File size: 8,477 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 |
import sys
from utils.dl.common.env import set_random_seed
set_random_seed(1)
from typing import List
from data.dataloader import build_dataloader
from data import Scenario
from methods.elasticdnn.api.online_model_v2 import ElasticDNN_OnlineModel
import torch
import sys
from torch import nn
from methods.elasticdnn.api.model import ElasticDNN_OfflineSegFMModel, ElasticDNN_OfflineSegMDModel
from methods.elasticdnn.api.algs.md_pretraining_wo_fbs import ElasticDNN_MDPretrainingWoFBSAlg
from methods.elasticdnn.model.base import ElasticDNNUtil
from methods.elasticdnn.pipeline.offline.fm_to_md.base import FM_to_MD_Util
from methods.elasticdnn.pipeline.offline.fm_to_md.vit import FM_to_MD_ViT_Util
from methods.elasticdnn.pipeline.offline.fm_lora.base import FMLoRA_Util
from methods.elasticdnn.pipeline.offline.fm_lora.vit import FMLoRA_ViT_Util
from methods.elasticdnn.model.vit import ElasticViTUtil
from utils.common.file import ensure_dir
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
from utils.dl.common.env import create_tbwriter
import os
import shutil
from utils.common.log import logger
from utils.common.data_record import write_json
# from methods.shot.shot import OnlineShotModel
from methods.feat_align.main import OnlineFeatAlignModel, FeatAlignAlg
import tqdm
from methods.feat_align.mmd import mmd_rbf
from methods.base.alg import BaseAlg
from methods.base.model import BaseModel
from copy import deepcopy
import time
def baseline_da(app_name: str,
scenario: Scenario,
da_alg: BaseAlg,
da_alg_hyp: dict,
da_model: BaseModel,
device,
__entry_file__,
tag=None):
# involve_fm = settings['involve_fm']
task_name = app_name
# online_model = elasticfm_model
log_dir = get_res_save_dir(__entry_file__, tag=tag)
tb_writer = create_tbwriter(os.path.join(log_dir, 'tb_log'), False)
res = []
global_avg_after_acc = 0.
global_iter = 0
for domain_index, _ in enumerate(scenario.target_domains_order):
cur_target_domain_name = scenario.target_domains_order[scenario.cur_domain_index]
if cur_target_domain_name in da_alg_hyp:
da_alg_hyp = da_alg_hyp[cur_target_domain_name]
logger.info(f'use dataset-specific da_alg_hyp')
da_metrics, after_da_model = da_alg(
{'main': da_model},
os.path.join(log_dir, f'{task_name}/{domain_index}')
).run(scenario, da_alg_hyp)
# os.remove(tmp_sd_path)
# 前面在当前域上训练，在这里压缩调优？
# print(da_model.models_dict['main'])
# 进行压缩
reducing_width_ratio = 8
samples = torch.rand(1, 3, 224, 224).to(device)
trained_fm_model = deepcopy(da_model.models_dict['main'])
fm_da_model = deepcopy(da_model) # 保存大模型
lora_util = FMLoRA_ViT_Util()
lora_absorbed_fm_model = lora_util.absorb_lora_and_recover_net_structure(trained_fm_model, samples)
compressed_fm_model = FM_to_MD_ViT_Util().init_md_from_fm_by_reducing_width_with_perf_test(lora_absorbed_fm_model, reducing_width_ratio, samples)
da_model.models_dict['main'] = compressed_fm_model
# 进行调优？之前那个da_metrics是FM的结果吧，调优也能得到一个精度结果换成这个？
datasets_for_training = scenario.get_online_cur_domain_datasets_for_training()
train_dataset = datasets_for_training[cur_target_domain_name]['train']
val_dataset = datasets_for_training[cur_target_domain_name]['val']
datasets_for_inference = scenario.get_online_cur_domain_datasets_for_inference()
test_dataset = datasets_for_inference
train_loader = iter(build_dataloader(train_dataset, da_alg_hyp['train_batch_size'], da_alg_hyp['num_workers'], True, None))
test_loader = build_dataloader(test_dataset, da_alg_hyp['val_batch_size'], da_alg_hyp['num_workers'], False, False)
for p in compressed_fm_model.parameters():
p.requires_grad = True
da_model.to_train_mode()
# 'distill_optimizer': 'AdamW',
# 'distill_optimizer_args': {'lr': 1e-4, 'betas': [0.9, 0.999], 'weight_decay': 0.01},
optimizer = torch.optim.__dict__['AdamW']([
{'params': da_model.models_dict['main'].parameters(), **{'lr': 1e-4, 'betas': [0.9, 0.999], 'weight_decay': 0.01}}
])
if da_alg_hyp['scheduler'] != '':
scheduler = torch.optim.lr_scheduler.__dict__[da_alg_hyp['scheduler']](optimizer, **da_alg_hyp['scheduler_args'])
else:
scheduler = None
pbar = tqdm.tqdm(range(da_alg_hyp['num_iters']), dynamic_ncols=True)
accs = []
total_train_time = 0.
