-
Notifications
You must be signed in to change notification settings - Fork 6
/
evolution_pre_train.py
606 lines (519 loc) · 25.5 KB
/
evolution_pre_train.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
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
import random
import numpy as np
import time
import torch
import torch.backends.cudnn as cudnn
from pathlib import Path
from lib.datasets import build_dataset
from lib import utils
from supernet_engine import evaluate
from model.supernet_transformer import Vision_TransformerSuper
import argparse
import os
import yaml
from lib.config import cfg, update_config_from_file
from lib.score_maker import ScoreMaker
import math
from itertools import combinations
import json
def decode_cand_tuple(cand_tuple):
depth = cand_tuple[0]
return depth, list(cand_tuple[1:depth+1]), list(cand_tuple[depth + 1: 2 * depth + 1]), cand_tuple[-1]
def get_max_min_model(choices):
max_depth = max(choices['depth'])
max_emb = max(choices['embed_dim'])
max_num_head = max(choices['num_heads'])
max_mlp_ratio = max(choices['mlp_ratio'])
min_depth = min(choices['depth'])
min_emb = min(choices['embed_dim'])
min_num_head = min(choices['num_heads'])
min_mlp_ratio = min(choices['mlp_ratio'])
max_model = tuple([max_depth] + [max_mlp_ratio] * max_depth + [max_num_head] * max_depth + [max_emb])
min_model = tuple([min_depth] + [min_mlp_ratio] * min_depth + [min_num_head] * min_depth + [min_emb])
return max_model, min_model
class Searcher(object):
def __init__(self, args, device, model, model_without_ddp, choices, output_dir, score_maker):
self.device = device
self.model = model
self.model_without_ddp = model_without_ddp
self.args = args
self.max_epochs = args.max_epochs
self.select_num = args.select_num
self.population_num = args.population_num
self.m_prob = args.m_prob
self.crossover_num = args.crossover_num
self.mutation_num = args.mutation_num
self.parameters_limits = args.param_limits
self.min_parameters_limits = args.min_param_limits
self.output_dir = output_dir
self.s_prob =args.s_prob
self.memory = []
self.vis_dict = {}
self.keep_top_k = {}
self.epoch = 0
self.checkpoint_path = args.resume
self.candidates = []
self.top_accuracies = []
self.cand_params = []
self.choices = choices
self.choices['num_heads'].sort()
self.choices['mlp_ratio'].sort()
self.score_maker = score_maker
self.eval_cnt = 0
self.update_num = 0
self.un_update_cnt = 0
self.all_cands = {}
min_param = self.min_parameters_limits
max_param = min_param + self.args.param_interval
while max_param < self.parameters_limits + 1e-6:
params = (max_param + min_param) / 2
self.all_cands[self.param_to_index(params)] = []
min_param = max_param
max_param = min_param + self.args.param_interval
self.cur_min_param = args.min_param_limits
self.cur_max_param = args.param_limits
self.interval_cands = {}
self.max_model, self.min_model = get_max_min_model(choices)
self.search_mode = args.search_mode
self.head_mlp_scores = {}
def get_params_range(self):
depth, mlp_ratio, num_heads, embed_dim = decode_cand_tuple(self.max_model)
sampled_config = {}
sampled_config['layer_num'] = depth
sampled_config['mlp_ratio'] = mlp_ratio
sampled_config['num_heads'] = num_heads
sampled_config['embed_dim'] = [embed_dim] * depth
n_parameters = self.model_without_ddp.get_sampled_params_numel(sampled_config)
max_params = n_parameters / 10. ** 6
depth, mlp_ratio, num_heads, embed_dim = decode_cand_tuple(self.min_model)
