/
search_model_shell.py
262 lines (187 loc) · 8.93 KB
/
search_model_shell.py
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
import argparse
import copy
import gc
import csv
import shutil
import pycls.core.config as config
from pycls.core.config import cfg
from pycls.datasets.loader import _DATASETS
import pycls.core.logging as logging
import yaml
import random
from typing import Union
import time
import os
import subprocess
import torch.nn.functional as F
from pycls.models.build import MODEL
import numpy as np
import pycls.datasets.loader as data_loader
import torch
from torch import Tensor
from pycls.predictor.pruners.predictive import find_measures
from autozc.structures import GraphStructure, LinearStructure, TreeStructure
from pycls.models.build import MODEL
from autozc.utils.rank_consistency import kendalltau, pearson, spearman
logger = logging.get_logger(__name__)
def autoformer_configs(trial_num):
trial_configs = []
choices = {'num_heads': [3, 4], 'mlp_ratio': [3.5, 4.0],
'hidden_dim': [192, 216, 240], 'depth': [12, 13, 14]}
dimensions = ['mlp_ratio', 'num_heads']
for idx in range(trial_num):
flag = False
while not flag:
depth = random.choice(choices['depth'])
config = {
dimension: [
random.choice(choices[dimension]) for _ in range(depth)
]
for dimension in dimensions
}
config['hidden_dim'] = random.choice(choices['hidden_dim'])
config['depth'] = depth
if config not in trial_configs:
flag = True
trial_configs.append(config)
# logger.info(f'generate {idx}-th config: {config}')
return trial_configs
def pit_configs(trial_num, args):
trial_configs = []
logger.info('Pit param limit: {}--{}'.format(args.pit_low, args.pit_up))
choices = {'base_dim': [16, 24, 32, 40], 'mlp_ratio': [2, 4, 6, 8],
'num_heads': [[2,2,2], [2,2,4], [2,2,8], [2,4,4], [2,4,8], [2,8,8], [4,4,4], [4,4,8], [4,8,8], [8,8,8]],
'depth': [[1,6,6], [1,8,4], [2,4,6], [2,6,4], [2,6,6], [2,8,2], [2,8,4], [3,4,6], [3,6,4], [3,8,2]]}
for idx in range(trial_num):
flag = False
while not flag:
config = {}
dimensions = ['mlp_ratio', 'num_heads', 'base_dim', 'depth']
for dimension in dimensions:
config[dimension] = random.choice(choices[dimension])
temp_model = MODEL.get('PiT')(arch_config = config)
temp_params= sum(p.numel() for p in temp_model.parameters() if p.requires_grad)
if config not in trial_configs and args.pit_low <= round(temp_params/1e6) <= args.pit_up:
flag = True
trial_configs.append(config)
logger.info('generate {}-th config: {}, param: {} M'.format(idx, config, round(temp_params / 1e6)))
else:
logger.info('not suitable, param is:{} M'.format(round(temp_params/1e6)))
return trial_configs
def sample_trial_configs(model_type, args):
pop= None
trial_num = args.trial_num
if model_type =='AutoFormerSub':
pop = autoformer_configs(trial_num)
elif model_type == 'PiT':
pop = pit_configs(trial_num, args)
return pop
def obtain_zc(cfg, csv_path):
res = []
with open(csv_path, 'r') as file:
reader = csv.DictReader(file)
temp = [eval(row[cfg.PROXY_DATASET]) for row in reader]
res.append(temp)
return res[0]
def prepare_trials(cfg, arch_pop, xargs, log_dir, temp_dir):
bash_file = ['#!/bin/bash']
if not os.path.exists(temp_dir):
os.makedirs(temp_dir, exist_ok=True)
for idx, cand in enumerate(arch_pop):
trial_name = '{}-{}'.format(cfg.MODEL.TYPE, idx)
with open(xargs.refer_cfg) as f:
refer_data = yaml.safe_load(f)
trial_data = copy.deepcopy(refer_data)
if cfg.MODEL.TYPE =='PIT':
trial_data['PIT_SUBNET']['BASE_DIM']= cand['base_dim']
trial_data['PIT_SUBNET']['MLP_RATIO'] = cand['mlp_ratio']
trial_data['PIT_SUBNET']['DEPTH'] = cand['depth']
trial_data['PIT_SUBNET']['NUM_HEADS'] = cand['num_heads']
elif cfg.MODEL.TYPE =='AutoFormerSub':
trial_data['AUTOFORMER_SUBNET']['HIDDEN_DIM'] = cand['hidden_dim']
trial_data['AUTOFORMER_SUBNET']['MLP_RATIO'] = cand['mlp_ratio']
trial_data['AUTOFORMER_SUBNET']['DEPTH'] = cand['depth']
