-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathsearcher.py
185 lines (146 loc) · 6.25 KB
/
searcher.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
import copy
import random
import logging
import hashlib
import warnings
import argparse
from tqdm import tqdm
from flops import TransformerHparams
warnings.filterwarnings("ignore")
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO
)
logger = logging.getLogger(__name__)
class Genome(object):
def __init__(self, gene_param=None):
self.fitness = 0.0
self.gene_param = gene_param
if not self.gene_param:
self.hash = 0
else:
self.update_hash()
def update_hash(self):
gene_string = str(self.gene_param["intermediate_size"])+ \
str(self.gene_param["vocab_size"]) + \
str(self.gene_param["attention_heads"]) + \
str(self.gene_param["hidden_dim"]) + \
str(self.gene_param["n_layers"])
self.hash = hashlib.md5(gene_string.encode("UTF-8")).hexdigest()
def mutation(self, search_space):
genome_len = len(self.gene_param)
loc = random.randint(0, genome_len - 1)
for x in range(genome_len):
if x >= loc:
mutated_gene = list(self.gene_param.keys())[x]
current_value = self.gene_param[mutated_gene]
possible_choices = copy.deepcopy(search_space[mutated_gene])
possible_choices.remove(current_value)
self.gene_param[mutated_gene] = random.choice(possible_choices)
self.update_hash()
class GA_search():
def __init__(self, args, search_space, cross_chance=0.6):
self.args = args
self.search_space = search_space
self.cross_chance = cross_chance
self.desired_length = args.population_size
self.population = []
self.best_gene = []
def is_duplicate(self, new_genome):
for genome in self.population:
if new_genome.hash == genome.hash:
return True
return False
def initialization(self):
count = 0
while count < self.args.population_size:
gene_param = {}
for key in self.search_space:
gene_param[key] = random.choice(self.search_space[key])
new_genome = Genome(gene_param)
if len(self.population) > 0:
while self.is_duplicate(new_genome):
new_genome.mutation(self.search_space)
self.population.append(copy.deepcopy(new_genome))
count += 1
def fitness(self, genome):
vocab_size = genome.gene_param["vocab_size"]
attention_heads = genome.gene_param["attention_heads"]
hidden_dim = genome.gene_param["hidden_dim"]
intermediate_size = genome.gene_param["intermediate_size"]
n_layers = genome.gene_param["n_layers"]
model = TransformerHparams(hidden_dim, n_layers, 514, vocab_size, intermediate_size, attention_heads)
flops = model.get_infer_flops()
params = model.get_params()
size_diff = abs(self.args.target_size - params)*4/1e6
genome.fitness = flops/1e9 - size_diff
def crossover_and_mutation(self, parents):
children = []
parent_1, parent_2 = parents
if self.cross_chance > random.random():
genome_len = len(self.search_space)
recomb_loc = random.randint(1, genome_len - 1)
child_1 = {}
child_2 = {}
keys = list(self.search_space)
keys = sorted(keys)
for x in range(0, genome_len):
if x < recomb_loc:
child_1[keys[x]] = parent_1.gene_param[keys[x]]
child_2[keys[x]] = parent_2.gene_param[keys[x]]
else:
child_1[keys[x]] = parent_2.gene_param[keys[x]]
child_2[keys[x]] = parent_1.gene_param[keys[x]]
genome_1 = Genome(child_1)
genome_2 = Genome(child_2)
else:
genome_1 = copy.deepcopy(parent_1)
genome_2 = copy.deepcopy(parent_2)
genome_1.mutation(self.search_space)
genome_2.mutation(self.search_space)
children.append(genome_1)
children.append(genome_2)
return children
def generation(self):
children = []
while len(children) < self.desired_length:
parents = random.sample(self.population, k=2)
children.extend(self.crossover_and_mutation(parents))
for genome in children:
while self.is_duplicate(genome):
genome.mutation(self.search_space)
self.population.append(copy.deepcopy(genome))
for genome in self.population:
self.fitness(genome)
graded_genome = [x for x in sorted(self.population, key=lambda x: x.fitness, reverse=True)]
self.best_gene.append((graded_genome[0].gene_param, graded_genome[0].fitness))
self.population = graded_genome[:self.args.population_size]
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--population_size", default=50, type=int)
parser.add_argument("--generation_size", default=100, type=int)
parser.add_argument("-t", "--target_size", default=3, type=float)
args = parser.parse_args()
search_space = {
"vocab_size": [*range(1000, 51000, 1000)],
"attention_heads": [1, 2, 4, 8],
"hidden_dim": [*range(16, 769, 16)],
"intermediate_size": [*range(32, 3072, 32)],
"n_layers": [*range(1, 13)]
}
args.target_size = args.target_size * 1e6/4
logger.info("***Start GA search for %d generations, %d population, target model size %d MB***" %
(args.generation_size, args.population_size, args.target_size*4/1e6))
searcher = GA_search(args, search_space)
searcher.initialization()
for gen in tqdm(range(args.generation_size), desc="Searching"):
# logger.info("***Start generate %d***" %(gen))
searcher.generation()
for genome in searcher.population:
searcher.fitness(genome)
graded_genome = [x for x in sorted(searcher.population, key=lambda x: x.fitness, reverse=True)]
logger.info("the best one:")
logger.info(graded_genome[0].gene_param)
if __name__ == "__main__":
main()