sampled_config = {}
sampled_config['layer_num'] = depth
sampled_config['mlp_ratio'] = mlp_ratio
sampled_config['num_heads'] = num_heads
sampled_config['embed_dim'] = [embed_dim] * depth
n_parameters = self.model_without_ddp.get_sampled_params_numel(sampled_config)
min_params = n_parameters / 10. ** 6
return min_params, max_params
def select_cands(self, *, key, reverse=True):
for k in self.all_cands.keys():
t = self.all_cands[k]
t.sort(key=key, reverse=reverse)
self.all_cands[k] = t[:self.args.cand_per_interval]
def param_to_index(self, param):
if param < self.min_parameters_limits:
return -1
if param >= self.parameters_limits:
return -1
return math.floor((param - self.min_parameters_limits) / self.args.param_interval)
def index_to_param_interval(self, index):
if index == -1:
return (0, self.min_parameters_limits)
if index == -2:
return (self.parameters_limits, 2*self.parameters_limits)
down = self.min_parameters_limits + index * self.args.param_interval
up = down + self.args.param_interval
return (down, up)
def stack_random_cand(self, random_func, *, batchsize=10):
while True:
cands = [random_func() for _ in range(batchsize)]
for cand in cands:
if cand not in self.vis_dict:
self.vis_dict[cand] = {}
info = self.vis_dict[cand]
for cand in cands:
yield cand
def get_random_cand_without_reallocate(self):
cand_tuple = list()
dimensions = ['mlp_ratio', 'num_heads']
depth = random.choice(self.choices['depth'])
cand_tuple.append(depth)
for dimension in dimensions:
idx = list(range(len(self.choices[dimension])))
random.shuffle(idx)
choice_cnt = {}
left_layers = depth
for i in idx[:-1]:
choice = self.choices[dimension][i]
cnt = random.choice(range(left_layers + 1))
left_layers = left_layers - cnt
choice_cnt[choice] = cnt
choice = self.choices[dimension][idx[-1]]
choice_cnt[choice] = left_layers
conf = [0] * depth
for choice in self.choices[dimension][1:][::-1]:
scores = np.random.rand(depth)
mask = np.where(np.array(conf) > 0, -1, 1)
mask_scores = scores * mask
for i in mask_scores.argsort()[::-1][:choice_cnt[choice]]:
conf[i] = choice
for i in range(len(conf)):
if conf[i] == 0:
conf[i] = self.choices[dimension][0]
cand_tuple.extend(conf)
cand_tuple.append(random.choice(self.choices['embed_dim']))
return tuple(cand_tuple)
def get_random_cand(self):
cand_tuple = list()
dimensions = ['mlp_ratio', 'num_heads']
score_names = ['mlp_scores', 'head_scores']
depth = random.choice(self.choices['depth'])
cand_tuple.append(depth)
emb_dim = random.choice(self.choices['embed_dim'])
max_dim = max(self.choices['embed_dim'])
for (dimension, score_name) in zip(dimensions, score_names):
idx = list(range(len(self.choices[dimension])))
random.shuffle(idx)
choice_cnt = {}
left_layers = depth
for i in idx[:-1]:
choice = self.choices[dimension][i]
cnt = random.choice(range(left_layers + 1))
left_layers = left_layers - cnt
choice_cnt[choice] = cnt
choice = self.choices[dimension][idx[-1]]
choice_cnt[choice] = left_layers
choice_cnt_list = [choice_cnt[choice] for choice in self.choices[dimension]]
method = None
if dimension == 'mlp_ratio':
method = self.args.block_score_method_for_mlp
else:
method = self.args.block_score_method_for_head
cand_tuple.extend(self.reallocate(depth,
emb_dim,
dimension,
self.head_mlp_scores[score_name],
choice_cnt_list,
method))
cand_tuple.append(emb_dim)
return tuple(cand_tuple)
def get_random(self, num):
print('random select ........')
