
encoded = [
self.encode(
gene_expr,
pad_to_fixed_length=pad_to_fixed_length,
max_length=max_length,
return_tensors=None,
**kwargs,
)
for gene_expr in batch_gene_expr
]
encoded = np.array(encoded, dtype=np.int64)
if return_tensors == "pt":
return {"input_ids": torch.tensor(encoded)}
return {"input_ids": encoded}
@property
def vocab_size(self) -> int:
return len(self.vocab)
def save_vocabulary(
self, save_directory: str, filename_prefix: Optional[str] = None
):
vocab_file = os.path.join(
save_directory,
(filename_prefix + "-" if filename_prefix else "") + "vocab.json",
)
with open(vocab_file, "w") as f:
json.dump(self.vocab, f)
return (vocab_file,)
class MOJOTokenizer(PreTrainedTokenizer):
def __init__(
self,
n_expressions_bins: dict[str, int],
min_omic_value: dict[str, float],
max_omic_value: dict[str, float],
use_max_normalization: dict[str, bool],
normalization_factor: dict[str, float],
prepend_cls_token: bool,
fixed_sequence_length: int,
unpadded_length: int,
**kwargs,
):
self.omics = n_expressions_bins.keys()
self.omic_tokenizers = {
omic: BinnedOmicTokenizer(
n_expressions_bins=n_expressions_bins[omic],
min_omic_value=min_omic_value[omic],
max_omic_value=max_omic_value[omic],
use_max_normalization=use_max_normalization[omic],
normalization_factor=normalization_factor[omic],
prepend_cls_token=prepend_cls_token,
fixed_sequence_length=fixed_sequence_length,
unpadded_length=unpadded_length,
**kwargs,
)
for omic in n_expressions_bins.keys()
}
self.vocab = {omic: self.omic_tokenizers[omic].vocab for omic in self.omics}
self.ids_to_tokens = {
omic: self.omic_tokenizers[omic].ids_to_tokens for omic in self.omics
}
super().__init__(**kwargs)
def _convert_token_to_id(self, token: dict[str, str]) -> dict[str, int]:
return {
omic: self.vocab[omic].get(token[omic], self.vocab[omic][self.unk_token])
for omic in token
}
def _convert_id_to_token(self, index: dict[str, int]) -> dict[str, str]:
return {
omic: self.omic_tokenizers[omic]._convert_id_to_token(index[omic])
for omic in index
}
def get_vocab(self) -> dict:
return self.vocab
def _tokenize(self, text, **kwargs):
raise NotImplementedError("Use `encode` or `batch_encode_plus` methods.")
def decode(self, token_ids: dict[str, list[int]], **kwargs):
return {
omic: self.omic_tokenizers[omic].decode(token_ids[omic])
for omic in token_ids
}
def encode(
self,
omic_array: Union[dict[str, np.ndarray], dict[str, List[float]]],
pad_to_fixed_length: bool = False,
max_length: Optional[int] = None,
return_tensors: Optional[str] = None,
**kwargs,
) -> Union[dict[str, List[int]], dict[str, torch.Tensor]]:
return {
omic: self.omic_tokenizers[omic].encode(
omic_array[omic],
pad_to_fixed_length=pad_to_fixed_length,
max_length=max_length,
return_tensors=return_tensors,
)
for omic in omic_array
}
def batch_encode_plus(
self,
batch_omic_array: Union[dict[str, np.ndarray], dict[str, List[np.ndarray]]],
pad_to_fixed_length: bool = False,
max_length: Optional[int] = None,
return_tensors: Optional[str] = None,
**kwargs,
):
return {
omic: self.omic_tokenizers[omic].batch_encode_plus(
batch_omic_array[omic],
pad_to_fixed_length=pad_to_fixed_length,
max_length=max_length,
return_tensors=return_tensors,
)
for omic in batch_omic_array
}
@property
def vocab_size(self) -> int:
return sum(len(self.vocab[omic]) for omic in self.vocab)
def save_vocabulary(
self, save_directory: str, filename_prefix: Optional[str] = None
):
vocab_files = []
for omic in self.omics:
vocab_file = os.path.join(
save_directory,
(filename_prefix + "-" if filename_prefix else "")
+ f"vocab_{omic}.json",
)
with open(vocab_file, "w") as f:
json.dump(self.vocab[omic], f)
vocab_files.append(vocab_file)
return tuple(vocab_files)
|