File size: 22,142 Bytes
663494c |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
class LoRALinear(nn.Module):
"""
LoRA layer: Low-Rank Adaptation.
This layer consists of a low-rank decomposition of weight updates.
"""
def __init__(self, in_features, out_features, r=8, alpha=1.0, dropout=0.1, **kwargs):
super(LoRALinear, self).__init__()
self.use_si = False
self.multi_domain = 0
if 'use_si' in kwargs.keys():
self.model = LoRALinearSI(
in_features, out_features, r, alpha, **kwargs
)
self.use_si = True
elif 'multi_domain' in kwargs.keys():
self.r = r
self.alpha = alpha
self.multi_domain = kwargs['multi_domain']
a_list, b_list, drop_list = [], [], []
for i in range(self.multi_domain):
a_list.append(nn.Linear(in_features, r, bias=False))
b_list.append(nn.Linear(r, out_features, bias=False))
drop_list.append(nn.Dropout(dropout))
self.A = nn.ModuleList(a_list)
self.B = nn.ModuleList(b_list)
self.drop =nn.ModuleList(drop_list)
self.scaling = alpha / r
self._init_weights()
else:
self.r = r
self.alpha = alpha
# Low-rank decomposition matrices
self.A = nn.Linear(in_features, r, bias=False) # Down-projection
self.drop = nn.Dropout(dropout)
self.B = nn.Linear(r, out_features, bias=False) # Up-projection
nn.init.zeros_(self.B.weight)
nn.init.normal_(self.A.weight, std=1 / r)
self.lora_name = "lora_layer" # Unique name
# Scaling factor for LoRA
self.scaling = alpha / r
def _init_weights(self):
for layer in self.A:
nn.init.normal_(layer.weight, std=1 / self.r)
for layer in self.B:
nn.init.zeros_(layer.weight)
def forward(self, x, task_mask=None, i=None,task_idx=None):
# Apply low-rank update: scaling * (A(x) * B)
if self.use_si:
return self.model(x)
return self.scaling * self.B(self.drop(self.A(x)))
def update_si_information(self):
if self.use_si:
self.model.update_si_information()
def finalize_si_importance(self):
if self.use_si:
self.model.finalize_si_importance()
class BayesianLinear(nn.Module):
def __init__(self, in_features, out_features, r=8, prior_std=0.1, dropout=0.1, **kwargs):
"""
Bayesian LoRA Layer: Instead of deterministic weights,
it learns a distribution over LoRA parameters using Bayesian inference.
Args:
in_features (int): Input dimension.
out_features (int): Output dimension.
rank (int): LoRA rank.
prior_std (float): Standard deviation of the Gaussian prior.
"""
super(BayesianLinear, self).__init__()
# Learnable means and log-variances (for stability)
self.scaling = 1 / r
self.A_mu = nn.Parameter(torch.randn(in_features, r) * (1 / r))
self.A_logvar = nn.Parameter(torch.randn(in_features, r) * (1 / r))
self.B_mu = nn.Parameter(torch.randn(r, out_features) * (1 / r))
self.B_logvar = nn.Parameter(torch.randn(r, out_features) * (1 / r))
self.drop = nn.Dropout(dropout)
# Gaussian prior (zero mean)
self.prior_std = prior_std
def sample_weights(self):
"""
Reparameterization Trick: Sample weights from Gaussian distribution.
"""
A_std = torch.exp(0.5 * self.A_logvar)
B_std = torch.exp(0.5 * self.B_logvar)
# Sample weights using reparameterization
B_sample = self.B_mu + B_std * torch.randn_like(B_std)
A_sample = self.A_mu + A_std * torch.randn_like(A_std)
return A_sample, B_sample
# def kl_divergence(self):
# """
# Compute KL divergence between learned weight distributions and the prior.
# """
# W_std = torch.exp(0.5 * self.W_logvar)
# A_std = torch.exp(0.5 * self.A_logvar)
# kl_W = (self.W_mu**2 + W_std**2 - 2 * torch.log(W_std) - 1).sum()
# kl_A = (self.A_mu**2 + A_std**2 - 2 * torch.log(A_std) - 1).sum()
# return 0.5 * (kl_W + kl_A)
def forward(self, x):
"""
Forward pass with Bayesian weight sampling.
