File size: 8,093 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 | """
@misc{wu2025sdlorascalabledecoupledlowrank,
title={SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning},
author={Yichen Wu and Hongming Piao and Long-Kai Huang and Renzhen Wang and Wanhua Li and Hanspeter Pfister and Deyu Meng and Kede Ma and Ying Wei},
year={2025},
eprint={2501.13198},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2501.13198},
}
Adapted from https://github.com/WuYichen-97/SD-Lora-CL
"""
import torch
import torch.nn as nn
import copy
import numpy as np
from torch.nn import functional as F
from .backbone.transformer import MultiHeadAttention_SDLoRA
class Model(nn.Module):
def __init__(self, backbone, device, **kwargs):
super().__init__()
self._cur_task_id = -1
self.backbone = backbone
self.device = device
self.embed_dim = kwargs["embd_dim"]
self.init_cls_num = kwargs['init_cls_num']
self.inc_cls_num = kwargs['inc_cls_num']
def update_fc(self):
self._cur_task_id += 1
if self._cur_task_id == 0:
classifier = nn.Linear(self.embed_dim, self.init_cls_num, bias=True)
nn.init.kaiming_uniform_(classifier.weight, nonlinearity='linear')
nn.init.constant_(classifier.bias, 0)
else:
classifier = nn.Linear(self.embed_dim, self.init_cls_num + self.inc_cls_num * (self._cur_task_id), bias=True)
nn.init.kaiming_uniform_(classifier.weight, nonlinearity='linear')
nn.init.constant_(classifier.bias, 0)
nb_output = self.classifier.out_features
classifier.weight.data[:nb_output] = copy.deepcopy(self.classifier.weight.data)
classifier.bias.data[:nb_output] = copy.deepcopy(self.classifier.bias.data)
del self.classifier
self.classifier = classifier
def forward(self, x, inference = False):
features = self.backbone(x)
logits = self.classifier(features)
return logits
class SD_LoRA(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.init_mag = kwargs['init_mag']
self.rank_reduction = kwargs['rank_reduction']
self.knowledge_dist = kwargs['knowledge_dist']
self._known_classes = 0
self._network = Model(backbone, device, **kwargs)
self.attention_modules = [module for module in self._network.modules() if isinstance(module, MultiHeadAttention_SDLoRA)]
def observe(self, data):
x, y = data['image'].to(self.device), data['label'].to(self.device)
logits = self._network(x)
# Masked previous classes
fake_y = y - self._known_classes
loss = F.cross_entropy(logits[:, self._known_classes:], fake_y)
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)
logits = self._network(x, inference = True)
preds = logits.max(1)[1]
correct_count = preds.eq(y).sum().item()
acc = correct_count / y.size(0)
return preds, acc
@torch.no_grad()
def before_task(self, task_idx, buffer, train_loader, test_loaders):
self._network.update_fc()
if self.rank_reduction[0]:
if task_idx == self.rank_reduction[1]:
for module in self.attention_modules:
module.lora_rank = self.rank_reduction[3]
elif task_idx == self.rank_reduction[2]:
for module in self.attention_modules:
module.lora_rank = self.rank_reduction[4]
# All blocks share same magnitude
mag = nn.ParameterList([nn.Parameter(torch.Tensor([self.init_mag])) for _ in range(task_idx + 1)])
for module in self.attention_modules:
module.mag_lora = mag
module.init_param()
self._network = self._network.to(self.device)
unfrezeed_params = []
for name, param in self._network.named_parameters():
param.requires_grad_(False)
if f'classifier' in name or \
f'lora' and f'list.{task_idx}' in name or \
('mag' in name and 'assimilated' not in name):
param.requires_grad_(True)
unfrezeed_params.append(name)
print(f"Current task : {task_idx}, Parameters to be updated: {len(unfrezeed_params)}")
@torch.no_grad()
def after_task(self, task_idx, buffer, train_loader, test_loaders):
self._known_classes += self.init_cls_num if task_idx == 0 else self.inc_cls_num
if self.knowledge_dist[0] and task_idx > 0:
for layer, module in enumerate(self.attention_modules):
dirs_q, dirs_v = [], []
for i in range(len(module.lora_A_q_list)):
norm_B = torch.norm(module.lora_B_q_list[i].weight)
norm_A = torch.norm(module.lora_A_q_list[i].weight)
if norm_A != 0 and norm_B != 0:
dirs_q.append(
(module.lora_B_q_list[i].weight @ module.lora_A_q_list[i].weight) / (norm_B * norm_A)
)
else: # zero-tensor, for consistency
dirs_q.append(
(module.lora_B_q_list[i].weight @ module.lora_A_q_list[i].weight)
)
norm_B = torch.norm(module.lora_B_v_list[i].weight)
norm_A = torch.norm(module.lora_A_v_list[i].weight)
if norm_A != 0 and norm_B != 0:
dirs_v.append(
(module.lora_B_v_list[i].weight @ module.lora_A_v_list[i].weight) / (norm_B * norm_A)
)
else: # zero-tensor, for consistency
dirs_v.append(
(module.lora_B_q_list[i].weight @ module.lora_A_q_list[i].weight)
)
flatten_dirs = [dir_q.flatten() for dir_q in dirs_q]
last_dir = flatten_dirs[-1].unsqueeze(1)
prev_dirs = torch.stack(flatten_dirs[:-1], dim=-1)
alphas = torch.linalg.lstsq(prev_dirs, last_dir)
if alphas.residuals < self.knowledge_dist[1]:
print(f'Layer {layer}: {alphas.residuals.item()} < {self.knowledge_dist[1]}, Q Merged with {alphas.solution}')
assert prev_dirs.shape[1] == len(module.assimilated_mag_lora_q) - 1
for ii in range(prev_dirs.shape[1]):
module.assimilated_mag_lora_q[ii] += alphas.solution[i]
nn.init.zeros_(module.lora_B_q_list[task_idx])
nn.init.zeros_(module.lora_A_q_list[task_idx])
flatten_dirs = [dir_v.flatten() for dir_v in dirs_v]
last_dir = flatten_dirs[-1].unsqueeze(1)
prev_dirs = torch.stack(flatten_dirs[:-1], dim=-1)
alphas = torch.linalg.lstsq(prev_dirs, last_dir)
if alphas.residuals < self.knowledge_dist[1]:
print(f'Layer {layer}: {alphas.residuals.item()} < {self.knowledge_dist[1]}, V Merged with {alphas.solution}')
assert prev_dirs.shape[1] == len(module.assimilated_mag_lora_v) - 1
for ii in range(prev_dirs.shape[1]):
module.assimilated_mag_lora_v[ii] += alphas.solution[i]
nn.init.zeros_(module.lora_B_v_list[task_idx])
nn.init.zeros_(module.lora_A_v_list[task_idx])
def get_parameters(self, config):
return self._network.parameters() |