-
Notifications
You must be signed in to change notification settings - Fork 1
/
prune_etenas.py
434 lines (384 loc) · 23.5 KB
/
prune_etenas.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
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
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
import os, sys, time, argparse
import json
import random
from easydict import EasyDict as edict
import numpy as np
import torch
from torch import nn
from tqdm import tqdm
from enum import Enum
from pathlib import Path
from typing import Dict, Any, Tuple
from collections import defaultdict
lib_dir = (Path(__file__).parent / 'lib').resolve()
if str(lib_dir) not in sys.path: sys.path.insert(0, str(lib_dir))
from datasets import get_datasets, get_nas_search_loaders
from procedures import prepare_seed, prepare_logger, MetricType, get_linear_region_counter_v2, get_nngp_n_v2, \
regional_division_counter, RegionDivisionScoreType, zen_score
from utils import get_model_infos, init_model, round_to, is_single_path
from log_utils import time_string
from models import get_cell_based_tiny_net, get_search_spaces
from nas_201_api import SimpleApi as API
INF = 1000 # used to mark prunned operators
class Ranker(Enum):
NNGP_frob = "Frobenius norm of NNGP"
NNGP_eig = "Eigenvalue score of NNGP"
NNGP_mean = "Mean value of NNGP"
NNGP_LGA = "Label-Gradient Alignment of NNGP"
NNGP_cond = "Condition number of NNGP"
NNGP = "NNGP"
REGS_num = "Expected number of ReLU regions"
REGS_dist = "ReLU regions distance"
ZEN = "Zen-Score"
def __str__(self):
return self.value
def is_override_batch(self):
if self is Ranker.REGS_num:
return True
if self is Ranker.REGS_dist:
return True
return False
def is_nngp_based(self):
if self.NNGP:
return True
if self.NNGP_frob:
return True
if self.NNGP_eig:
return True
if self.NNGP_mean:
return True
if self.NNGP_LGA:
return True
if self.NNGP_cond:
return True
return False
def prune_allowed_args(self, args: Dict[str, Any]):
output_args = {"sign": args.get("sign", 1)}
if self in [Ranker.NNGP_frob, Ranker.NNGP_eig, Ranker.NNGP_mean, Ranker.NNGP_LGA, Ranker.NNGP_cond]:
output_args["num_batch"] = args.get("num_batch", 2)
output_args["train_mode"] = args.get("train_mode", True)
if self in [Ranker.REGS_num, Ranker.REGS_dist]:
output_args["num_batch"] = 1
output_args["train_mode"] = args.get("train_mode", False)
if self is Ranker.REGS_num:
output_args["batch"] = args.get("batch", 1000)
if self is Ranker.REGS_dist:
output_args["batch"] = args.get("batch", 150)
if self is Ranker.ZEN:
output_args["train_mode"] = args.get("train_mode", False)
return output_args
def parse_rankers_config(path):
with open(path, "r") as file:
config = json.load(file)
rankers = []
for r in config:
r_type = Ranker(r["type"])
kwargs = Ranker.prune_allowed_args(r_type, r.get("args", {}))
if r_type is Ranker.NNGP_frob:
func = lambda train_loader, valid_loader, nets: get_nngp_n_v2(train_loader, valid_loader, nets,
metric=MetricType.FRO,
train_mode=kwargs["train_mode"],
as_correlation=True, use_logits=True,
num_batch=kwargs["num_batch"])
elif r_type is Ranker.NNGP_frob:
func = lambda train_loader, valid_loader, nets: get_nngp_n_v2(train_loader, valid_loader, nets,
metric=MetricType.FRO,
train_mode=kwargs["train_mode"],
as_correlation=True, use_logits=True,
num_batch=kwargs["num_batch"])
elif r_type is Ranker.NNGP_eig:
func = lambda train_loader, valid_loader, nets: get_nngp_n_v2(train_loader, valid_loader, nets,
metric=MetricType.EIG,
train_mode=kwargs["train_mode"],
as_correlation=False, use_logits=False,
num_batch=kwargs["num_batch"])
elif r_type is Ranker.NNGP_mean:
func = lambda train_loader, valid_loader, nets: get_nngp_n_v2(train_loader, valid_loader, nets,
metric=MetricType.MEAN,
train_mode=kwargs["train_mode"],
as_correlation=True, use_logits=True,
num_batch=kwargs["num_batch"])
elif r_type is Ranker.NNGP_LGA:
func = lambda train_loader, valid_loader, nets: get_nngp_n_v2(train_loader, valid_loader, nets,
metric=MetricType.LGA,
train_mode=kwargs["train_mode"],
as_correlation=False, use_logits=True,
num_batch=kwargs["num_batch"])
elif r_type is Ranker.NNGP_cond:
func = lambda train_loader, valid_loader, nets: get_nngp_n_v2(train_loader, valid_loader, nets,
metric=MetricType.COND,
train_mode=kwargs["train_mode"],
as_correlation=False, use_logits=True,
num_batch=kwargs["num_batch"])
elif r_type is Ranker.REGS_num:
func = lambda train_loader, _, nets: get_linear_region_counter_v2(train_loader, nets,
train_mode=kwargs["train_mode"],
num_batch=kwargs["num_batch"])
elif r_type is Ranker.REGS_dist:
