File size: 16,515 Bytes
5fee096 | 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 | """
@inproceedings{liu2025lora,
title={LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual Learning},
author={Liu, Xuan and Chang, Xiaobin},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year={2025}
}
Adapted from https://github.com/scarlet0703/LoRA-Sub-DRS
"""
import torch
import torch.nn as nn
import copy
import numpy as np
import math
from copy import deepcopy
from torch.optim.optimizer import Optimizer
from torch.nn import functional as F
from tqdm import tqdm
from collections import defaultdict
from scipy.spatial.distance import cdist
from .backbone.transformer import MultiHeadAttention_LoRA_Sub
class AugmentedTripletLoss(nn.Module):
def __init__(self, margin=1.0, norm=2):
super(AugmentedTripletLoss, self).__init__()
self.margin = margin
self.norm = norm
self.ranking_loss = nn.MarginRankingLoss(margin=margin)
def forward(self, inputs, targets, center):
device = (torch.device('cuda')
if inputs.is_cuda
else torch.device('cpu'))
n = inputs.size(0) # batch_size
# Compute pairwise distance, replace by the official when merged
dist = torch.pow(inputs, 2).sum(dim=1, keepdim=True).expand(n, n)
dist = dist + dist.t()
dist.addmm_(inputs, inputs.t(), beta=1, alpha=-2)
dist = dist.clamp(min=1e-12).sqrt() # for numerical stability
# For each anchor, find the hardest positive and negative
mask = targets.expand(n, n).eq(targets.expand(n, n).t())
num_proto = len(center)
dist_ap, dist_an = [], []
for i in range(n):
dist_ap.append(dist[i][mask[i]].max().unsqueeze(0))
if dist[i][mask[i] == 0].numel() == 0:
dist_an.append((dist[i][mask[i]].max()+self.margin).unsqueeze(0))
else:
dist_an.append(dist[i][mask[i] == 0].min().unsqueeze(0))
dist_ap = torch.cat(dist_ap)
if num_proto > 0:
center = torch.from_numpy(center / np.linalg.norm(center, axis=1)[:, None]).to(device)
for i in range(n):
for j in range(num_proto):
distp = torch.norm(inputs[i].unsqueeze(0) - center[j], self.norm).clamp(min=1e-12)
dist_an[i] = min(dist_an[i].squeeze(0), distp).unsqueeze(0)
dist_an = torch.cat(dist_an)
# Compute ranking hinge loss
y = torch.ones_like(dist_an)
loss = self.ranking_loss(dist_an, dist_ap, y)
return loss
class Adam(Optimizer):
r"""Implements Adam algorithm.
It has been proposed in `Adam: A Method for Stochastic Optimization`_.
Arguments:
params (iterable): iterable of parameters to optimize or dicts defining
parameter groups
lr (float, optional): learning rate (default: 1e-3)
betas (Tuple[float, float], optional): coefficients used for computing
running averages of gradient and its square (default: (0.9, 0.999))
eps (float, optional): term added to the denominator to improve
numerical stability (default: 1e-8)
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
amsgrad (boolean, optional): whether to use the AMSGrad variant of this
algorithm from the paper `On the Convergence of Adam and Beyond`_
(default: False)
.. _Adam\: A Method for Stochastic Optimization:
https://arxiv.org/abs/1412.6980
.. _On the Convergence of Adam and Beyond:
https://openreview.net/forum?id=ryQu7f-RZ
"""
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8, svd=False, thres=1.001,
weight_decay=0, amsgrad=False):
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {}".format(eps))
if not 0.0 <= betas[0] < 1.0:
raise ValueError(
"Invalid beta parameter at index 0: {}".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError(
"Invalid beta parameter at index 1: {}".format(betas[1]))
defaults = dict(lr=lr, betas=betas, eps=eps,
weight_decay=weight_decay, amsgrad=amsgrad, svd=svd,
thres=thres)
super(Adam, self).__init__(params, defaults)
self.eigens = defaultdict(dict)
self.transforms = defaultdict(dict)
def __setstate__(self, state):
super(Adam, self).__setstate__(state)
for group in self.param_groups:
group.setdefault('amsgrad', False)
group.setdefault('svd', False)
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
svd = group['svd']
for p in group['params']:
if p.grad is None:
continue
grad = p.grad.data
if grad.is_sparse:
raise RuntimeError(
'Adam does not support sparse gradients, please consider SparseAdam instead')
