-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathnas_ea_fa_v2_train_gcn_nasbench101_dnc.py
361 lines (290 loc) · 18 KB
/
nas_ea_fa_v2_train_gcn_nasbench101_dnc.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
from nord.neural_nets import BenchmarkEvaluator
import numpy as np
import os
import copy
import time
from params import EXP_REPEAT_TIMES, MAX_TIME_BUDGET, POPULATION_SIZE, NUM_GEN, K, H, T
from nasbench101_utils_dnc import MAX_CONNECTIONS
from nasbench101_utils_dnc import Architecture
from nasbench101_utils_dnc import randomly_sample_architecture, create_nord_architecture, build_connection_matrix, \
get_all_isomorphic_sequences, get_min_distance, get_model_sequences, tournament_selection, bitwise_mutation
from performance_evaluation import progress_update, save_performance
from save_individual import save_individual_101_dnc, save_individual_fitness_approximation
import torch
import dgl
from dgl_utils import data_preparation, model_configuration, ModelHandler
from dgl_params import *
def NAS_EA_FA_V2_train_gcn_101():
# Instantiate the evaluator
evaluator = BenchmarkEvaluator()
dir_name_results = 'results_nas_ea_fa_v2_dnc101_train_gcn_{}epochs_trainperc{}_K{}_H{}_populationsize{}_unique'.format(NUM_EPOCHS,
int(TRAIN_PERCENTAGE * 100),
K, H, POPULATION_SIZE)
dir_name_dgl_train_stats = 'nas_ea_fa_v2_dnc101_dgl_model_train_stats_{}epochs_trainperc{}_K{}_H{}_populationsize{}_unique'.format(NUM_EPOCHS,
int(TRAIN_PERCENTAGE * 100),
K, H, POPULATION_SIZE)
dir_name_dgl_eval_stats = 'nas_ea_fa_v2_dnc101_dgl_model_eval_stats_{}epochs_trainperc{}_K{}_H{}_populationsize{}_unique'.format(NUM_EPOCHS,
int(TRAIN_PERCENTAGE * 100),
K, H, POPULATION_SIZE)
if not os.path.exists(dir_name_results):
os.mkdir(dir_name_results)
for exp_repeat_index in range(EXP_REPEAT_TIMES):
start_time = time.time()
folder_name = os.path.join(dir_name_results, 'results' + str(exp_repeat_index + 1))
if not os.path.exists(folder_name):
os.mkdir(folder_name)
best_val_acc = []
best_test_acc_based_on_val_acc = []
train_times = []
total_train_time = []
best_test_acc = []
x_train = []
x_train_graphs = []
y_train = []
current_time_budget = 0
# Randomly sample POPULATION_SIZE architectures with an initial fitness of 0
total_population = []
for _ in range(POPULATION_SIZE):
is_valid_architecture = False
while not is_valid_architecture:
architecture = randomly_sample_architecture()
# check if connection number is ok for nasbench-101
if sum(architecture.connections) <= MAX_CONNECTIONS and architecture.valid_architecture:
total_population.append(architecture)
is_valid_architecture = True
num_file = 0
t = 0 # iteration count
# while current_time_budget <= MAX_TIME_BUDGET:
while t < T:
tic = time.time()
t += 1
# sort in descending order by fitness
population = sorted(total_population, key=lambda x: x.fitness, reverse=True)
new_population = []
num_arch = 0
start_index = 0
# train and evaluate top K individuals
for arch_index in range(len(population)):
architecture = population[arch_index]
d = create_nord_architecture(architecture)
# evaluate architecture
val_acc, train_time = evaluator.descriptor_evaluate(d, acc='validation_accuracy')
test_acc, train_time = evaluator.descriptor_evaluate(d, acc='test_accuracy')
architecture.fitness = val_acc
architecture.test_acc = test_acc
architecture.train_time = train_time
# # check whether architecture is already in new population
# architecture_exists = False
# if new_population != []:
# for arch in new_population:
# if architecture.layers == arch.layers and np.array_equal(architecture.connections, arch.connections) and architecture.fitness == arch.fitness:
# architecture_exists = True
# break
# if architecture_exists:
# continue
# convert neural descriptor to networkx
G = d.to_networkx(benchmark_dataset='nasbench101')
# convert networkx to dgl object
architecture.dgl_obj = dgl.from_networkx(G, node_attrs=['layer'])
# add architecture to training set
if architecture.connections.tolist() not in x_train:
x_train.append(architecture.connections.tolist())
x_train_graphs.append(architecture.dgl_obj)
y_train.append(architecture.fitness)
if time == 0.0:
continue
new_population.append(architecture)
# get isomorphic sequences
isomorphic_sequences = get_all_isomorphic_sequences(architecture)
