-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathhelp.py
493 lines (438 loc) · 20.6 KB
/
help.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
####################################################################################################
# HELP: hardware-adaptive efficient latency prediction for nas via meta-learning, NeurIPS 2021
# Hayeon Lee, Sewoong Lee, Song Chong, Sung Ju Hwang
# github: https://github.com/HayeonLee/HELP, email: hayeon926@kaist.ac.kr
####################################################################################################
import os
import logging
from collections import OrderedDict
from collections import defaultdict
import csv
from tqdm import tqdm
import json
#import wandb
import torch
import torch.nn as nn
from torch.utils.tensorboard import SummaryWriter
from net import MetaLearner
from net import Net
from net import InferenceNetwork
from loader import Data
from utils import *
class HELP:
def __init__(self, args):
self.args = args
self.mode = args.mode
self.metrics = args.metrics
self.search_space = args.search_space
self.load_path = args.load_path
# Log
self.save_summary_steps = args.save_summary_steps
self.save_path = args.save_path
# Data & Meta-learning Settings
self.meta_train_devices = args.meta_train_devices
self.meta_valid_devices = args.meta_valid_devices
self.meta_test_devices = args.meta_test_devices
self.num_inner_tasks = args.num_inner_tasks
self.meta_lr = args.meta_lr
self.num_episodes = args.num_episodes
self.num_train_updates = args.num_train_updates
self.num_eval_updates = args.num_eval_updates
self.alpha_on = args.alpha_on
self.inner_lr = args.inner_lr
self.second_order = args.second_order
# Meta-learner
self.hw_emb_dim = args.hw_embed_dim
self.layer_size = args.layer_size
# Inference Network
self.z_on = args.z_on
self.determ = args.determ
self.kl_scaling = args.kl_scaling
self.z_scaling = args.z_scaling
self.mc_sampling = args.mc_sampling
# End to End NAS
if self.mode == 'nas' and not self.search_space in ['nasbench201', 'ofa']:
raise NotImplementedError
self.nas_target_device = args.nas_target_device
self.latency_constraint = args.latency_constraint
# Data
self.data = Data(args.mode,
args.data_path,
args.search_space,
args.meta_train_devices,
args.meta_valid_devices,
args.meta_test_devices,
args.num_inner_tasks,
args.num_meta_train_sample,
args.num_samples,
args.num_query,
args.sampled_arch_path)
# Model
self.model = MetaLearner(args.search_space,
args.hw_embed_on,
args.hw_embed_dim,
args.layer_size).cuda()
self.model_params = list(self.model.parameters())
if self.alpha_on:
self.define_task_lr_params()
self.model_params += list(self.task_lr.values())
else: self.task_lr = None
if self.z_on:
self.inference_network = InferenceNetwork(args.hw_embed_on,
args.hw_embed_dim,
args.layer_size,
args.determ).cuda()
self.model_params += list(self.inference_network.parameters())
self.loss_fn = loss_fn['mse']
if self.mode == 'meta-train':
self.meta_optimizer = torch.optim.Adam(self.model_params, lr=self.meta_lr)
self.scheduler = None
# Set the logger
set_logger(os.path.join(self.save_path, 'log.txt'))
if args.use_wandb:
wandb.init(entity="hayeonlee",
project=args.project,
name=args.exp_name,
group=args.group,
reinit=True)
wandb.config.update(args)
writer = None
else:
writer = SummaryWriter(log_dir=self.save_path)
self.log = {
'meta_train': Log(self.save_path,
self.save_summary_steps,
self.metrics,
self.meta_train_devices,
'meta_train',
writer, args.use_wandb),
'meta_valid': Log(self.save_path,
self.save_summary_steps,
self.metrics,
self.meta_valid_devices,
'meta_valid',
writer,
args.use_wandb),
}
def define_task_lr_params(self):
