import torch from new_impl.cv.elasticdnn.api.model import ElasticDNN_OfflineVQAFMModel, ElasticDNN_OfflineVQAMDModel from new_impl.cv.elasticdnn.api.algs.fm_lora import ElasticDNN_FMLoRAAlg from new_impl.cv.elasticdnn.model.base import ElasticDNNUtil from new_impl.cv.elasticdnn.pipeline.offline.fm_lora.base import FMLoRA_Util from blip import FMLoRA_blip_Util from new_impl.cv.elasticdnn.pipeline.offline.fm_to_md.base import FM_to_MD_Util from blip import FM_to_MD_blip_Util from utils.dl.common.model import LayerActivation, get_module, get_parameter, set_module from utils.common.exp import save_models_dict_for_init, get_res_save_dir from data import build_scenario from utils.common.log import logger import torch.nn.functional as F import sys from torch import nn from new_impl.cv.elasticdnn.api.algs.md_pretraining_wo_fbs import ElasticDNN_MDPretrainingWoFBSAlg class ElasticDNN_blip_OfflineVQAFMModel(ElasticDNN_OfflineVQAFMModel): def generate_md_by_reducing_width(self, reducing_width_ratio, samples: torch.Tensor): return FM_to_MD_blip_Util().init_md_from_fm_by_reducing_width_with_perf_test(self.models_dict['main'], reducing_width_ratio, samples) # raise NotImplementedError def get_feature_hook(self) -> LayerActivation: return LayerActivation(get_module(self.models_dict['main'], 'text_decoder.cls.predictions.decoder'), True, self.device) def get_elastic_dnn_util(self) -> ElasticDNNUtil: raise NotImplementedError def forward_to_get_task_loss(self, x, y, *args, **kwargs): self.to_train_mode() # print(x['input_ids'].size(), x['pixel_values'].size(), ) o = self.infer(x).logits # print(o.size(), y.size(), o, y) return F.binary_cross_entropy_with_logits(o, y) * y.shape[1] def get_lora_util(self) -> FMLoRA_Util: return FMLoRA_blip_Util() def get_task_head_params(self): head = get_module(self.models_dict['main'], 'text_decoder.cls.predictions.decoder') params_name = {k for k, v in head.named_parameters()} logger.info(f'task head params: {params_name}') return list(head.parameters()) class ElasticDNN_blip_OfflineVQAMDModel(ElasticDNN_OfflineVQAMDModel): def get_feature_hook(self) -> LayerActivation: return LayerActivation(get_module(self.models_dict['main'], 'text_decoder.cls.predictions.decoder'), True, self.device) # def forward_to_get_task_loss(self, x, y, *args, **kwargs): # self.to_train_mode() # # print(x['input_ids'].size(), x['pixel_values'].size(), ) # o = self.infer(x) # # print(o.size(), y.size(), o, y) # return F.binary_cross_entropy_with_logits(o, y) * y.shape[1] def forward_to_get_task_loss(self, x, y, *args, **kwargs): self.to_train_mode() o = self.models_dict['main'](**y) return o.loss def get_distill_loss(self, student_output, teacher_output): #print(student_output.shape, teacher_output.shape) return F.mse_loss(student_output, teacher_output.detach()) #return F.cross_entropy(student_output, teacher_output.detach()) def get_trained_params(self): return self.models_dict['main'].parameters() def get_matched_param_of_fm(self, self_param_name, fm: nn.Module): # TODO: if any([k in self_param_name for k in ['fbs', 'ab', 'embeddings']]): 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: # if self_param_name.startswith('norm'): # return None return get_parameter(fm, self_param_name) # 1. xx.query.weight -> xx.query.fc.weight and xx.query.ab.0/1 if ('query' in self_param_name or 'key' in self_param_name or \ 'value' in self_param_name) and ('weight' in self_param_name): ss = self_param_name.split('.') # raise NotImplementedError() # TODO: fm_qkv_name = '.'.join(ss[0: -1]) + '.fc' fm_qkv = get_module(fm, fm_qkv_name) fm_abs_name = '.'.join(ss[0: -1]) + '.ab' fm_abs = get_module(fm, fm_abs_name) return torch.cat([ fm_qkv.weight.data, # task-agnositc params fm_abs[1].weight @ fm_abs[0].weight ], dim=0) elif ('query' in self_param_name or 'key' in self_param_name or \ 'value' in self_param_name) and ('bias' in self_param_name): ss = self_param_name.split('.') fm_qkv_name = '.'.join(ss[0: -1]) + '.fc.bias' return get_parameter(fm, fm_qkv_name) elif 'intermediate.dense' 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 __name__ == '__main__': from utils.dl.common.env import set_random_seed set_random_seed(1) scenario = build_scenario( source_datasets_name=['VQA_split1'], target_datasets_order=['VQA_split1_c'] * 1, # TODO da_mode='close_set', data_dirs={ 'VQA_split1': '/data/zql/datasets/vqav2', 'VQA_split1_c': '/data/zql/datasets/vqav2' }, ) # 1. init model fm_models_dict_path = 'new_impl/mm/Vis_bert/QuestionAnswering/results/blip_lora.py/20231018/999999-095006-/data/zql/concept-drift-in-edge-projects/UniversalElasticNet/new_impl/mm/Vis_bert/QuestionAnswering/blip_lora.py/models/fm_best.pt' fm_models = torch.load(fm_models_dict_path) fm_models_dict_path = save_models_dict_for_init(fm_models, __file__, 'fm_blip_vqa_lora') md_models_dict_path = save_models_dict_for_init({ 'main': -1 }, __file__, 'md_blip_none') torch.cuda.set_device(1) device = 'cuda' fm_model = ElasticDNN_blip_OfflineVQAFMModel('fm', fm_models_dict_path, device) md_model = ElasticDNN_blip_OfflineVQAMDModel('md', md_models_dict_path, device) # 2. init alg models = { 'fm': fm_model, 'md': md_model } import sys # from experiments.elasticdnn.clip.offline.fm_to_md.cls.md_pretraining_wo_fbs import ElasticDNN_MDPretrainingWoFBSAlg # from methods.elasticdnn.api.algs.md_pretraining_wo_fbs_clip_debug import ElasticDNN_MDPretrainingWoFBSAlg fm_to_md_alg = ElasticDNN_MDPretrainingWoFBSAlg(models, get_res_save_dir(__file__, sys.argv[0])) # sample_dataset = list(scenario.get_offline_datasets().values())[0]['train'] sample_dataset = list(scenario.get_offline_datasets().values())[0]['train'] sample = sample_dataset[0][0] for k, v in sample.items(): sample[k] = v.unsqueeze(0) from utils.dl.common.lr_scheduler import get_linear_schedule_with_warmup fm_to_md_alg.run(scenario, hyps={ 'launch_tbboard': False, 'samples_size': sample, 'generate_md_width_ratio': 4, 'train_batch_size':32, 'val_batch_size': 512, 'num_workers': 16, 'optimizer': 'AdamW', 'optimizer_args': {'lr': 1e-5, 'betas': [0.9, 0.999], 'weight_decay': 0.01}, 'scheduler': 'LambdaLR', 'scheduler_args': {'lr_lambda': get_linear_schedule_with_warmup(10000, 70000)}, 'num_iters': 80000, 'val_freq': 1000, 'distill_loss_weight': 1. })