"""
if self.training:
A, B = self.sample_weights()
else:
A, B = self.A_mu, self.B_mu # Use deterministic weights for testing
out = self.drop(x @ A)
return out @ B # LoRA forward pass
class LoRALinearSI(nn.Module):
def __init__(self, in_features, out_features, r=8,
alpha=1.0, lambda_si=0.1, si_decay=0.99, dropout=0.1,
plasticity_base=0.5, sparsity_threshold=1e-3):
super().__init__()
self.r = r
self.alpha = alpha # Base scaling factor for LoRA updates
self.lambda_si = lambda_si # Strength of SI regularization
self.si_decay = si_decay # Decay factor for importance updates
self.plasticity_base = plasticity_base # Base plasticity level
self.sparsity_threshold = sparsity_threshold # Threshold for detecting sparse weights
# LoRA trainable parameters
self.lora_A = nn.Parameter(torch.randn(in_features, r))
self.lora_B = nn.Parameter(torch.randn(r, out_features))
self.drop = nn.Dropout(dropout)
nn.init.zeros_(self.lora_B)
nn.init.normal_(self.lora_A, std=1 / r)
# Synaptic Intelligence (SI) buffers
self.register_buffer("omega_A", torch.zeros_like(self.lora_A)) # Importance of lora_A
self.register_buffer("omega_B", torch.zeros_like(self.lora_B)) # Importance of lora_B
self.register_buffer("prev_params_A", self.lora_A.clone().detach())
self.register_buffer("prev_params_B", self.lora_B.clone().detach())
self.register_buffer("trajectory_A", torch.zeros_like(self.lora_A)) # Tracks updates for lora_A
self.register_buffer("trajectory_B", torch.zeros_like(self.lora_B)) # Tracks updates for lora_B
# self.register_buffer("plasticity", torch.ones_like(self.lora_A) * self.plasticity_base) # Dynamic plasticity control
def forward(self, x):
adaptive_alpha = self.alpha #* self.plasticity # Scale LoRA update based on plasticity
lora_update = torch.matmul(x, self.lora_A)
lora_update = self.drop(lora_update)
lora_update = torch.matmul(lora_update, self.lora_B)
return adaptive_alpha * lora_update # Dynamic scaling
def update_si_information(self):
"""Update Synaptic Intelligence importance online."""
if self.lora_A.grad is not None:
delta_theta_A = self.lora_A - self.prev_params_A
self.trajectory_A += delta_theta_A * self.lora_A.grad # Path integral for A
self.prev_params_A = self.lora_A.detach().clone()
if self.lora_B.grad is not None:
delta_theta_B = self.lora_B - self.prev_params_B
self.trajectory_B += delta_theta_B * self.lora_B.grad # Path integral for B
self.prev_params_B = self.lora_B.detach().clone()
def compute_sparsity(self, param):
"""Compute the sparsity score: fraction of near-zero values."""
return torch.mean((torch.abs(param) < self.sparsity_threshold).float())
def finalize_si_importance(self):
"""Compute final importance after training a task and adjust plasticity."""
self.omega_A = self.si_decay * self.omega_A + (1 - self.si_decay) * (self.trajectory_A / (self.lora_A**2 + 1e-6)).detach()
self.omega_B = self.si_decay * self.omega_B + (1 - self.si_decay) * (self.trajectory_B / (self.lora_B**2 + 1e-6)).detach()
self.trajectory_A.zero_()
self.trajectory_B.zero_()
# Compute sparsity scores
# sparsity_A = self.compute_sparsity(self.lora_A)
# sparsity_B = self.compute_sparsity(self.lora_B)
# Adjust plasticity dynamically based on sparsity
# self.plasticity = torch.exp(-self.omega_A) * (1 - sparsity_A)
def si_loss(self):
"""Compute the SI loss term for both LoRA parameters."""
loss_A = torch.sum(self.omega_A * (self.lora_A - self.prev_params_A) ** 2)
loss_B = torch.sum(self.omega_B * (self.lora_B - self.prev_params_B) ** 2)
return self.lambda_si * (loss_A + loss_B)
def set_plasticity(self, value: float):
"""Manually set a global plasticity value if needed."""
self.plasticity.fill_(value)
class MOELoRALinear(nn.Module):
"""
LoRA layer: Low-Rank Adaptation.
