File size: 15,329 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 |
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 LayerActivation, get_module
from utils.common.log import logger
import matplotlib.pyplot as plt
class ElasticDNN_MDPretrainingIndexAlg(BaseAlg):
"""
construct indexes between a filter/row of MD and all filters/rows of FM in the same layer
too huge indexes (~1GB), train so slow, hard to optimize
"""
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': (int, int, int, int),
'train_batch_size': int,
'val_batch_size': int,
'num_workers': int,
'optimizer': str,
'indexes_optimizer_args': dict,
'scheduler': str,
'scheduler_args': dict,
'num_iters': int,
'val_freq': int,
'index_loss_l1_weight': float,
'val_num_sparsities': int,
'bn_cal_num_iters': int,
'index_guided_linear_comb_split_size': Or(int, None)
})
def upsample_2d_tensor(self, p: torch.Tensor, target_len: int):
assert p.dim() == 2 # regard 2d weight as (batch_size, 1d_vector_dim)
return F.upsample(p.unsqueeze(1).unsqueeze(3),
size=(target_len, 1),
mode='bilinear').squeeze(3).squeeze(1)
def two_params_diff_fast(self, trained_p: torch.Tensor, ref_p: torch.Tensor,
index: torch.Tensor,
split_size: int):
assert trained_p.dim() == ref_p.dim()
assert index.size(0) == trained_p.size(0) and index.size(1) == ref_p.size(0)
# print(trained_p.size(), ref_p.size(), index.size())
ref_p = ref_p.detach()
if trained_p.dim() > 1:
trained_p = trained_p.flatten(1)
ref_p = ref_p.flatten(1)
# the weight size of master DNN and foundation model may be totally different
# MD -> FM: upsample first
# FM -> MD: downsample first
if trained_p.size(1) < ref_p.size(1):
trained_p = self.upsample_2d_tensor(trained_p, ref_p.size(1))
index = index.unsqueeze(-1)
# linear_combed_ref_p = (ref_p.unsqueeze(0) * index).sum(1)
# else:
# print(trained_p.size(), ref_p.size(), index.size())
if split_size is None:
# old version: huge memory consumption, not recommended (although this is fastest)
# print('old version')
linear_combed_ref_p = (ref_p.unsqueeze(0) * index).sum(1)
else:
# new version
linear_combed_ref_p = 0
cur_split_size = split_size
while index.size(1) % cur_split_size != 0:
cur_split_size -= 1
# print(cur_split_size)
for i in range(0, index.size(1), cur_split_size):
# if not isinstance(linear_combed_ref_p, int):
# print(linear_combed_ref_p.size(), ref_p.unsqueeze(0)[:, i: i + cur_split_size].size(), index[:, i: i + cur_split_size].size())
linear_combed_ref_p += ref_p.unsqueeze(0)[:, i: i + cur_split_size] * index[:, i: i + cur_split_size]
linear_combed_ref_p = linear_combed_ref_p.sum(1)
diff = (linear_combed_ref_p - trained_p).norm(2) ** 2
return diff
def get_index_loss(self, fm, md, indexes, match_fn, split_size):
res = 0.
for name, p in md.named_parameters():
if name not in indexes.keys():
continue
# if p.dim() == 0:
# continue
raw_p = match_fn(name, fm)
# if raw_p is None:
# continue
index = indexes[name]
# print(name)
res += self.two_params_diff_fast(p, raw_p, index, split_size)
return res
def bn_cal(self, model: nn.Module, train_loader, num_iters, device):
has_bn = False
for n, m in model.named_modules():
if isinstance(m, nn.BatchNorm2d):
has_bn = True
break
if not has_bn:
return {}
def bn_calibration_init(m):
""" calculating post-statistics of batch normalization """
if getattr(m, 'track_running_stats', False):
# reset all values for post-statistics
m.reset_running_stats()
# set bn in training mode to update post-statistics
m.training = True
with torch.no_grad():
model.eval()
model.apply(bn_calibration_init)
for _ in range(num_iters):
x, _ = next(train_loader)
model(x.to(device))
model.eval()
bn_stats = {}
for n, m in model.named_modules():
if isinstance(m, nn.BatchNorm2d):
bn_stats[n] = m
return bn_stats
def run(self, scenario: Scenario, hyps: Dict) -> Dict[str, Any]:
super().run(scenario, hyps)
# sanity check
# a= torch.tensor([[1, 2, 3], [1, 2, 4]])
# index = torch.tensor([[1, 2, 3],
# [1, 2, 4]])
# b = torch.tensor([[1, 2, 3], [1, 2, 4], [2, 3, 4]])
# print(self.two_params_diff_fast(a, b, index, hyps['index_guided_linear_comb_split_size']))
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
# 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()
# lora_absorbed_fm_model = lora_util.absorb_lora_and_recover_net_structure(self.models['fm'].models_dict['main'],
# torch.rand(hyps['samples_size']).to(device))
# 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'],
# torch.rand(hyps['samples_size']).to(device))
# self.models['fm'].models_dict['main'] = before_fm_model
elastic_dnn_util = self.models['fm'].get_elastic_dnn_util()
# master_dnn = elastic_dnn_util.convert_raw_dnn_to_master_dnn_with_perf_test(master_dnn,
# hyps['FBS_r'], hyps['FBS_ignore_layers'])
# self.models['md'].models_dict['main'] = master_dnn
# self.models['md'].to(device)
# master_dnn = self.models['md'].models_dict['main']
# 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)
# 2.1 train only FBS (skipped because current md cannot do proper inference)
# 2.2 train whole master DNN (knowledge distillation, index relationship)
# for p in master_dnn.parameters():
# p.requires_grad = True
# self.models['md'].to_train_mode()
indexes = {}
for name, p in self.models['md'].models_dict['main'].named_parameters():
if p.dim() > 1:
matched_p_in_fm = self.models['md'].get_matched_param_of_fm(name, self.models['fm'].models_dict['main'])
if matched_p_in_fm is None:
continue
indexes[name] = torch.zeros((p.size(0), matched_p_in_fm.size(0))).to(device)
indexes[name].requires_grad = True
logger.info(f'construct index in layer {name}')
tmp_indexes_file_path = os.path.join(self.res_save_dir, 'tmp-indexes.pt')
torch.save(indexes, tmp_indexes_file_path)
logger.info(f'generate indexes ({(os.path.getsize(tmp_indexes_file_path) / 1024**2):.3f}MB)')
os.remove(tmp_indexes_file_path)
optimizer = torch.optim.__dict__[hyps['optimizer']]([
# {'params': self.models['md'].models_dict['main'].parameters(), **hyps['optimizer_args']},
{'params': [v for v in indexes.values()], **hyps['indexes_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.