for sequence in isomorphic_sequences:
a = Architecture(architecture.layers, np.array(sequence, dtype=int), build_connection_matrix(sequence),
architecture.fitness)
if a.valid_architecture:
d = create_nord_architecture(a)
# convert neural descriptor to networkx
G = d.to_networkx(benchmark_dataset='nasbench101')
# convert networkx to dgl object
a.dgl_obj = dgl.from_networkx(G, node_attrs=['layer'])
if sequence not in x_train:
x_train.append(sequence)
x_train_graphs.append(a.dgl_obj)
y_train.append(val_acc)
# x_train.extend(isomorphic_sequences)
# for _ in range(len(isomorphic_sequences)):
# y_train.append(val_acc)
best_val_acc, best_test_acc_based_on_val_acc, best_test_acc, train_times, total_train_time = \
progress_update(val_acc=val_acc, test_acc=test_acc, train_time=train_time,
best_val_acc=best_val_acc,
best_test_acc_based_on_val_acc=best_test_acc_based_on_val_acc,
best_test_acc=best_test_acc, train_times=train_times,
total_train_time=total_train_time, fitness='val_acc')
current_time_budget += train_time
num_arch += 1
if current_time_budget > MAX_TIME_BUDGET or num_arch >= K:
start_index = arch_index
break
num_file += 1
with open(os.path.join(folder_name, 'topK_iteration' + str(num_file) + '.txt'), 'w') as f:
ind_num = 0
for ind in new_population:
ind_num += 1
save_individual_101_dnc(f, ind, ind_num, 'val_acc')
num_topK = len(new_population)
# train and evaluate top H individuals
tic1 = time.time()
# get min distance between each of the remaining individuals and the training set
dist_list = [get_min_distance(x_train, get_model_sequences(architecture)) for architecture in
population[start_index + 1:]]
toc1 = time.time()
print('x_train length:', len(x_train))
print('dist_list calculation time:', toc1 - tic1, 'sec')
while num_arch < K + H and current_time_budget <= MAX_TIME_BUDGET:
# find architecture with max distance from training set
max_distance = 0
max_dist_arch_index = start_index
for i in range(len(dist_list)):
if dist_list[i] > max_distance:
max_distance = dist_list[i]
max_dist_arch_index = i
architecture = population[start_index + 1 + max_dist_arch_index]
dist_list[max_dist_arch_index] = 0 # architecture already added to x_train
d = create_nord_architecture(architecture)
# evaluate architecture
val_acc, train_time = evaluator.descriptor_evaluate(d, acc='validation_accuracy')
test_acc, train_time = evaluator.descriptor_evaluate(d, acc='test_accuracy')
architecture.fitness = val_acc
architecture.test_acc = test_acc
architecture.train_time = train_time
# # check whether architecture is already in new population
# architecture_exists = False
# if new_population != []:
# for arch in new_population:
# if architecture.layers == arch.layers and np.array_equal(architecture.connections,
# arch.connections) \
# and architecture.fitness == arch.fitness:
# architecture_exists = True
# break
# if architecture_exists:
# continue
# convert neural descriptor to networkx
G = d.to_networkx(benchmark_dataset='nasbench101')
# convert networkx to dgl object
architecture.dgl_obj = dgl.from_networkx(G, node_attrs=['layer'])
# add architecture to training set
if architecture.connections.tolist() not in x_train:
x_train.append(architecture.connections.tolist())
x_train_graphs.append(architecture.dgl_obj)
y_train.append(architecture.fitness)
if time == 0.0:
continue
new_population.append(architecture)
# get isomorphic sequences
isomorphic_sequences = get_all_isomorphic_sequences(architecture)
for sequence in isomorphic_sequences:
a = Architecture(architecture.layers, np.array(sequence, dtype=int), build_connection_matrix(sequence),
architecture.fitness)
if a.valid_architecture:
d = create_nord_architecture(a)
# convert neural descriptor to networkx
G = d.to_networkx(benchmark_dataset='nasbench101')
# convert networkx to dgl object
a.dgl_obj = dgl.from_networkx(G, node_attrs=['layer'])
if sequence not in x_train:
x_train.append(sequence)
x_train_graphs.append(a.dgl_obj)
y_train.append(val_acc)
# x_train.extend(isomorphic_sequences)
# for _ in range(len(isomorphic_sequences)):
# y_train.append(val_acc)
best_val_acc, best_test_acc_based_on_val_acc, best_test_acc, train_times, total_train_time = \
progress_update(val_acc=val_acc, test_acc=test_acc, train_time=train_time,
