File size: 12,546 Bytes
b84549f |
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 |
import torch
import sys
from torch import nn
from dnns.vit import make_softmax_prunable
from new_impl.cv.elasticdnn.api.model import ElasticDNN_OfflineClsFMModel, ElasticDNN_OfflineClsMDModel
# from methods.elasticdnn.api.algs.md_pretraining_w_fbs import ElasticDNN_MDPretrainingWFBSAlg
from new_impl.cv.elasticdnn.model.base import ElasticDNNUtil
from new_impl.cv.elasticdnn.pipeline.offline.fm_to_md.base import FM_to_MD_Util
from new_impl.cv.elasticdnn.pipeline.offline.fm_to_md.vit import FM_to_MD_ViT_Util
from new_impl.cv.elasticdnn.pipeline.offline.fm_lora.base import FMLoRA_Util
from new_impl.cv.elasticdnn.pipeline.offline.fm_lora.vit import FMLoRA_ViT_Util
from new_impl.cv.elasticdnn.model.vit import ElasticViTUtil
from utils.dl.common.model import LayerActivation, get_module, get_parameter
from utils.common.exp import save_models_dict_for_init, get_res_save_dir
from data import build_scenario
from utils.dl.common.loss import CrossEntropyLossSoft
import torch.nn.functional as F
class ElasticDNN_ViT_OfflineClsFMModel(ElasticDNN_OfflineClsFMModel):
def generate_md_by_reducing_width(self, reducing_width_ratio, samples: torch.Tensor):
return FM_to_MD_ViT_Util().init_md_from_fm_by_reducing_width_with_perf_test(self.models_dict['main'],
reducing_width_ratio, samples).to(self.device)
def get_feature_hook(self) -> LayerActivation:
return LayerActivation(get_module(self.models_dict['main'], 'classifier'), True, self.device)
def get_elastic_dnn_util(self) -> ElasticDNNUtil:
return ElasticViTUtil()
def forward_to_get_task_loss(self, x, y, *args, **kwargs):
return F.cross_entropy(self.infer(x).logits, y)
def get_lora_util(self) -> FMLoRA_Util:
return FMLoRA_ViT_Util()
def get_task_head_params(self):
head = get_module(self.models_dict['main'], 'classifier')
return list(head.parameters())
class ElasticDNN_ViT_OfflineClsMDModel(ElasticDNN_OfflineClsMDModel):
def __init__(self, name: str, models_dict_path: str, device: str):
super().__init__(name, models_dict_path, device)
self.distill_criterion = CrossEntropyLossSoft()
def get_feature_hook(self) -> LayerActivation:
return LayerActivation(get_module(self.models_dict['main'], 'classifier'), True, self.device)
def forward_to_get_task_loss(self, x, y, *args, **kwargs):
return F.cross_entropy(self.infer(x).logits, y)
def get_distill_loss(self, student_output, teacher_output):
return self.distill_criterion(student_output, teacher_output)
def get_matched_param_of_fm(self, self_param_name, fm: nn.Module):
# only between qkv.weight, norm.weight/bias
if any([k in self_param_name for k in ['fbs', 'cls_token', 'pos_embed']]):
return None
p = get_parameter(self.models_dict['main'], self_param_name)
if p.dim() == 0:
return None
elif p.dim() == 1 and 'layernorm' in self_param_name and 'weight' in self_param_name:
return get_parameter(fm, self_param_name)