for p in self.models['md'].models_dict['main'].parameters():
p.requires_grad = False
for p in self.models['fm'].models_dict['main'].parameters():
p.requires_grad = False
for iter_index in pbar:
self.models['md'].to_eval_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)
# x, y = next(train_loader)
# x, y = x.to(device), y.to(device)
# task_loss = self.models['md'].forward_to_get_task_loss(x, y)
# l1_reg_loss = hyps['l1_reg_loss_weight'] * elastic_dnn_util.get_accu_l1_reg_of_raw_channel_attention_in_master_dnn(master_dnn)
index_loss = self.get_index_loss(self.models['fm'].models_dict['main'],
self.models['md'].models_dict['main'],
indexes,
self.models['md'].get_matched_param_of_fm,
hyps['index_guided_linear_comb_split_size'])
index_l1_loss = hyps['index_loss_l1_weight'] * torch.FloatTensor([v.abs().sum() for v in indexes.values()]).sum()
total_loss = index_loss + index_l1_loss
optimizer.zero_grad()
total_loss.backward()
optimizer.step()
scheduler.step()
if (iter_index + 1) % (10) == 0:
# visualize indexes histgoram
# do not use add_histogram because indexes is huge
os.makedirs(os.path.join(self.res_save_dir, 'index_hist'), exist_ok=True)
with torch.no_grad():
for p_name, index in indexes.items():
index_hist = index.view(-1).histc(bins=20).detach().cpu().numpy()
plt.bar(list(range(20)), index_hist)
plt.savefig(os.path.join(self.res_save_dir, f'index_hist/{p_name}.png'))
plt.clf()
if (iter_index + 1) % hyps['val_freq'] == 0:
elastic_dnn_util.clear_cached_channel_attention_in_master_dnn(self.models['md'].models_dict['main'])
cur_md = self.models['md'].models_dict['main']
md_for_test = deepcopy(self.models['md'].models_dict['main'])
val_accs = {}
avg_val_acc = 0.
bn_stats = {}
for val_sparsity in [0.0, 0.2, 0.4, 0.8]:
elastic_dnn_util.set_master_dnn_sparsity(md_for_test, val_sparsity)
bn_stats[f'{val_sparsity:.4f}'] = self.bn_cal(md_for_test, train_loader, hyps['bn_cal_num_iters'], device)
self.models['md'].models_dict['main'] = md_for_test
self.models['md'].to_eval_mode()
val_acc = self.models['md'].get_accuracy(val_loader)
val_accs[f'{val_sparsity:.4f}'] = val_acc
avg_val_acc += val_acc
avg_val_acc /= hyps['val_num_sparsities']
self.models['md'].models_dict['main'] = cur_md
self.models['md'].models_dict['indexes'] = indexes
self.models['md'].models_dict['bn_stats'] = bn_stats
self.models['fm'].models_dict['indexes'] = indexes
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 avg_val_acc > best_avg_val_acc:
best_avg_val_acc = avg_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(index=index_loss, index_l1=index_l1_loss, total=total_loss), iter_index)
pbar.set_description(f'loss: {total_loss:.6f}, index_loss: {index_loss:.6f}, index_l1_loss: {index_l1_loss:.6f}')
if (iter_index + 1) >= hyps['val_freq']:
tb_writer.add_scalars(f'accs/val_accs', val_accs, iter_index)
tb_writer.add_scalar(f'accs/avg_val_acc', avg_val_acc, iter_index)
pbar.set_description(f'loss: {total_loss:.6f}, index_loss: {index_loss:.6f}, index_l1_loss: {index_l1_loss:.6f}, '
f'avg_val_acc: {avg_val_acc:.4f}')
|