func = lambda train_loader, _, nets: regional_division_counter(train_loader, nets,
train_mode=kwargs["train_mode"],
score_type=RegionDivisionScoreType.FULL,
num_batch=kwargs["num_batch"], verbose=False)
elif r_type is Ranker.ZEN:
func = lambda train_loader, _, nets: zen_score(train_loader, nets, train_mode=kwargs["train_mode"])
else:
raise RuntimeError("invalid ranker")
rankers.append((r_type, func, kwargs))
return rankers
def prune_func_rank(rankers_list, arch_parameters, model_config, special_model_configs, train_loader, valid_loader,
special_dataloaders,
edge_groups=None, init="kaiming_norm", repeat=1, prune_number=1):
for alpha in arch_parameters:
alpha[:, 0] = -INF
# set neural networks
network_origin = get_cell_based_tiny_net(model_config).cuda().train()
# init_model(network_origin, init)
network_origin.set_alphas(arch_parameters)
special_networks_origins = {}
for r, c in special_model_configs.items():
special_networks_origins[r] = get_cell_based_tiny_net(c).cuda().train()
# init_model(special_networks_origins[r], init)
special_networks_origins[r].set_alphas(arch_parameters)
alpha_active = [(nn.functional.softmax(alpha, 1) > 0.01).float() for alpha in arch_parameters]
if edge_groups is None:
prune_number = min(prune_number,
alpha_active[0][0].sum() - 1) # adjust prune_number based on current remaining ops on each edge
active_parameter_indexes = [(idx_ct, idx_edge, idx_op)
for idx_ct in range(len(arch_parameters))
for idx_edge in range(arch_parameters[idx_ct].shape[0])
for idx_op in range(arch_parameters[idx_ct].shape[1])
if alpha_active[idx_ct][idx_edge].sum() != 1 # more than one op remaining
if alpha_active[idx_ct][idx_edge, idx_op] > 0 # op is active
]
else:
active_parameter_indexes = [(idx_ct, idx_edge, idx_op)
for idx_ct in range(len(arch_parameters))
for group_st, group_end in edge_groups
for idx_edge in range(group_st, group_end)
for idx_op in range(arch_parameters[idx_ct].shape[1])
if group_end - group_st > prune_number
if alpha_active[idx_ct][idx_edge, idx_op] > 0 # op is active
]
edge_scores = []
for edge_indexes in tqdm(active_parameter_indexes):
(idx_ct, idx_edge, idx_op) = edge_indexes
_arch_param = [alpha.detach().clone() for alpha in arch_parameters]
_arch_param[idx_ct][idx_edge, idx_op] = -INF
network = get_cell_based_tiny_net(model_config).cuda().train()
special_networks = {r: get_cell_based_tiny_net(special_model_configs[r]).cuda().train()
for r in special_model_configs
}
_scores = []
for _ in range(repeat):
ranker_scores = []
for ranker, rank_func, kwargs in rankers_list:
### initializing networks for backward
init_model(network_origin, init + "_fanout" if init.startswith('kaiming') else init)
# make sure network_origin and network are identical
for param_ori, param in zip(network_origin.parameters(), network.parameters()):
param.data.copy_(param_ori.data)
network.set_alphas(_arch_param)
for k in special_networks_origins:
init_model(special_networks_origins[k], init + "_fanout" if init.startswith('kaiming') else init)
# make sure network_thin and network_thin_origin are identical
for param_ori, param in zip(special_networks_origins[k].parameters(),
special_networks[k].parameters()):
param.data.copy_(param_ori.data)
special_networks[k].set_alphas(_arch_param)
origin_net = special_networks_origins.get(ranker, network_origin)
net = special_networks.get(ranker, network)
train_loader_, valid_loader_ = special_dataloaders.get(ranker, (train_loader, valid_loader))
origin_rank, rank = rank_func(train_loader_, valid_loader_, [origin_net, net])
percent_shift = kwargs["sign"] * (origin_rank - rank) / origin_rank
ranker_scores.append(percent_shift)
_scores.append(np.sum(ranker_scores))
edge_scores.append((edge_indexes, np.mean(_scores)))
edge_scores = sorted(edge_scores,
key=lambda tup: tup[1]) # list of (cell_idx, edge_idx, op_idx), [ntk_rank, regions_rank]
edge2choice = defaultdict(list)
for (cell_idx, edge_idx, op_idx), _ in edge_scores:
if edge_groups is None:
if len(edge2choice[(cell_idx, edge_idx)]) < prune_number:
edge2choice[(cell_idx, edge_idx)].append((cell_idx, edge_idx, op_idx))
else:
for group_st, group_end in edge_groups:
if group_st <= edge_idx < group_end and len(edge2choice[(group_st, group_end)]) < (group_end - group_st - prune_number):
edge2choice[(group_st, group_end)].append((cell_idx, edge_idx, op_idx))
for choices in edge2choice.values():
for (cell_idx, edge_idx, op_idx) in choices:
arch_parameters[cell_idx].data[edge_idx, op_idx] = -INF
return arch_parameters
def main(xargs):
PID = os.getpid()
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
prepare_seed(xargs.rand_seed)