# 1. xx.qkv.to_qkv.yy to xx.qkv.qkv.aa and xx.qkv.abs.zz
# if 'qkv.weight' in self_param_name:
# ss = self_param_name.split('.')
# fm_qkv_name = '.'.join(ss[0: -2]) + '.qkv'
# fm_qkv = get_module(fm, fm_qkv_name)
# fm_abs_name = '.'.join(ss[0: -2]) + '.abs'
# fm_abs = get_module(fm, fm_abs_name)
# return torch.cat([
# fm_qkv.weight.data, # task-agnositc params
# torch.cat([(_abs[0].weight.T @ _abs[1].weight.T).T for _abs in fm_abs], dim=0) # task-specific params (LoRA)
# ], dim=0)
# # elif 'to_qkv.bias' in self_param_name:
# # ss = self_param_name.split('.')
# # fm_qkv_name = '.'.join(ss[0: -2]) + '.qkv.bias'
# # return get_parameter(fm, fm_qkv_name)
# elif 'mlp.fc1' in self_param_name and 'weight' in self_param_name:
# fm_param_name = self_param_name.replace('.linear', '')
# return get_parameter(fm, fm_param_name)
# elif 'mlp.fc2' in self_param_name and 'weight' in self_param_name:
# fm_param_name = self_param_name
# return get_parameter(fm, fm_param_name)
# else:
# # return get_parameter(fm, self_param_name)
# return None
# if 'qkv.weight' in self_param_name:
# ss = self_param_name.split('.')
# fm_qkv_name = '.'.join(ss[0: -1]) + '.qkv'
# fm_qkv = get_module(fm, fm_qkv_name)
# fm_abs_name = '.'.join(ss[0: -1]) + '.abs'
# fm_abs = get_module(fm, fm_abs_name)
# return torch.cat([
# fm_qkv.weight.data, # task-agnositc params
# torch.cat([(_abs[0].weight.T @ _abs[1].weight.T).T for _abs in fm_abs], dim=0) # task-specific params (LoRA)
# ], dim=0)
# # elif 'to_qkv.bias' in self_param_name:
# # ss = self_param_name.split('.')
# # fm_qkv_name = '.'.join(ss[0: -2]) + '.qkv.bias'
# # return get_parameter(fm, fm_qkv_name)
# elif 'mlp.fc1' in self_param_name and 'weight' in self_param_name:
# fm_param_name = self_param_name.replace('.linear', '')
# return get_parameter(fm, fm_param_name)
# elif 'mlp.fc2' in self_param_name and 'weight' in self_param_name:
# fm_param_name = self_param_name
# res = get_parameter(fm, fm_param_name)
# # print('mlp fc2 debug', fm_param_name, res is None)
# return res
# else:
# # return get_parameter(fm, self_param_name)
# return None
# def get_matched_param_of_fm(self, self_param_name, fm: nn.Module):
# if any([k in self_param_name for k in ['fbs', 'cls_token', 'pos_embed']]):
# return None
# # 1. xx.qkv.to_qkv.yy to xx.qkv.qkv.aa and xx.qkv.abs.zz
# if 'to_qkv.weight' in self_param_name:
# ss = self_param_name.split('.')
# fm_qkv_name = '.'.join(ss[0: -2]) + '.qkv'
# fm_qkv = get_module(fm, fm_qkv_name)
# fm_abs_name = '.'.join(ss[0: -2]) + '.abs'
# fm_abs = get_module(fm, fm_abs_name)
# return torch.cat([
# fm_qkv.weight.data, # task-agnositc params
# torch.cat([(_abs[0].weight.T @ _abs[1].weight.T).T for _abs in fm_abs], dim=0) # task-specific params (LoRA)
# ], dim=0)
# elif 'to_qkv.bias' in self_param_name:
# ss = self_param_name.split('.')
# fm_qkv_name = '.'.join(ss[0: -2]) + '.qkv.bias'
# return get_parameter(fm, fm_qkv_name)
# elif 'mlp.fc1' in self_param_name:
# fm_param_name = self_param_name.replace('.linear', '')
# return get_parameter(fm, fm_param_name)
# else:
# return get_parameter(fm, self_param_name)
if ('attention.attention.projection_query' in self_param_name or 'attention.attention.projection_key' in self_param_name or \
'attention.attention.projection_value' in self_param_name) and ('weight' in self_param_name):
ss = self_param_name.split('.')
fm_qkv_name = '.'.join(ss[0: -1]) + '.qkv'
fm_qkv = get_module(fm, fm_qkv_name)
fm_abs_name = '.'.join(ss[0: -1]) + '.abs'
fm_abs = get_module(fm, fm_abs_name)
return torch.cat([
fm_qkv.weight.data, # task-agnositc params
torch.cat([(_abs[0].weight.T @ _abs[1].weight.T).T for _abs in fm_abs], dim=0) # task-specific params (LoRA)
], dim=0)
elif ('attention.attention.projection_query' in self_param_name or 'attention.attention.projection_key' in self_param_name or \
'attention.attention.projection_value' in self_param_name) and ('bias' in self_param_name):
ss = self_param_name.split('.')