cur_acc = 0.
for iter_index in pbar:
cur_start_time = time.time()
da_model.to_train_mode()
fm_da_model.to_eval_mode()
x, y = next(train_loader)
if isinstance(x, dict):
for k, v in x.items():
if isinstance(v, torch.Tensor):
x[k] = v.to(device)
y = y.to(device)
else:
x, y = x.to(device), y.to(device)
with torch.no_grad():
fm_output = fm_da_model.infer(x)
md_output = da_model.infer(x)
distill_criterion = CrossEntropyLossSoft()
total_loss = distill_criterion(md_output, fm_output)
optimizer.zero_grad()
total_loss.backward()
optimizer.step()
if scheduler is not None:
scheduler.step()
total_train_time += time.time() - cur_start_time
if (iter_index + 1) % da_alg_hyp['val_freq'] == 0:
from data import split_dataset
cur_md = da_model.models_dict['main']
md_for_test = deepcopy(da_model.models_dict['main'])
da_model.models_dict['main'] = md_for_test
cur_test_batch_dataset = split_dataset(test_dataset, da_alg_hyp['val_batch_size'], iter_index + 1)[0]
cur_test_batch_dataloader = build_dataloader(cur_test_batch_dataset, da_alg_hyp['train_batch_size'], da_alg_hyp['num_workers'], False, False)
da_model.to_eval_mode()
cur_acc = da_model.get_accuracy(cur_test_batch_dataloader)
accs += [{
'iter': iter_index + 1,
'acc': cur_acc
}]
pbar.set_description(f'loss: {total_loss:.6f}, cur_acc: {cur_acc:.4f}')
time_usage = total_train_time
da_metrics = {
'accs': accs,
'time': time_usage
}
da_model = fm_da_model # 恢复大模型
# 蒸馏结束
if domain_index > 0:
shutil.rmtree(os.path.join(log_dir, f'{task_name}/{domain_index}/backup_codes'))
accs = da_metrics['accs']
before_acc = accs[0]['acc']
after_acc = accs[-1]['acc']
tb_writer.add_scalars(f'accs/{task_name}', dict(before=before_acc, after=after_acc), domain_index)
tb_writer.add_scalar(f'times/{task_name}', da_metrics['time'], domain_index)
for _acc in accs:
tb_writer.add_scalar('total_acc', _acc['acc'], _acc['iter'] + global_iter)
global_iter += _acc['iter'] + 1
scenario.next_domain()
logger.info(f"task: {task_name}, domain {domain_index}, acc: {before_acc:.4f} -> "
f"{after_acc:.4f} ({da_metrics['time']:.2f}s)")
global_avg_after_acc += after_acc
cur_res = da_metrics
res += [cur_res]
write_json(os.path.join(log_dir, 'res.json'), res, backup=False)
global_avg_after_acc /= (domain_index + 1)
logger.info(f'-----> final metric: {global_avg_after_acc:.4f}')
write_json(os.path.join(log_dir, f'res_{global_avg_after_acc:.4f}.json'), res, backup=False)
|