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 })