fm_qkv_name = '.'.join(ss[0: -1]) + '.qkv.bias'
return get_parameter(fm, fm_qkv_name)
elif 'intermediate.dense' in self_param_name and 'weight' in self_param_name:
fm_param_name = self_param_name.replace('.linear', '')
return get_parameter(fm, fm_param_name)
elif 'output.dense' in self_param_name and 'weight' in self_param_name:
fm_param_name = self_param_name.replace('.linear', '')
return get_parameter(fm, fm_param_name)
else:
#return get_parameter(fm, self_param_name)
return None
if __name__ == '__main__':
from utils.dl.common.env import set_random_seed
set_random_seed(1)
# 1. init model
from dnns.vit import vit_b_16
fm_models_dict_path = 'new_impl/cv/results/cvt_md_wo_fbs.py/20231019/999998-135139-/data/zql/concept-drift-in-edge-projects/UniversalElasticNet/new_impl/cv/cvt_md_wo_fbs.py/models/fm_best.pt'
fm_models_dict_path = save_models_dict_for_init(torch.load(fm_models_dict_path), __file__, 'fm_cvt_cls_lora')
pretrained_md_models_dict_path = 'new_impl/cv/results/cvt_md_wo_fbs.py/20231019/999998-135139-/data/zql/concept-drift-in-edge-projects/UniversalElasticNet/new_impl/cv/cvt_md_wo_fbs.py/models/md_best.pt'
md_models_dict = torch.load(pretrained_md_models_dict_path)
md_models_dict_path = save_models_dict_for_init(md_models_dict, __file__, 'md_cvt_cls_pretrained_wo_fbs')
torch.cuda.set_device(1)
device = 'cuda'
fm_model = ElasticDNN_ViT_OfflineClsFMModel('fm', fm_models_dict_path, device)
md_model = ElasticDNN_ViT_OfflineClsMDModel('md', md_models_dict_path, device)
# 2. init alg
models = {
'fm': fm_model,
'md': md_model
}
from new_impl.cv.elasticdnn.api.algs.md_pretraining_index_v2_train_index_and_md import ElasticDNN_MDPretrainingIndexAlg
fm_to_md_alg = ElasticDNN_MDPretrainingIndexAlg(models, get_res_save_dir(__file__, sys.argv[0]))
# 3. init scenario
scenario = build_scenario(
source_datasets_name=['GTA5Cls', 'SuperviselyPersonCls'],
target_datasets_order=['CityscapesCls', 'BaiduPersonCls'] * 15,
da_mode='close_set',
data_dirs={
'GTA5Cls': '/data/zql/datasets/gta5_for_cls_task',
'SuperviselyPersonCls': '/data/zql/datasets/supervisely_person_for_cls_task',
'CityscapesCls': '/data/zql/datasets/cityscapes_for_cls_task',
'BaiduPersonCls': '/data/zql/datasets/baidu_person_for_cls_task'
},
)
from utils.dl.common.lr_scheduler import get_linear_schedule_with_warmup
fm_to_md_alg.run(scenario, hyps={
'launch_tbboard': False,
'samples_size': (1, 3, 224, 224),
'FBS_r': 16,
'FBS_ignore_layers': [],
'train_batch_size': 128,
'val_batch_size': 512,
'num_workers': 16,
'optimizer': 'AdamW',
# 'optimizer_args': {'lr': 1e-4, 'betas': [0.9, 0.999], 'weight_decay': 0.01},
'indexes_optimizer_args': {'lr': 3e-3, 'betas': [0.9, 0.999], 'weight_decay': 0.1},
# 'scheduler': 'StepLR',
# 'scheduler_args': {'step_size': 20000, 'gamma': 0.1},
# 'optimizer': 'AdamW',
# 'optimizer_args': {'lr': 1e-4, 'betas': [0.9, 0.999], 'weight_decay': 0.01},
'optimizer_args': {'lr': 1e-3, 'betas': [0.9, 0.999], 'weight_decay': 0.01},#注意学习率的调整，不同的模型不一样。
'scheduler': 'LambdaLR',
'scheduler_args': {'lr_lambda': get_linear_schedule_with_warmup(10000, 70000)},
'max_sparsity': 0.9,
'min_sparsity': 0.0,
'num_iters': 60000,
'val_freq': 1000,
'index_loss_weight': 1e-4,
'l1_reg_loss_weight': 1e-9,
'val_num_sparsities': 4,
'bn_cal_num_iters': 800,#有bn层注意需要加上这个
'index_init': 'zero',
'index_guided_linear_comb_split_size': 512
})
|