if self.args.search_mode == 'iteration' or self.args.reallocate:
cand_iter = self.stack_random_cand(self.get_random_cand)
else:
cand_iter = self.stack_random_cand(self.get_random_cand_without_reallocate)
while len(self.candidates) < num:
cand = next(cand_iter)
if not self.is_legal(cand):
continue
self.candidates.append(cand)
print('random {}/{}'.format(len(self.candidates), num))
print('random_num = {}'.format(len(self.candidates)))
def is_legal(self, cand):
assert isinstance(cand, tuple)
if cand not in self.vis_dict:
self.vis_dict[cand] = {}
info = self.vis_dict[cand]
if 'visited' in info:
return False
depth, mlp_ratio, num_heads, embed_dim = decode_cand_tuple(cand)
sampled_config = {}
sampled_config['layer_num'] = depth
sampled_config['mlp_ratio'] = mlp_ratio
sampled_config['num_heads'] = num_heads
sampled_config['embed_dim'] = [embed_dim]*depth
n_parameters = self.model_without_ddp.get_sampled_params_numel(sampled_config)
info['params'] = n_parameters / 10.**6
if info['params'] > self.cur_max_param:
print('parameters limit exceed {}'.format(self.cur_max_param))
return False
if info['params'] < self.cur_min_param:
print('under minimum parameters limit {}'.format(self.cur_min_param))
return False
info['visited'] = True
return True
def conf_to_cnt_list(self, conf, part):
cnt_list = [0]*len(self.choices[part])
for choice in conf:
cnt_list[self.choices[part].index(choice)] += 1
return cnt_list
def reallocate(self, depth, embed_dim, part, scores, choice_cnt, method):
if method == 'deeper_is_better':
conf = []
for choice, cnt in zip(self.choices[part], choice_cnt):
conf = conf + ([choice] * cnt)
return conf
if 'max_dim' in method:
embed_dim = max(self.choices['embed_dim'])
conf = [0] * depth
for choice, cnt in zip(self.choices[part][1:][::-1], choice_cnt[1:][::-1]):
cur_scores = np.array(scores[(f"{embed_dim},{choice}")][:depth])
mask = np.where(np.array(conf) > 0, -1, 1)
mask_scores = cur_scores * mask
for i in mask_scores.argsort()[::-1][:cnt]:
conf[i] = choice
for i in range(len(conf)):
if conf[i] == 0:
conf[i] = self.choices[part][0]
return conf
def get_score(self):
for cand in self.candidates:
info = self.vis_dict[cand]
if self.args.score_method == 'params':
info['score'] = info['params']
else:
depth, mlp_ratio, num_heads, embed_dim = decode_cand_tuple(cand)
sampled_config = {}
sampled_config['layer_num'] = depth
sampled_config['mlp_ratio'] = mlp_ratio
sampled_config['num_heads'] = num_heads
sampled_config['embed_dim'] = [embed_dim] * depth
score = self.score_maker.get_score(self.model, self.args.score_method, config=sampled_config)
info['score'] = score
def update_top_k(self, candidates, *, k, key, reverse=True, get_update_num=False):
assert k in self.keep_top_k
print('select ......')
t = self.keep_top_k[k]
t += candidates
t.sort(key=key, reverse=reverse)
self.keep_top_k[k] = t[:k]
if get_update_num:
self.update_num = 0
for cand in self.keep_top_k[k]:
if cand in candidates:
self.update_num += 1
print('update {} models in top {}.'.format(self.update_num, k))
if self.update_num == 0:
self.un_update_cnt += 1
def get_mutation(self, k, mutation_num, m_prob, s_prob):
assert k in self.keep_top_k
print('mutation ......')
res = []
iter = 0
max_iters = mutation_num * 10
def random_func():
cand = list(random.choice(self.keep_top_k[k]))
depth, mlp_ratio, num_heads, embed_dim = decode_cand_tuple(cand)
random_s = random.random()
# depth
if random_s < s_prob:
new_depth = random.choice(self.choices['depth'])
if new_depth > depth:
mlp_ratio = mlp_ratio + [random.choice(self.choices['mlp_ratio']) for _ in range(new_depth - depth)]
num_heads = num_heads + [random.choice(self.choices['num_heads']) for _ in range(new_depth - depth)]
else:
mlp_ratio = mlp_ratio[:new_depth]
num_heads = num_heads[:new_depth]
depth = new_depth
# mlp_ratio
for i in range(depth):
random_s = random.random()
if random_s < m_prob:
mlp_ratio[i] = random.choice(self.choices['mlp_ratio'])
# num_heads
for i in range(depth):
random_s = random.random()
if random_s < m_prob:
num_heads[i] = random.choice(self.choices['num_heads'])
# embed_dim
random_s = random.random()
if random_s < s_prob:
embed_dim = random.choice(self.choices['embed_dim'])
mlp_cnt = self.conf_to_cnt_list(mlp_ratio, 'mlp_ratio')
head_cnt = self.conf_to_cnt_list(num_heads, 'num_heads')
mlp_ratio = self.reallocate(depth,
embed_dim,
'mlp_ratio',
self.head_mlp_scores['mlp_scores'],
mlp_cnt,
self.args.block_score_method_for_mlp)
num_heads = self.reallocate(depth,
embed_dim,
'num_heads',
self.head_mlp_scores['head_scores'],
head_cnt,
self.args.block_score_method_for_head)
result_cand = [depth] + mlp_ratio + num_heads + [embed_dim]
return tuple(result_cand)
cand_iter = self.stack_random_cand(random_func)
while len(res) < mutation_num and max_iters > 0:
max_iters -= 1
cand = next(cand_iter)
if not self.is_legal(cand):
continue
res.append(cand)
print('mutation {}/{}'.format(len(res), mutation_num))
print('mutation_num = {}'.format(len(res)))
return res
def get_crossover(self, k, crossover_num):
assert k in self.keep_top_k
print('crossover ......')