if xargs.timestamp == 'none':
xargs.timestamp = "{:}".format(time.strftime('%h-%d-%C_%H-%M-%s', time.gmtime(time.time())))
xargs.save_dir = os.path.join(xargs.save_dir,
"repeat{}-prunNum{}-{}-batch{}".format(xargs.repeat, xargs.prune_number, xargs.init, xargs.batch_size),
str(xargs.timestamp),
"seed{}".format(xargs.rand_seed))
# checking ranking list
rankers_list = parse_rankers_config(xargs.rankers_config)
batch_size = xargs.batch_size
train_data, valid_data, xshape, class_num = get_datasets(xargs.dataset, xargs.data_path, cutout=-1)
_, train_loader, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset,
'configs/dataset_split/', batch_size, xargs.workers)
special_dataloaders = {}
for r, _, kwargs in rankers_list:
if Ranker.is_override_batch(r):
if kwargs["batch"] != batch_size:
input_size = None
if r is Ranker.REGS_num:
input_size = (kwargs["batch"], 1, 3, 3)
train_data_, valid_data_, _, _ = get_datasets(xargs.dataset, xargs.data_path,
input_size=input_size, cutout=-1)
_, train_loader_, valid_loader_ = get_nas_search_loaders(train_data_, valid_data_, xargs.dataset,
'configs/dataset_split/', kwargs["batch"],
xargs.workers)
special_dataloaders[r] = (train_loader_, valid_loader_)
##### config & logging #####
logger = prepare_logger(xargs)
logger.log(f'Batch size : {xargs.batch_size}')
logger.log(f'Input image shape : {xshape[1:]}')
logger.log(f'Saving dir : {xargs.save_dir}')
logger.log(f'Rankers :')
for r, _, args in rankers_list:
logger.log(f'\t{r.value}: {args}')
logger.log('||||||| {:10s} ||||||| Train-Loader-Num={:}'.format(xargs.dataset, len(train_loader)))
###############
search_space = get_search_spaces('cell', xargs.search_space_name)
special_model_configs = {}
if xargs.search_space_name == 'nas-bench-201':
model_config = edict({'name': 'DARTS-V1',
'C': 3, 'N': 1, 'depth': -1, 'use_stem': True,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space': search_space,
'affine': True, 'track_running_stats': bool(xargs.track_running_stats),
})
if Ranker.REGS_num in special_dataloaders:
special_model_configs[Ranker.REGS_num] = edict({'name': 'DARTS-V1',
'C': 1, 'N': 1, 'depth': 1, 'use_stem': False,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space': search_space,
'affine': True,
'track_running_stats': bool(xargs.track_running_stats),
})
elif xargs.search_space_name == 'darts':
model_config = edict({'name': 'DARTS-V1',
'C': 1, 'N': 1, 'depth': 2, 'use_stem': True, 'stem_multiplier': 1,
'num_classes': class_num,
'space': search_space,
'affine': True, 'track_running_stats': bool(xargs.track_running_stats),
'super_type': xargs.super_type,
'steps': 4,
'multiplier': 4,
})
if Ranker.REGS_num in special_dataloaders:
special_model_configs[Ranker.REGS_num] = edict({'name': 'DARTS-V1',
'C': 1, 'N': 1, 'depth': 2, 'use_stem': False,
'stem_multiplier': 1,
'max_nodes': xargs.max_nodes, 'num_classes': class_num,
'space': search_space,
'affine': True,
'track_running_stats': bool(xargs.track_running_stats),
'super_type': xargs.super_type,
'steps': 4,
'multiplier': 4,
})
else:
raise RuntimeError(f"{xargs.search_space_name} is not a valid search space")
logger.log('model-config : {:}'.format(model_config))
logger.log('special-model-configs : {:}'.format(special_model_configs))
network = get_cell_based_tiny_net(model_config)
# ### all params trainable (except train_bn) #########################
flop, param = get_model_infos(network, xshape)
logger.log('FLOP = {:.2f} M, Params = {:.2f} MB'.format(flop, param))
logger.log('search-space [{:} ops] : {:}'.format(len(search_space), search_space))
network = network.cuda()
genotypes = {-1: network.genotype()}
arch_parameters = [alpha.detach().clone() for alpha in network.get_alphas()]
for alpha in arch_parameters:
alpha[:, 1:] = 0
alpha[:, 0] = -INF
start_time = time.time()
epoch = -1
while not is_single_path(network):
epoch += 1
torch.cuda.empty_cache()
print("<< ============== JOB (PID = %d) %s ============== >>" % (PID, '/'.join(xargs.save_dir.split("/")[-6:])))
arch_parameters = prune_func_rank(rankers_list, arch_parameters, model_config, special_model_configs,
train_loader, valid_loader, special_dataloaders,
init=xargs.init,
repeat=xargs.repeat,
prune_number=xargs.prune_number
)
# rebuild supernet
network = get_cell_based_tiny_net(model_config)
network = network.cuda()
network.set_alphas(arch_parameters)
genotypes[epoch] = network.genotype()
logger.log('operators remaining (1s) and prunned (0s)\n{:}'.format(
'\n'.join([str((alpha > -INF).int()) for alpha in network.get_alphas()])))
if xargs.search_space_name == 'darts':
print("===>>> Prune Edge Groups...")