This layer consists of a low-rank decomposition of weight updates.
"""
def __init__(self, in_features, out_features, r=8, alpha=1.0, dropout=0.1, num_task=3, **kwargs):
super(MOELoRALinear, self).__init__()
self.loras = nn.ModuleList([
LoRALinear(
in_features,
out_features,
r, alpha, dropout, **kwargs) for _ in range(num_task)
])
self.num_task=num_task
def forward(self, x, i):
if isinstance(i, int):
return self.loras[i](x)
elif i.dtype == torch.float:
orig_shape = x.shape
b = orig_shape[0]
new_shape = (b//self.num_task, self.num_task) + orig_shape[1:]
x = x.reshape(new_shape)
mask_shape = i.shape + (1,)*len(orig_shape[1:])
i = i.reshape(mask_shape)
res_list = torch.stack([
self.loras[t](x[:, t]) for t in range(self.num_task)
], dim=1) #[b, task, class, dim]
res_list = res_list * i
res_list = res_list.reshape(orig_shape[:-1]+(-1,))
return res_list
res_list = torch.stack([
self.loras[t](x) for t in range(self.num_task)
], dim=1) #[b, task, class, dim]
b = res_list.shape[0]
res = res_list[torch.arange(b), i]
# print(res.shape, i.shape)
return res
class ZeroAdapter(nn.Module):
"""
LoRA layer: Low-Rank Adaptation.
This layer consists of multiple LoRA mitigating catastrophic forgetting
"""
def __init__(self, in_features, out_feature, dropout=0.1, **kwargs):
super(ZeroAdapter, self).__init__()
mid_feature = in_features // 2
self.down_linear = nn.Linear(in_features, mid_feature)
self.up_linear = nn.Linear(mid_feature, out_feature)
nn.init.zeros_(self.down_linear.weight)
nn.init.zeros_(self.down_linear.bias)
nn.init.zeros_(self.up_linear.weight)
nn.init.zeros_(self.up_linear.bias)
self.act = nn.ReLU()
self.drop = nn.Dropout(dropout)
self.lora_name = "lora_layer" # Unique name
def forward(self, x):
x = self.down_linear(x)
x = self.drop(self.act(x))
x = self.up_linear(x)
return x
class LoRAMoECLAdapter(nn.Module):
def __init__(self, in_features, mid_feature, out_feature,
num_task=6, r=8, alpha=1.0, dropout=0.1, **kwargs):
super(LoRAMoECLAdapter, self).__init__()
self.r = r
self.alpha = alpha
self.num_task = num_task
self.adapters = nn.ModuleList([
nn.Sequential(
LoRALinear(in_features, mid_feature, r, alpha, dropout),
nn.Dropout(dropout),
nn.ReLU(),
LoRALinear(mid_feature, out_feature, r, alpha, dropout),
)
for _ in range(num_task)
])
self.router = nn.Linear(in_features, num_task)
self.out_drop = nn.Dropout(dropout)
self.lora_name = "lora_layer" # Unique name
def forward(self, x, i=None):
outputs = []
logits = self.router(x)
route_prob = logits.softmax(-1)
for i in range(self.num_task):
outputs.append(self.adapters[i](x))
outputs = torch.stack(outputs, dim=-2)
outputs = torch.sum(outputs * route_prob[..., None], dim=-2)
outputs = self.out_drop(outputs)
return outputs
class LoRACLAdapter(nn.Module):
"""
LoRA layer: Low-Rank Adaptation.