update = self.get_update(group, grad, p)
if svd and len(self.transforms) > 0:
if len(update.shape) == 4:
# the transpose of the manuscript
update_ = torch.mm(update.view(update.size(
0), -1), self.transforms[p]).view_as(update)
else:
if self.transforms[p].shape[0]==update.shape[0]:
update_ = torch.mm(self.transforms[p], update)
else:
update_ = torch.mm(update, self.transforms[p])
else:
update_ = update
p.data.add_(update_)
return loss
def get_transforms(self):
for group in self.param_groups:
svd = group['svd']
if svd is False:
continue
for p in group['params']:
thres = group['thres']
if p.requires_grad == False or thres == 1.0:
continue
eigen_values = self.eigens[p]['eigen_value']
cumulative_sum = eigen_values.cumsum(dim=0) / eigen_values.sum()
num_vectors = (cumulative_sum >= thres).nonzero(as_tuple=True)[0][0] + 1
print('reserving basis {}/{}; cond: {}, ratio:{}'.format(
num_vectors, eigen_values.shape[0],
eigen_values[0] / eigen_values[-1],
cumulative_sum[num_vectors - 1]
))
basis = self.eigens[p]['eigen_vector'][:, :num_vectors]
transform = torch.mm(basis, basis.transpose(1, 0))
self.transforms[p] = transform / torch.norm(transform)
self.transforms[p].detach_()
def get_eigens(self, fea_in):
for group in self.param_groups:
if group['svd']:
for p in group['params']:
if p.requires_grad:
eigen = self.eigens[p]
_, eigen_value, eigen_vector = torch.svd(fea_in[p], some=False)
eigen['eigen_value'] = eigen_value
eigen['eigen_vector'] = eigen_vector
def get_update(self, group, grad, p):
amsgrad = group['amsgrad']
state = self.state[p]
# State initialization
if len(state) == 0:
state['step'] = 0
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p.data)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p.data)
if amsgrad:
# Maintains max of all exp. moving avg. of sq. grad. values
state['max_exp_avg_sq'] = torch.zeros_like(p.data)
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
if amsgrad:
max_exp_avg_sq = state['max_exp_avg_sq']
beta1, beta2 = group['betas']
state['step'] += 1
if group['weight_decay'] != 0:
grad.add_(group['weight_decay'], p.data)
# Decay the first and second moment running average coefficient
exp_avg.mul_(beta1).add_(1 - beta1, grad)
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
if amsgrad:
# Maintains the maximum of all 2nd moment running avg. till now
torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
# Use the max. for normalizing running avg. of gradient
denom = max_exp_avg_sq.sqrt().add_(group['eps'])
else:
denom = exp_avg_sq.sqrt().add_(group['eps'])
bias_correction1 = 1 - beta1 ** state['step']
bias_correction2 = 1 - beta2 ** state['step']
step_size = group['lr'] * \
math.sqrt(bias_correction2) / bias_correction1
update = - step_size * exp_avg / denom
return update
class Model(nn.Module):
def __init__(self, backbone, device, **kwargs):
super().__init__()
self._cur_task_id = -1
self.backbone = backbone
self.device = device
self.classifier_pool = nn.ModuleList([
nn.Linear(kwargs["embd_dim"], kwargs['init_cls_num'], bias=True)] +
[nn.Linear(kwargs["embd_dim"], kwargs['inc_cls_num'], bias=True) for _ in range(kwargs['task_num'] - 1)]
)
def update_fc(self):
self._cur_task_id += 1
def update_input_matrix(self, x):
self.backbone(x, get_input_matrix = True)
def extract_features(self, x):
return self.backbone(x)
def forward(self, x):
logits = []
features = self.backbone(x)
for prompts in [self.classifier_pool[self._cur_task_id]]:
logits.append(prompts(features))
return {
'logits': torch.cat(logits, dim=1),
'features': features
}
class LoRAsub_DRS(nn.Module):
def __init__(self, backbone, device, **kwargs):
super().__init__()
self.device = device
self.init_cls_num = kwargs["init_cls_num"]
self.inc_cls_num = kwargs["inc_cls_num"]
self.task_num = kwargs["task_num"]
self.fc_lrate = kwargs["fc_lrate"]
self.margin_inter = kwargs["margin_inter"]
self.lambada = kwargs["lambada"]
self._known_classes = 0
self._total_classes = 0
self._cur_task = 0
self._network = Model(backbone, device, **kwargs)
self.attention_modules = [module for module in self._network.modules() if isinstance(module, MultiHeadAttention_LoRA_Sub)]
