File size: 17,722 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 |
# from typing import Any, Dict
# from schema import Schema, Or
# import schema
# from data import Scenario, MergedDataset
# from methods.base.alg import BaseAlg
# from data import build_dataloader
# from ..model import ElasticDNN_OfflineFMModel, ElasticDNN_OfflineMDModel
# from ...model.base import ElasticDNNUtil
# import torch.optim
# import tqdm
# import torch.nn.functional as F
# from torch import nn
# from utils.dl.common.env import create_tbwriter
# import os
# import random
# import numpy as np
# from copy import deepcopy
# from utils.dl.common.model import LayerActivation2, get_module
# from utils.common.log import logger
# class ElasticDNN_MDPretrainingWoFBSAlg(BaseAlg):
# """
# TODO: fine-tuned FM -> init MD -> trained MD -> construct indexes (only between similar weights) and fine-tune
# """
# def get_required_models_schema(self) -> Schema:
# return Schema({
# 'fm': ElasticDNN_OfflineFMModel,
# 'md': ElasticDNN_OfflineMDModel
# })
# def get_required_hyp_schema(self) -> Schema:
# return Schema({
# 'launch_tbboard': bool,
# 'samples_size': object,
# 'generate_md_width_ratio': int,
# 'train_batch_size': int,
# 'val_batch_size': int,
# 'num_workers': int,
# 'optimizer': str,
# 'optimizer_args': dict,
# 'scheduler': str,
# 'scheduler_args': dict,
# 'num_iters': int,
# 'val_freq': int,
# 'distill_loss_weight': float
# })
# def run(self, scenario: Scenario, hyps: Dict) -> Dict[str, Any]:
# super().run(scenario, hyps)
# assert isinstance(self.models['md'], ElasticDNN_OfflineMDModel) # for auto completion
# assert isinstance(self.models['fm'], ElasticDNN_OfflineFMModel) # for auto completion
# # 1. add FBS
# device = self.models['md'].device
# if self.models['md'].models_dict['main'] == -1:
# logger.info(f'init master DNN by reducing width of an adapted foundation model (already tuned by LoRA)...')
# before_fm_model = deepcopy(self.models['fm'].models_dict['main'])
# lora_util = self.models['fm'].get_lora_util()
# sample = hyps['samples_size']
# if isinstance(sample, (tuple, list)) and isinstance(sample[0], int):
# sample = torch.rand(hyps['samples_size']).to(device)
# lora_absorbed_fm_model = lora_util.absorb_lora_and_recover_net_structure(self.models['fm'].models_dict['main'],
# sample)
# self.models['fm'].models_dict['main'] = lora_absorbed_fm_model
# master_dnn = self.models['fm'].generate_md_by_reducing_width(hyps['generate_md_width_ratio'],
# sample)
# self.models['fm'].models_dict['main'] = before_fm_model
# self.models['md'].models_dict['main'] = master_dnn
# self.models['md'].to(device)
# # 2. train (knowledge distillation, index relationship)
# offline_datasets = scenario.get_offline_datasets()
# train_dataset = MergedDataset([d['train'] for d in offline_datasets.values()])
# val_dataset = MergedDataset([d['val'] for d in offline_datasets.values()])
# train_loader = iter(build_dataloader(train_dataset, hyps['train_batch_size'], hyps['num_workers'],
# True, None))
# val_loader = build_dataloader(val_dataset, hyps['val_batch_size'], hyps['num_workers'],
# False, False)
# # val_acc = self.models['md'].get_accuracy(val_loader)
# # print(val_acc)
# # exit()
# # 2.1 train whole master DNN (knowledge distillation)
# self.models['md'].to_train_mode()
# for p in master_dnn.parameters():
# p.requires_grad = True
# if hasattr(self.models['md'], 'get_trained_params'):
# trained_p = self.models['md'].get_trained_params()
# logger.info(f'use custom trained parameters!!')
# else:
# trained_p = self.models['md'].models_dict['main'].parameters()
# for p in trained_p:
# p.requires_grad = True
# optimizer = torch.optim.__dict__[hyps['optimizer']]([
# {'params': trained_p, **hyps['optimizer_args']}
# ])
# scheduler = torch.optim.lr_scheduler.__dict__[hyps['scheduler']](optimizer, **hyps['scheduler_args'])
# tb_writer = create_tbwriter(os.path.join(self.res_save_dir, 'tb_log'), launch_tbboard=hyps['launch_tbboard'])
# pbar = tqdm.tqdm(range(hyps['num_iters']), dynamic_ncols=True)
# best_avg_val_acc = 0.
# md_output_hook = None
# for iter_index in pbar:
# self.models['md'].to_train_mode()
# self.models['fm'].to_eval_mode()
# # rand_sparsity = random.random() * (hyps['max_sparsity'] - hyps['min_sparsity']) + hyps['min_sparsity']
# # elastic_dnn_util.set_master_dnn_sparsity(self.models['md'].models_dict['main'], rand_sparsity)
# if md_output_hook is None:
# md_output_hook = self.models['md'].get_feature_hook()
# fm_output_hook = self.models['fm'].get_feature_hook()
# 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 = self.models['fm'].infer(x)
# task_loss = self.models['md'].forward_to_get_task_loss(x, y)
# if isinstance(md_output_hook, (tuple, list)):
# distill_loss = 0.