best_val_acc=best_val_acc,
best_test_acc_based_on_val_acc=best_test_acc_based_on_val_acc,
best_test_acc=best_test_acc, train_times=train_times,
total_train_time=total_train_time, fitness='val_acc')
current_time_budget += train_time
num_arch += 1
with open(os.path.join(folder_name, 'topH_iteration' + str(num_file) + '.txt'), 'w') as f:
ind_num = num_topK
for index in range(num_topK, len(new_population)):
ind = new_population[index]
ind_num += 1
save_individual_101_dnc(f, ind, ind_num, 'val_acc')
# update population
if len(new_population) != 0:
population = new_population
# # train fitness approximation
# with open(os.path.join(folder_name, 'xgb_stats_iteration' + str(num_file) + '.txt'), 'w') as f:
# with redirect_stdout(f):
# # xgb_model = XGBRegressor(objective='reg:squarederror', learning_rate=0.1)
# xgb_model = XGBRegressor(eta=0.1)
# if t > 1:
# xgb_model.fit(np.array(x_train), np.array(y_train), eval_set=[(x_train, y_train), (x_val, y_val)],
# eval_metric='rmse')
# else:
# xgb_model.fit(np.array(x_train), np.array(y_train), eval_set=[(x_train, y_train)],
# eval_metric='rmse')
# xgb_stats = xgb_model.evals_result()
# print(xgb_stats)
# data setup
dataset, train_data_loader, val_data_loader = data_preparation(train_percentage=TRAIN_PERCENTAGE,
train_batch_size=TRAIN_BATCH_SIZE,
val_batch_size=VAL_BATCH_SIZE,
num_arch=len(x_train_graphs),
graphs=x_train_graphs,
labels=torch.Tensor(y_train))
# dataset.labels = dataset.labels - 0.5
# model setup
model, optimizer, loss_fn = model_configuration(dataset=dataset, num_filters=NUM_FILTERS,
learning_rate=LEARNING_RATE,
dropout_probability=DROPOUT_PROBABILITY)
model_handler = ModelHandler(model=model, optimizer=optimizer, loss_fn=loss_fn)
model_handler.set_loaders(train_loader=train_data_loader, val_loader=val_data_loader)
# model_handler.to('cpu')
model_handler.train(n_epochs=NUM_EPOCHS,
stats_save_dir=os.path.join(dir_name_dgl_train_stats, 'experiment' + str(exp_repeat_index + 1)),
model_name='iteration' + str(t))
model_handler.plot_losses(stats_save_dir=os.path.join(dir_name_dgl_train_stats, 'experiment' + str(exp_repeat_index + 1)),
model_name='iteration' + str(t))
# model_handler.load_checkpoint(os.path.join('dgl_model', 'gcn_150filters_5sageconv_lstm_scaled_relu_trainperc30_100epochs_500nasbench101archs_checkpoint_epoch{}.pth'.format(NUM_EPOCHS)))
model_handler.evaluate(stats_save_dir=os.path.join(dir_name_dgl_eval_stats, 'experiment' + str(exp_repeat_index + 1)),
model_name='iteration' + str(t))
with open(os.path.join(folder_name, 'len_data' + str(num_file) + '.txt'), 'w') as f:
f.write(str(len(x_train_graphs)) + '\n')
# evolutionary algorithm
total_population = []
for epoch in range(NUM_GEN):
new_population = []
for i in range(POPULATION_SIZE):
individual = copy.deepcopy(tournament_selection(population))
new_individual = bitwise_mutation(individual)
d = create_nord_architecture(new_individual)
# convert neural descriptor to networkx
G = d.to_networkx(benchmark_dataset='nasbench101')
# convert networkx to dgl object
new_individual.dgl_obj = dgl.from_networkx(G, node_attrs=['layer'])
# new_individual.fitness = xgb_model.predict(np.array([get_model_sequences(new_individual)]))[0]
new_individual.fitness = model_handler.predict(new_individual.dgl_obj)[0][0] # + 0.5
new_population.append(new_individual)
total_population.append(new_individual)
population = new_population
with open(os.path.join(folder_name, 'population_iteration' + str(num_file) + '_epoch' + str(epoch + 1) +
'.txt'), 'w') as f:
ind_num = 0
for ind in population:
ind_num += 1
save_individual_fitness_approximation(f, ind, ind_num, 'val_acc')
# # validation set for next iteration's xgboost model
# x_val = x_train
# y_val = y_train
toc = time.time()
print('experiment index:', exp_repeat_index+1, 'time needed for iteration t=' + str(t) + ':', toc - tic,
'sec')
print('current time budget:', current_time_budget, 'max time budget:', MAX_TIME_BUDGET)
end_time = time.time()
save_performance(folder_name, exp_repeat_index, start_time, end_time, best_val_acc,
best_test_acc_based_on_val_acc, best_test_acc, train_times, total_train_time,
'val_acc')
if __name__ == '__main__':
np.random.seed(42)
NAS_EA_FA_V2_train_gcn_101()