This layer consists of multiple LoRA mitigating catastrophic forgetting
"""
def __init__(self, in_features, out_feature,
num_task=6, r=8, alpha=1.0, dropout=0.1, **kwargs):
super(LoRACLAdapter, self).__init__()
self.r = r
self.alpha = alpha
self.loras = nn.ModuleList([
LoRALinear(in_features, out_feature, r, alpha, dropout) for _ in range(num_task)
])
self.attn_weights = nn.ModuleList([nn.Linear(out_feature, 1) for _ in range(num_task)])
self.attn_drop = nn.Dropout(dropout)
self.num_task = num_task
# Scaling factor for LoRA
self.scaling = alpha / r
self.lora_name = "lora_layer" # Unique name
def forward(self, x, task_mask=None):
# Apply low-rank update: scaling * (A(x) * B)
#x:[b, 1, d]
assert task_mask is not None
outputs = []
output_weights = []
for i in range(self.num_task):
out = self.loras[i](x)
weight_out = self.attn_weights[i](out)
outputs.append(out)
output_weights.append(weight_out)
outputs = torch.cat(outputs, dim=1)
output_weights = torch.cat(output_weights, dim=1)
output_weights = output_weights.softmax(1)
outputs = outputs * self.attn_drop(output_weights)
# detach invalid outputs:
task_mask = task_mask[0]
task_mask = task_mask.unsqueeze(-1).expand(outputs.shape[0], -1, outputs.shape[2])
# print(task_mask.shape, outputs.shape)
outputs[task_mask==0] = outputs[task_mask==0].detach()
outputs = outputs.sum(1)
return outputs[:, None]
valid_lora_list = (LoRALinear, LoRACLAdapter, ZeroAdapter, LoRAMoECLAdapter, MOELoRALinear)
def lora_wrapper(
module,
LoraLayer=LoRALinear,
rank=8, alpha=1.0, dropout=0.1,
num_task=6,
**kwargs):
"""
Creates a separate LoRA module that mirrors the Linear layers in the original model.
"""
if isinstance(module, nn.ModuleList):
lora_module = nn.ModuleList()
for m in module:
lora_module.append(lora_wrapper(
m, LoraLayer,
rank=rank, alpha=alpha, dropout=dropout,num_task=num_task
))
return lora_module
if isinstance(module, nn.ModuleDict):
lora_module = nn.ModuleDict()
for k,v in module.items():
lora_module[f'lora_{k}'] = lora_wrapper(
v, LoraLayer,
rank=rank, alpha=alpha, dropout=dropout,num_task=num_task
)
return lora_module
if len(list(module.named_modules())) == 1 :
if not isinstance(module, nn.Linear):
print(f'Wrap non nn.Linear unit{type(module)}, skipping with Identity')
return nn.Identity()
lora_module = LoraLayer(module.in_features, module.out_features,
r=rank, alpha=alpha,dropout=dropout, num_task=num_task,**kwargs)
return lora_module
# sequential case
lora_module = nn.Sequential()
for name, child in module.named_children():
if isinstance(child, nn.Linear):
lora_layer = LoraLayer(child.in_features, child.out_features,
r=rank, alpha=alpha,dropout=dropout, num_task=num_task, **kwargs)
lora_module.add_module(f'lora_{name}', lora_layer)
elif isinstance(child, nn.Sequential):
lora_module.add_module(f'lora_{name}',
lora_wrapper(child,
LoraLayer,
rank=rank, alpha=alpha, dropout=dropout,num_task=num_task,
)
)
else:
lora_module.add_module(f'lora_{name}', nn.Identity())
return lora_module
def single_peft_forward(x, model, lora_model, lora_only=False, idx=None):
if lora_only:
return lora_model(x, i=idx)
return model(x) + lora_model(x, i=idx)
def peft_wrapper_forward(x, model, lora_model, use_lora=True,
layer_idx=-1, layer_name="", lora_only=False, task_idx=None):
"""
Custom forward function to combine original model output with LoRA output.