arch_parameters = prune_func_rank(rankers_list, arch_parameters, model_config, special_model_configs,
train_loader, valid_loader, special_dataloaders,
init=xargs.init,
repeat=xargs.repeat,
edge_groups=[(0, 2), (2, 5), (5, 9), (9, 14)],
prune_number=2
)
network = get_cell_based_tiny_net(model_config)
network = network.cuda()
network.set_alphas(arch_parameters)
logger.log('<<<--->>> End: {:}'.format(network.genotype()))
logger.log('operators remaining (1s) and prunned (0s)\n{:}'.format(
'\n'.join([str((alpha > -INF).int()) for alpha in network.get_alphas()])))
end_time = time.time()
logger.log('\n' + '-' * 100)
logger.log(f"Time spent: {end_time - start_time} sec")
# write final parameters into file
arch_parameters_npy = [[alpha.detach().clone().cpu().numpy() for alpha in arch_parameters]]
np.save(os.path.join(xargs.save_dir, "arch_parameters_history.npy"), arch_parameters_npy)
logger.log(genotypes[epoch] if type(genotypes[epoch]) is str else genotypes[epoch].tostr())
# check the performance from the architecture dataset (for NAS-Bench-201)
if xargs.arch_nas_dataset is not None and xargs.search_space_name != 'darts':
api = API(xargs.arch_nas_dataset)
logger.log('{:} create API = {:} done'.format(time_string(), api))
idx = api.query_index_by_arch(genotypes[epoch])
acc = api.query_by_index(idx)[xargs.dataset]
logger.log('Test Accuracy {} on {}'.format(acc, xargs.dataset))
logger.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser("ETENAS")
parser.add_argument('--data_path', type=str, help='Path to dataset')
parser.add_argument('--rankers_config', type=str, help='path to rankers config')
parser.add_argument('--dataset', type=str, choices=['cifar10', 'cifar100', 'ImageNet16-120'],
help='Choose between cifar10/100/ImageNet16-120/imagenet-1k')
parser.add_argument('--search_space_name', type=str, default='nas-bench-201',
help='space of operator candidates: nas-bench-201 or darts.')
parser.add_argument('--max_nodes', type=int, help='The maximum number of nodes.')
parser.add_argument('--track_running_stats', type=int, choices=[0, 1],
help='Whether use track_running_stats or not in the BN layer.')
parser.add_argument('--workers', type=int, default=0, help='number of data loading workers (default: 0)')
parser.add_argument('--batch_size', type=int, default=16, help='batch size for ntk')
parser.add_argument('--save_dir', type=str, help='Folder to save checkpoints and log.')
parser.add_argument('--arch_nas_dataset', type=str,
help='The path to load the nas-bench-201 architecture dataset (tiny-nas-benchmark).')
parser.add_argument('--rand_seed', type=int, help='manual seed')
parser.add_argument('--prune_number', type=int, default=1,
help='number of operator to prune on each edge per round')
parser.add_argument('--repeat', type=int, default=3, help='repeat calculation of ranking functions')
parser.add_argument('--timestamp', default='none', type=str, help='timestamp for logging naming')
parser.add_argument('--init', default='kaiming_norm',
choices=['kaiming_norm', 'kaiming_norm_fanin', 'kaiming_norm_fanout'],
help='initialization to use')
parser.add_argument('--super_type', type=str, default='basic', help='type of supernet: basic or nasnet-super')
args = parser.parse_args()
if args.rand_seed is None or args.rand_seed < 0:
args.rand_seed = random.randint(1, 100000)
main(args)