File size: 8,006 Bytes
4ec5e47 | 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 | import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers.modeling_outputs import (
BaseModelOutput,
MaskedLMOutput,
)
from transformers.models.fnet.modeling_fnet import (
FNetEmbeddings,
FNetIntermediate,
FNetOnlyMLMHead,
FNetOutput,
FNetPreTrainedModel,
)
from .configuration_splinet import SpliNetConfig
class SplineMixer(nn.Module):
def __init__(self, config):
super().__init__()
self.radius = int(
config.splinet_radius
)
z = -3.0 + 2.0 * math.sqrt(2.0)
# Fixed analytic spline coefficients.
#
# IMPORTANT:
# Do NOT register these as a non-persistent tensor buffer.
# Hugging Face low-memory/meta-device loading can materialize
# such a buffer without its analytically initialized values.
#
# Store the 33 coefficients as ordinary Python floats instead.
# They are recreated on the actual input device and dtype in
# forward(). They are not learned model state.
self.kernel_values = tuple(
float(
math.sqrt(2.0)
* (z ** abs(k))
)
for k in range(
-self.radius,
self.radius + 1,
)
)
def forward(self, x):
if x.shape[1] != 512:
raise ValueError(
"This released SpliNet checkpoint was pretrained "
"and validated on fixed 512-token blocks. "
f"Received sequence length {x.shape[1]}. "
"Tokenize/pack the input to exactly 512 tokens."
)
d = x.shape[-1]
y = (
x.transpose(1, 2)
.contiguous()
)
y = F.pad(
y,
(
self.radius,
self.radius,
),
mode="reflect",
)
kernel = (
torch.tensor(
self.kernel_values,
device=y.device,
dtype=y.dtype,
)
.view(1, 1, -1)
.expand(d, 1, -1)
.contiguous()
)
y = F.conv1d(
y,
kernel,
groups=d,
)
return (
y.transpose(1, 2)
.contiguous()
)
class SpliNetMixingBlock(nn.Module):
def __init__(self, config):
super().__init__()
self.mixer = SplineMixer(
config
)
self.LayerNorm = nn.LayerNorm(
config.hidden_size,
eps=config.layer_norm_eps,
)
def forward(self, x):
return self.LayerNorm(
x + self.mixer(x)
)
class SpliNetLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.mixing = (
SpliNetMixingBlock(
config
)
)
self.intermediate = (
FNetIntermediate(
config
)
)
self.output = (
FNetOutput(
config
)
)
def forward(self, x):
x = self.mixing(x)
return self.output(
self.intermediate(x),
x,
)
class SpliNetEncoder(nn.Module):
def __init__(self, config):
super().__init__()
self.layer = nn.ModuleList(
[
SpliNetLayer(config)
for _ in range(
config.num_hidden_layers
)
]
)
def forward(
self,
x,
output_hidden_states=False,
):
hidden_states = (
()
if output_hidden_states
else None
)
for layer in self.layer:
if output_hidden_states:
hidden_states += (x,)
x = layer(x)
if output_hidden_states:
hidden_states += (x,)
return BaseModelOutput(
last_hidden_state=x,
hidden_states=hidden_states,
)
class SpliNetModel(FNetPreTrainedModel):
config_class = SpliNetConfig
base_model_prefix = "splinet"
def __init__(self, config):
super().__init__(config)
self.embeddings = (
FNetEmbeddings(config)
)
self.encoder = (
SpliNetEncoder(config)
)
self.post_init()
def get_input_embeddings(self):
return (
self.embeddings
.word_embeddings
)
def set_input_embeddings(
self,
value,
):
self.embeddings.word_embeddings = value
def forward(
self,
input_ids=None,
token_type_ids=None,
position_ids=None,
inputs_embeds=None,
output_hidden_states=False,
**kwargs,
):
if input_ids is not None:
shape = input_ids.shape
device = input_ids.device
elif inputs_embeds is not None:
shape = inputs_embeds.shape[:-1]
device = inputs_embeds.device
else:
raise ValueError(
"input_ids or inputs_embeds required."
)
if shape[1] != 512:
raise ValueError(
"This SpliNet checkpoint requires exactly "
f"512 tokens; received {shape[1]}."
)
if token_type_ids is None:
token_type_ids = torch.zeros(
shape,
dtype=torch.long,
device=device,
)
x = self.embeddings(
input_ids=input_ids,
token_type_ids=token_type_ids,
position_ids=position_ids,
inputs_embeds=inputs_embeds,
)
return self.encoder(
x,
output_hidden_states=output_hidden_states,
)
class SpliNetForMaskedLM(FNetPreTrainedModel):
config_class = SpliNetConfig
base_model_prefix = "splinet"
_tied_weights_keys = {
"cls.predictions.decoder.bias":
"cls.predictions.bias",
"cls.predictions.decoder.weight":
"splinet.embeddings.word_embeddings.weight",
}
def __init__(self, config):
super().__init__(config)
self.splinet = (
SpliNetModel(config)
)
self.cls = (
FNetOnlyMLMHead(config)
)
self.post_init()
if config.tie_word_embeddings:
(
self.cls
.predictions
.decoder.weight
) = (
self.splinet
.embeddings
.word_embeddings
.weight
)
def get_input_embeddings(self):
return (
self.splinet
.embeddings
.word_embeddings
)
def set_input_embeddings(
self,
value,
):
(
self.splinet
.embeddings
.word_embeddings
) = value
def get_output_embeddings(self):
return (
self.cls
.predictions
.decoder
)
def set_output_embeddings(
self,
value,
):
self.cls.predictions.decoder = value
def forward(
self,
input_ids=None,
labels=None,
**kwargs,
):
out = self.splinet(
input_ids=input_ids,
**kwargs,
)
logits = self.cls(
out.last_hidden_state
)
loss = None
if labels is not None:
loss = F.cross_entropy(
logits.reshape(
-1,
self.config.vocab_size,
),
labels.reshape(-1),
ignore_index=-100,
)
return MaskedLMOutput(
loss=loss,
logits=logits,
hidden_states=out.hidden_states,
)
|