# for h1, h2 in zip(md_output_hook, fm_output_hook):
# md_output = h1.output
# fm_output = h2.output
# distill_loss += hyps['distill_loss_weight'] * self.models['md'].get_distill_loss(md_output, fm_output)
# else:
# md_output = md_output_hook.output
# fm_output = fm_output_hook.output
# distill_loss = hyps['distill_loss_weight'] * self.models['md'].get_distill_loss(md_output, fm_output)
# total_loss = task_loss + distill_loss
# optimizer.zero_grad()
# total_loss.backward()
# # for n, p in self.models['md'].models_dict['main'].named_parameters():
# # if p.grad is not None:
# # print(n)
# # exit()
# optimizer.step()
# scheduler.step()
# if (iter_index + 1) % hyps['val_freq'] == 0:
# # elastic_dnn_util.clear_cached_channel_attention_in_master_dnn(self.models['md'].models_dict['main'])
# if isinstance(md_output_hook, (tuple, list)):
# [h.remove() for h in md_output_hook]
# [h.remove() for h in fm_output_hook]
# else:
# md_output_hook.remove()
# fm_output_hook.remove()
# md_output_hook = None
# fm_output_hook = None
# cur_md = self.models['md'].models_dict['main']
# md_for_test = deepcopy(self.models['md'].models_dict['main'])
# val_acc = 0.
# self.models['md'].models_dict['main'] = md_for_test
# self.models['md'].to_eval_mode()
# val_acc = self.models['md'].get_accuracy(val_loader)
# self.models['md'].models_dict['main'] = cur_md
# self.models['md'].save_model(os.path.join(self.res_save_dir, 'models/md_last.pt'))
# self.models['fm'].save_model(os.path.join(self.res_save_dir, 'models/fm_last.pt'))
# if val_acc > best_avg_val_acc:
# best_avg_val_acc = val_acc
# self.models['md'].save_model(os.path.join(self.res_save_dir, 'models/md_best.pt'))
# self.models['fm'].save_model(os.path.join(self.res_save_dir, 'models/fm_best.pt'))
# tb_writer.add_scalars(f'losses', dict(task=task_loss, distill=distill_loss, total=total_loss), iter_index)
# pbar.set_description(f'loss: {total_loss:.6f}')
# if (iter_index + 1) >= hyps['val_freq']:
# tb_writer.add_scalar(f'accs/val_acc', val_acc, iter_index)
# pbar.set_description(f'loss: {total_loss:.6f}, val_acc: {val_acc:.4f}')
# code below is commented on 0716 17:49, because of a bug that the loss cannot be gradient decented
# (bug confirmed, why? I dont know :)
from typing import Any, Dict
from schema import Schema, Or
import schema
from data import Scenario, MergedDataset
from methods.base.alg import BaseAlg
from data import build_dataloader
from ..model import ElasticDNN_OfflineFMModel, ElasticDNN_OfflineMDModel
from ...model.base import ElasticDNNUtil
import torch.optim
import tqdm
import torch.nn.functional as F
from torch import nn
from utils.dl.common.env import create_tbwriter
import os
import random
import numpy as np
from copy import deepcopy
from utils.dl.common.model import LayerActivation2, get_module
from utils.common.log import logger
from torchvision.transforms import Compose
class ElasticDNN_MDPretrainingWoFBSAlg(BaseAlg):
"""
TODO: fine-tuned FM -> init MD -> trained MD -> construct indexes (only between similar weights) and fine-tune
"""
def get_required_models_schema(self) -> Schema:
return Schema({
'fm': ElasticDNN_OfflineFMModel,
'md': ElasticDNN_OfflineMDModel
})
def get_required_hyp_schema(self) -> Schema:
from schema import Optional
return Schema({
'launch_tbboard': bool,
'samples_size': any,
'generate_md_width_ratio': int,
'train_batch_size': int,
'val_batch_size': int,
'num_workers': int,
'optimizer': str,
'optimizer_args': dict,
'scheduler': str,
'scheduler_args': dict,
'num_iters': int,
'val_freq': int,
'distill_loss_weight': float,
Optional('transform'): Compose,
})
def run(self, scenario: Scenario, hyps: Dict, collate_fn=None) -> Dict[str, Any]:
super().run(scenario, hyps)
assert isinstance(self.models['md'], ElasticDNN_OfflineMDModel) # for auto completion
assert isinstance(self.models['fm'], ElasticDNN_OfflineFMModel) # for auto completion
# 1. add FBS
device = self.models['md'].device
if self.models['md'].models_dict['main'] == -1:
logger.info(f'init master DNN by reducing width of an adapted foundation model (already tuned by LoRA)...')