trial_data['AUTOFORMER_SUBNET']['NUM_HEADS'] = cand['num_heads']
with open(temp_dir+'/{}.yaml'.format(trial_name), 'w') as f:
yaml.safe_dump(trial_data, f, default_flow_style=False)
if cfg.AUTO_PROX.type == None:
execution_line = "CUDA_VISIBLE_DEVICES={} python proxy_zc.py --save_dir {} --csv --csv_dir {} --refer_cfg {}/{}.yaml --other_zc {} ".format(
xargs.gpu_idx, temp_dir , log_dir, temp_dir, trial_name, xargs.other_zc)
else:
execution_line = "CUDA_VISIBLE_DEVICES={} python proxy_zc.py --save_dir {} --csv --csv_dir {} --refer_cfg {}/{}.yaml".format(
xargs.gpu_idx, temp_dir, log_dir, temp_dir, trial_name)
bash_file.append(execution_line)
with open(os.path.join(temp_dir, 'run_bash.sh'), 'w') as handle:
for line in bash_file:
handle.write(line + os.linesep)
subprocess.call("sh {}/run_bash.sh".format(temp_dir), shell=True)
def obtain_optimal_model(cfg, args, arch_pop, log_dir):
if cfg.AUTO_PROX.type == None:
csv_path = os.path.join(log_dir, '{}_{}_{}.csv'.format(cfg.MODEL.TYPE, args.other_zc, cfg.PROXY_DATASET))
zc_name = args.other_zc
else:
csv_path = os.path.join(log_dir,
'{}_{}_{}.csv'.format(cfg.MODEL.TYPE, cfg.AUTO_PROX.type, cfg.PROXY_DATASET))
zc_name = cfg.AUTO_PROX.type
zc_score = obtain_zc(cfg, csv_path)
best_index = zc_score.index(max(zc_score))
best_score = zc_score[best_index]
best_cfg = arch_pop[best_index]
for arch_id, arch_cfg in enumerate(arch_pop):
logger.info(f'config: {arch_cfg} ')
logger.info(f'{zc_name} score: {zc_score[arch_id]}')
return best_cfg, best_score
if __name__ == '__main__':
parser = argparse.ArgumentParser( description='evo search zc',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--trial_num', default=1000,
type=int, help='number of mutate')
parser.add_argument('--save_dir', type=str, default='work_dirs/search_model_shell')
parser.add_argument('--gpu_idx', type=int, default=None)
parser.add_argument(
'--refer_cfg', default='./configs/auto/autoformer/autoformer-ti-subnet_c100_base.yaml', type=str,
help='save output path')
parser.add_argument(
'--other_zc',
default=None,
type=str,
help= 'size, epe_nas, nwot, grasp, snip, ntk, fisher, synflow, dss'
)
parser.add_argument("--pit_up", default=22, type=float, help="pit param upper limit")
parser.add_argument("--pit_low", default=4, type=float, help="pit param lower limit")
args = parser.parse_args()
config.load_cfg(args.refer_cfg)
config.assert_cfg()
logging.setup_logging()
time_str = time.strftime("%Y-%m-%d_%H-%M-%S", time.localtime(time.time()))
if cfg.AUTO_PROX.type == None:
log_dir = os.path.join(args.save_dir,
'{}_{}_{}_{}'.format(cfg.MODEL.TYPE, args.other_zc, cfg.PROXY_DATASET, time_str))
if not os.path.exists(log_dir):
os.makedirs(log_dir, exist_ok=True)
log_file = os.path.join(log_dir,
'rs_model_{}_{}_{}_{}.txt'.format(cfg.MODEL.TYPE, args.other_zc, cfg.PROXY_DATASET, time_str))
else:
log_dir = os.path.join(args.save_dir,
'{}_{}_{}_{}'.format(cfg.MODEL.TYPE, cfg.AUTO_PROX.type, cfg.PROXY_DATASET, time_str))
if not os.path.exists(log_dir):
os.makedirs(log_dir, exist_ok=True)
log_file = os.path.join(log_dir ,
'rs_model_{}_{}_{}_{}.txt'.format(cfg.MODEL.TYPE, cfg.AUTO_PROX.type, cfg.PROXY_DATASET,
time_str))
temp_dir = log_dir+'/temp'
file_handler = logging.FileHandler(log_file, 'w')
file_handler.setLevel(logging.INFO)
logger.addHandler(file_handler)
message = '\n'.join([f'{k:<20}: {v}' for k, v in vars(args).items()])
logger.info(message)
arch_pop = sample_trial_configs(cfg.MODEL.TYPE, args)
data_loader = data_loader.construct_proxy_loader()
t1 = time.time()
prepare_trials(cfg, arch_pop, args, log_dir, temp_dir)
# log_dir = os.path.join(args.save_dir,'AutoFormerSub_nwot_cifar100_2023-03-01_00-11-53')
best_cfg, best_score = obtain_optimal_model(cfg, args, arch_pop, log_dir)
t2 = time.time()
logger.info('Finished, time cost: {} hour.'.format((t2 - t1) / 3600))
logger.info(f'best arch config: {best_cfg}, best score: {best_score}')
shutil.rmtree(temp_dir)