layer_idx: can be specified for (nn.ModuleList) model; Default: running sequentially through whole ModuleList
layer_name: can be specified for (nn.ModuleDict) model; Default:running sequentially through whole ModuleDict
lora_only: if lora_only=True, forward function will only pass through the lora layer when meet with matched Linear
"""
if isinstance(model, nn.ModuleList):
if layer_idx > -1:
return single_peft_forward(x, model[layer_idx], lora_model[layer_idx], lora_only, task_idx)
if isinstance(model, nn.ModuleDict):
if layer_name != "":
return single_peft_forward(x, model[layer_name], lora_model[layer_name], lora_only, task_idx)
if len(list(model.named_modules())) == 1:
return single_peft_forward(x, model, lora_model, lora_only, task_idx)
def process_layer(orig_layer, lora_layer, x):
""" Recursively process nested nn.Sequential layers """
if isinstance(orig_layer, nn.Sequential) and isinstance(lora_layer, nn.Sequential):
for o_layer, l_layer in zip(orig_layer.children(), lora_layer.children()):
x = process_layer(o_layer, l_layer, x)
return x
else:
if use_lora and not isinstance(lora_layer, nn.Identity):
return single_peft_forward(x, orig_layer, lora_layer, lora_only, task_idx)
else:
return orig_layer(x)
for orig_layer, lora_layer in zip(model.children(), lora_model.children()):
x = process_layer(orig_layer, lora_layer, x)
return x
def finetuning_detach(model):
'''
work with a detach for customed layer
ensure if some sublayer inside containing such LoRA layer
or adapter with "lora_name" attribute,
also have this finetuning function and lora_name attr
'''
for name, module in model.named_modules():
if 'lora' in name:
for param in module.parameters():
param.requires_grad = True
else:
for param in module.parameters():
param.requires_grad = False # disable param
if isinstance(module, (nn.Dropout, nn.Dropout2d, nn.Dropout3d)):
module.eval()
def frozen_grad(model):
for param in model.parameters():
param.requires_grad = False
return model
class TestModule(nn.Module):
def __init__(self):
super(TestModule, self).__init__()
# self.model = nn.Sequential(
# nn.Linear(10, 20),
# nn.ReLU(),
# nn.Sequential(
# nn.Linear(20, 30),
# nn.ReLU(),
# nn.Linear(30, 40)
# )
# )
# self.model = nn.ModuleList([nn.Linear(10, 10) for _ in range(3)])
self.model = nn.ModuleDict()
for i in range(3):
self.model[f'{i}'] = nn.Linear(10,10)
self.lora_layer = lora_wrapper(
self.model,
ZeroAdapter,
rank=4, alpha=1.0)
def forward(self, x):
x = peft_wrapper_forward(x, self.model, self.lora_layer)
return x
def retreive_bayesian_lora_param(module):
'''
input, any nn.Module
searching for all Bayesian Lora param
return: lora_dict: Dict[sub_name: Dict['A_mu','B_mu','A_logvar','B_logvar']]
'''
lora_dict = {}
lora_list = set(['A_mu','B_mu','A_logvar','B_logvar'])
if isinstance(module, BayesianLinear):
lora_dict['.'] = dict()
for name,m in module.named_parameters():
lora_dict['.'][name] = m
return lora_dict
for name,m in module.named_parameters():
name_list = name.split('.')
if name_list[-2] in lora_list:
m_prefix = ".".join(name_list[:-2])
if m_prefix not in lora_dict:
lora_dict[m_prefix] = dict()
lora_dict[m_prefix][name.split('.')[-1]] = m
return lora_dict
def test_lora_si():
from time import time
import numpy as np
lora_model = LoRALinearSI(
256, 256, 16
)
t = []
for _ in range(10):
s = time()
x = torch.randn(2, 256)
y = lora_model(x)
loss = lora_model.si_loss()
t.append(time()-s)
print(loss, np.mean(t))
def test_kl_lora():
lora_layer = BayesianLinear(
32, 32, r=8
)
inputs = torch.randn(4, 10, 32)
out = lora_layer(inputs)
bayesian_params = retreive_bayesian_lora_param(lora_layer)
loss = 0.
for v_dict in bayesian_params.values():
print(v_dict.keys())
B_std = torch.exp(0.5 * v_dict['B_logvar'])
A_std = torch.exp(0.5 * v_dict['A_logvar'])
kl_B = (v_dict['B_mu']**2 + B_std**2 - 2 * torch.log(B_std) - 1).sum()
kl_A = (v_dict['A_mu']**2 + A_std**2 - 2 * torch.log(A_std) - 1).sum()
module_loss = 0.5 * (kl_B + kl_A)
loss += module_loss
print(out.shape, loss)
# Example usage
if __name__ == "__main__":
# Define a nested Sequential model
# model = TestModule()
# finetuning_detach(model)
# x = torch.randn(4, 10)
# print(model(x).shape)
# # Print the model structure after attaching LoRA layers
# print("Model structure after attaching LoRA layers:\n", model)
# for name, param in model.named_parameters():
# print(name, param.shape, param.requires_grad)
# test_lora_si()
test_kl_lora()
|