res = []
iter = 0
max_iters = 10 * crossover_num
def random_func():
p1 = random.choice(self.keep_top_k[k])
p2 = random.choice(self.keep_top_k[k])
max_iters_tmp = 50
while len(p1) != len(p2) and max_iters_tmp > 0:
max_iters_tmp -= 1
p1 = random.choice(self.keep_top_k[k])
p2 = random.choice(self.keep_top_k[k])
cand = tuple(random.choice([i, j]) for i, j in zip(p1, p2))
depth, mlp_ratio, num_heads, embed_dim = decode_cand_tuple(cand)
mlp_cnt = self.conf_to_cnt_list(mlp_ratio, 'mlp_ratio')
head_cnt = self.conf_to_cnt_list(num_heads, 'num_heads')
mlp_ratio = self.reallocate(depth,
embed_dim,
'mlp_ratio',
self.head_mlp_scores['mlp_scores'],
mlp_cnt,
self.args.block_score_method_for_mlp)
num_heads = self.reallocate(depth,
embed_dim,
'num_heads',
self.head_mlp_scores['head_scores'],
head_cnt,
self.args.block_score_method_for_head)
result_cand = [depth] + mlp_ratio + num_heads + [embed_dim]
return tuple(result_cand)
cand_iter = self.stack_random_cand(random_func)
while len(res) < crossover_num and max_iters > 0:
max_iters -= 1
cand = next(cand_iter)
if not self.is_legal(cand):
continue
res.append(cand)
print('crossover {}/{}'.format(len(res), crossover_num))
print('crossover_num = {}'.format(len(res)))
return res
def search(self, out_file_name=None):
print('searching...')
if not self.args.block_score_method_for_mlp == 'deeper_is_better' or not self.args.block_score_method_for_head == 'deeper_is_better':
self.head_mlp_scores = self.score_maker.get_block_scores(self.model, self.args, self.choices)
# random search
if self.args.search_mode == 'random':
self.cur_min_param = self.min_parameters_limits
self.cur_max_param = self.cur_min_param + self.args.param_interval
while self.cur_max_param < self.parameters_limits + 1e-6:
self.candidates = []
self.keep_top_k = {100: []}
self.get_random(self.population_num)
self.get_score()
self.update_top_k(
self.candidates, k=100, key=lambda x: self.vis_dict[x]['score'])
for i, cand in enumerate(self.keep_top_k[100]):
print('No.{} {} score = {}, params = {}'.format(
i + 1, cand, self.vis_dict[cand]['score'], self.vis_dict[cand]['params']))
self.interval_cands[(self.cur_min_param, self.cur_max_param)] = self.keep_top_k[100][:self.args.cand_per_interval]
self.cur_min_param = self.cur_max_param
self.cur_max_param = self.cur_min_param + self.args.param_interval
# evolution search
elif self.args.search_mode == 'evolution':
self.cur_min_param = self.min_parameters_limits
self.cur_max_param = self.cur_min_param + self.args.param_interval
while self.cur_max_param < self.parameters_limits + 1e-6:
self.update_num = 0
self.un_update_cnt = 0
self.epoch = 0
self.candidates = []
self.keep_top_k = {self.select_num: [], 100: []}
self.get_random(self.population_num)
while self.epoch < self.max_epochs:
print('epoch = {} for param {} to param {}'.format(self.epoch, self.cur_min_param, self.cur_max_param))
if self.un_update_cnt == 2:
self.epoch += 1
continue
self.get_score()
self.update_top_k(
self.candidates, k=self.select_num, key=lambda x: self.vis_dict[x]['score'], get_update_num=True)
self.update_top_k(
self.candidates, k=100, key=lambda x: self.vis_dict[x]['score'])
print('epoch = {} for param {} to param {} : top {} result'.format(
self.epoch, self.cur_min_param, self.cur_max_param, len(self.keep_top_k[100])))
for i, cand in enumerate(self.keep_top_k[100]):
print('No.{} {} score = {}, params = {}'.format(
i + 1, cand, self.vis_dict[cand]['score'], self.vis_dict[cand]['params']))
self.epoch += 1