before_fm_model = deepcopy(self.models['fm'].models_dict['main'])
lora_util = self.models['fm'].get_lora_util()
sample = hyps['samples_size']
if isinstance(sample, (tuple, list)) and isinstance(sample[0], int):
sample = torch.rand(hyps['samples_size']).to(device)
lora_absorbed_fm_model = lora_util.absorb_lora_and_recover_net_structure(self.models['fm'].models_dict['main'],
sample)
self.models['fm'].models_dict['main'] = lora_absorbed_fm_model
master_dnn = self.models['fm'].generate_md_by_reducing_width(hyps['generate_md_width_ratio'],
sample)
self.models['fm'].models_dict['main'] = before_fm_model
self.models['md'].models_dict['main'] = master_dnn
self.models['md'].to(device)
# 2. train (knowledge distillation, index relationship)
if 'transform' in hyps.keys():
offline_datasets = scenario.get_offline_datasets(transform=hyps['transform'])
else:
offline_datasets = scenario.get_offline_datasets()
train_dataset = MergedDataset([d['train'] for d in offline_datasets.values()])
val_dataset = MergedDataset([d['val'] for d in offline_datasets.values()])
train_loader = iter(build_dataloader(train_dataset, hyps['train_batch_size'], hyps['num_workers'],
True, None, collate_fn=collate_fn))
val_loader = build_dataloader(val_dataset, hyps['val_batch_size'], hyps['num_workers'],
False, False, collate_fn=collate_fn)
# logger.info(f'FM acc: {self.models["fm"].get_accuracy(val_loader):.4f}')
# 2.1 train whole master DNN (knowledge distillation)
for p in master_dnn.parameters():
p.requires_grad = True
self.models['md'].to_train_mode()
optimizer = torch.optim.__dict__[hyps['optimizer']]([
{'params': self.models['md'].models_dict['main'].parameters(), **hyps['optimizer_args']}
])
scheduler = torch.optim.lr_scheduler.__dict__[hyps['scheduler']](optimizer, **hyps['scheduler_args'])
tb_writer = create_tbwriter(os.path.join(self.res_save_dir, 'tb_log'), launch_tbboard=hyps['launch_tbboard'])
pbar = tqdm.tqdm(range(hyps['num_iters']), dynamic_ncols=True)
best_avg_val_acc = 0.
md_output_hook = None
for iter_index in pbar:
self.models['md'].to_train_mode()
self.models['fm'].to_eval_mode()
# rand_sparsity = random.random() * (hyps['max_sparsity'] - hyps['min_sparsity']) + hyps['min_sparsity']
# elastic_dnn_util.set_master_dnn_sparsity(self.models['md'].models_dict['main'], rand_sparsity)
if md_output_hook is None:
md_output_hook = self.models['md'].get_feature_hook()
fm_output_hook = self.models['fm'].get_feature_hook()
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 = self.models['fm'].infer(x)
task_loss = self.models['md'].forward_to_get_task_loss(x, y)
md_output = md_output_hook.output
fm_output = fm_output_hook.output
distill_loss = hyps['distill_loss_weight'] * self.models['md'].get_distill_loss(md_output, fm_output)
total_loss = task_loss + distill_loss
optimizer.zero_grad()
total_loss.backward()
optimizer.step()
scheduler.step()
if (iter_index + 1) % hyps['val_freq'] == 0:
# elastic_dnn_util.clear_cached_channel_attention_in_master_dnn(self.models['md'].models_dict['main'])
md_output_hook.remove()
md_output_hook = None
fm_output_hook.remove()
fm_output_hook = None
cur_md = self.models['md'].models_dict['main']
md_for_test = deepcopy(self.models['md'].models_dict['main'])
val_acc = 0.
self.models['md'].models_dict['main'] = md_for_test
self.models['md'].to_eval_mode()
val_acc = self.models['md'].get_accuracy(val_loader)
self.models['md'].models_dict['main'] = cur_md
self.models['md'].save_model(os.path.join(self.res_save_dir, 'models/md_last.pt'))
self.models['fm'].save_model(os.path.join(self.res_save_dir, 'models/fm_last.pt'))
if val_acc > best_avg_val_acc:
best_avg_val_acc = val_acc
self.models['md'].save_model(os.path.join(self.res_save_dir, 'models/md_best.pt'))
self.models['fm'].save_model(os.path.join(self.res_save_dir, 'models/fm_best.pt'))
tb_writer.add_scalars(f'losses', dict(task=task_loss, distill=distill_loss, total=total_loss), iter_index)
pbar.set_description(f'loss: {total_loss:.6f}')
if (iter_index + 1) >= hyps['val_freq']:
tb_writer.add_scalar(f'accs/val_acc', val_acc, iter_index)
pbar.set_description(f'loss: {total_loss:.6f}, val_acc: {val_acc:.4f}')
|