self.criterion = AugmentedTripletLoss(margin=self.margin_inter).to(self.device)
self._protos = []
def observe(self, data):
x, y = data['image'].to(self.device), data['label'].to(self.device) - self._known_classes
outputs = self._network(x)
logits, features = outputs['logits'], outputs['features']
ATL = self.criterion(
features / features.norm(dim=-1, keepdim=True),
y,
self._protos
)
loss = F.cross_entropy(logits, y) + self.lambada * ATL
preds = logits.max(1)[1]
correct_count = preds.eq(y).sum().item()
acc = correct_count / y.size(0)
return preds, acc, loss
def inference(self, data):
x, y = data['image'].to(self.device), data['label'].to(self.device)
features = self._network.extract_features(x)
features = (features.T / (np.linalg.norm(features.T, axis=0) + 1e-8)).T
class_means = self._protos / np.linalg.norm(self._protos, axis=1)[:, None]
dists = cdist(class_means, features, 'sqeuclidean')
scores = dists.T
#preds = np.argsort(scores, axis=1)[:, :1]
preds = np.argmin(scores, axis=1)
correct_count = (preds == y.cpu().numpy()).sum()
acc = correct_count / y.size(0)
return preds, acc
@torch.no_grad()
def before_task(self, task_idx, buffer, train_loader, test_loaders):
self._known_classes = self._total_classes
self._total_classes += self.init_cls_num if task_idx == 0 else self.inc_cls_num
self._network.update_fc()
self._network = self._network.to(self.device)
for module in self.attention_modules:
module.init_param()
unfrezeed_params = []
for name, param in self._network.named_parameters():
param.requires_grad_(False)
if f'classifier_pool.{self._cur_task}.' in name or \
f'lora'in name:
param.requires_grad_(True)
unfrezeed_params.append(name)
print(f"Current task : {task_idx}, Parameters to be updated: {len(unfrezeed_params)}")
if task_idx > 0:
for batch in tqdm(train_loader, desc="Forwarding to get input matrix"):
self._network.update_input_matrix(x = batch['image'].to(self.device))
self.fea_in = {}
for module in self.attention_modules:
self.fea_in[module.lora_A_k.weight] = deepcopy(module.cur_matrix).to(self.device)
self.fea_in[module.lora_A_v.weight] = deepcopy(module.cur_matrix).to(self.device)
self.fea_in[module.lora_B_k.weight] = deepcopy(module.cur_matrix).to(self.device)
self.fea_in[module.lora_B_v.weight] = deepcopy(module.cur_matrix).to(self.device)
module.reset_input_matrix()
@torch.no_grad()
def after_task(self, task_idx, buffer, train_loader, test_loaders):
for module in self.attention_modules:
module.save_weight()
# Build Proto
for class_idx in range(self._known_classes, self._total_classes):
inputs_list = []
for batch in train_loader:
x, y = batch['image'].to(self.device), batch['label'].to(self.device)
inputs_list.append(x[y == class_idx])
class_inputs = torch.cat(inputs_list, dim=0)
features_list = []
for start_idx in range(0, class_inputs.shape[0], 128):
end_idx = min(start_idx + 128, class_inputs.shape[0])
batch_inputs = class_inputs[start_idx:end_idx].to(self.device)
feats = self._network.extract_features(batch_inputs)
features_list.append(feats.detach().cpu().numpy())
features = np.concatenate(features_list, axis=0)
class_mean = np.mean(features, axis=0)
self._protos.append(class_mean)
assert len(self._protos) > 0
self._known_classes += self.init_cls_num if task_idx == 0 else self.inc_cls_num
self._cur_task += 1
def get_parameters(self, config):
return self._network.parameters()
def get_optimizer(self, lr, weight_decay):
fea_params = []
for module in self.attention_modules:
fea_params.append(module.lora_A_k.weight)
fea_params.append(module.lora_A_v.weight)
fea_params.append(module.lora_B_k.weight)
fea_params.append(module.lora_B_v.weight)
cls_params = [
self._network.classifier_pool[self._cur_task].weight,
self._network.classifier_pool[self._cur_task].bias,
]
model_optimizer_arg = {'params': [{'params': fea_params, 'svd': True, 'lr': lr,
'thres': 0.99},
{'params': cls_params, 'weight_decay': weight_decay,
'lr': self.fc_lrate}],
'weight_decay': weight_decay,
'betas': (0.9, 0.999)
}
optim = Adam(**model_optimizer_arg)
if self._cur_task > 0:
optim.get_eigens(self.fea_in)
optim.get_transforms()
return optim |