if self.epoch >= self.max_epochs:
break
# check
mutation = self.get_mutation(
self.select_num, self.mutation_num, self.m_prob, self.s_prob)
crossover = self.get_crossover(self.select_num, self.crossover_num)
self.candidates = mutation + crossover
self.get_random(self.population_num)
self.interval_cands[(self.cur_min_param, self.cur_max_param)] = self.keep_top_k[100][:self.args.cand_per_interval]
self.cur_min_param = self.cur_max_param
self.cur_max_param = self.cur_min_param + self.args.param_interval
# force search
else:
max_dim = max(self.choices['embed_dim'])
iter_cnt = 0
for embed_dim in self.choices['embed_dim']:
for depth in self.choices['depth']:
depth_ids = list(range(depth+1))
num_head_choice = len(self.choices['num_heads'])
num_mlp_choice = len(self.choices['mlp_ratio'])
mlp_confs = []
head_confs = []
for mlp_dist in combinations(depth_ids, num_mlp_choice - 1):
mlp_dist = [0] + list(mlp_dist) + [depth]
mlp_cnt = [mlp_dist[i+1] - mlp_dist[i] for i in range(len(mlp_dist)-1)]
mlp_confs.append(self.reallocate(depth,
embed_dim,
'mlp_ratio',
self.head_mlp_scores['mlp_scores'],
mlp_cnt,
self.args.block_score_method_for_mlp))
for head_dist in combinations(depth_ids, num_head_choice - 1):
head_dist = [0] + list(head_dist) + [depth]
head_cnt = [head_dist[i+1] - head_dist[i] for i in range(len(head_dist)-1)]
head_confs.append(self.reallocate(depth,
embed_dim,
'num_heads',
self.head_mlp_scores['head_scores'],
head_cnt,
self.args.block_score_method_for_head))
for mlp_conf in mlp_confs:
iter_cnt += 1
for head_conf in head_confs:
cand = tuple([depth] + mlp_conf + head_conf + [embed_dim])
depth, mlp_ratio, num_heads, embed_dim = decode_cand_tuple(cand)
sampled_config = {}
sampled_config['layer_num'] = depth
sampled_config['mlp_ratio'] = mlp_ratio
sampled_config['num_heads'] = num_heads
sampled_config['embed_dim'] = [embed_dim] * depth
n_parameters = self.model_without_ddp.get_sampled_params_numel(sampled_config)
params = n_parameters / 10. ** 6
index = self.param_to_index(params)
if self.args.score_method == 'params':
score = params
else:
score = self.score_maker.get_score(self.model, self.args.score_method, config=sampled_config)
info = {'cand': cand, 'score': score, 'params': params}
self.vis_dict[cand] = info
if index in self.all_cands.keys():
self.all_cands[index].append(info)
self.select_cands(key=lambda x: x['score'])
for index in self.all_cands.keys():
k = self.index_to_param_interval(index)
self.interval_cands[k] = [item['cand'] for item in self.all_cands[index]]
if out_file_name is None:
out_file_name = f'out/interval_cands_{self.args.super_model_size}_{self.args.score_method}_{self.args.block_score_method_for_mlp}_for_mlp_{self.args.block_score_method_for_head}_for_head'
out_file_name += f'_i{self.args.param_interval}_top_{self.args.cand_per_interval}.pt'
torch.save(self.interval_cands, out_file_name)
else:
json_dict = {}
for interval in self.interval_cands.keys():
cand_list = []
for cand in self.interval_cands[interval]:
depth, mlp_ratio, num_heads, embed_dim = decode_cand_tuple(cand)
info = {
'layer_num': depth,
'mlp_ratio': mlp_ratio,
'num_heads': num_heads,
'embed_dim': [embed_dim]*depth,
'num_params': float(self.vis_dict[cand]['params']),
'score': float(self.vis_dict[cand]['score'])
}
cand_list.append(info)
if len(cand_list) > 0:
json_dict[str(interval[1])] = cand_list
print("selected candidates:")
print(json_dict)
with open(out_file_name, "w") as fp:
json.dump(json_dict, fp, indent=2)
fp.close()
return self.interval_cands