self.task_lr = OrderedDict()
for key, val in self.model.named_parameters():
self.task_lr[key] = nn.Parameter(
1e-3 * torch.ones_like(val))
def get_params_z(self, xs, ys, hw_embed):
params = self.model.cloned_params()
z, kl = self.inference_network((xs, ys, hw_embed))
zs = self.z_scaling
for i, (name, weight) in enumerate(params.items()):
if 'weight' in name:
if 'fc3' in name:
idx = 0
elif 'fc4' in name:
idx = 1
elif 'fc5' in name:
idx = 2
else:
continue
layer_size = 2*self.layer_size
params[name] = weight * (1 + zs*z['w'][idx*layer_size:(idx+1)*layer_size])
elif 'bias' in name:
if 'fc3' in name:
idx = 0
elif 'fc4' in name:
idx = 1
elif 'fc5' in name:
idx = 2
else:
continue
params[name] = weight + zs*z['b'][idx]
else: raise ValueError(name)
return params, kl, z
def train_single_task(self, hw_embed, xs, ys, num_updates):
self.model.train()
if self.search_space in ['fbnet', 'ofa']:
xs, ys = xs.cuda(), ys.cuda()
elif self.search_space == 'nasbench201':
xs, ys = (xs[0].cuda(), xs[1].cuda()), ys.cuda()
hw_embed = hw_embed.cuda()
if self.z_on:
params, kl, z = self.get_params_z(xs, ys, hw_embed)
else:
params = self.model.cloned_params()
kl = 0.0
adapted_params = params
for n in range(num_updates):
ys_hat = self.model(xs, hw_embed, adapted_params)
loss = self.loss_fn(ys_hat, ys)
grads = torch.autograd.grad(
loss, adapted_params.values(), create_graph=(self.second_order))
for (key, val), grad in zip(adapted_params.items(), grads):
if self.task_lr is not None: # Meta-SGD
task_lr = self.task_lr[key]
else:
task_lr = self.inner_lr # MAML
adapted_params[key] = val - task_lr * grad
return adapted_params, kl
def meta_train(self):
print('==> start training...')
max_valid_corr = -1
if self.z_on:
self.inference_network.train()
with tqdm(total=self.num_episodes) as t:
for i_epi in range(self.num_episodes):
# Run inner loops to get adapted parameters (theta_t`)
adapted_state_dicts = []
query_list = []
episode = self.data.generate_episode()
for i_task in range(self.num_inner_tasks):
# Perform a gradient descent to meta-learner on the task
(hw_embed, xs, ys, xq, yq, _) = episode[i_task]
adapted_state_dict, kl_loss = \
self.train_single_task(hw_embed, xs, ys, self.num_train_updates)
# Store adapted parameters
# Store dataloaders for meta-update and evaluation
adapted_state_dicts.append(adapted_state_dict)
query_list.append((hw_embed, xq, yq))
# Update the parameters of meta-learner
# Compute losses with adapted parameters along with corresponding tasks
# Updated the parameters of meta-learner using sum of the losses
meta_loss = 0
for i_task in range(self.num_inner_tasks):
hw_embed, xq, yq = query_list[i_task]
if self.search_space in ['fbnet', 'ofa']:
xq, yq = xq.cuda(), yq.cuda()
elif self.search_space == 'nasbench201':
xq, yq = (xq[0].cuda(), xq[1].cuda()), yq.cuda()
hw_embed = hw_embed.cuda()
adapted_state_dict = adapted_state_dicts[i_task]
yq_hat = self.model(xq, hw_embed, adapted_state_dict)
loss_t = self.loss_fn(yq_hat, yq)
meta_loss += loss_t / float(self.num_inner_tasks) \
+ self.kl_scaling * kl_loss
self.meta_optimizer.zero_grad()
meta_loss.backward()
self.meta_optimizer.step()
if self.scheduler is not None:
self.scheduler.step(meta_loss)
# Evaluate model on new tasks
# Evaluate on train and test dataset given a number of tasks (args.num_steps)
if (i_epi + 1) % self.save_summary_steps == 0:
logging.info(f"Episode {i_epi+1}/{self.num_episodes}")
postfix = {}
for split in ['meta_train', 'meta_valid']:
msg = f"[{split.upper()}] "
self._test_predictor(split, i_epi)
self.log[split].update_epi(i_epi)
for m in self.metrics + ['mse_loss', 'kl_loss']:
v = self.log[split].avg(i_epi, m)
postfix[f'{split}/{m}'] = f'{v:05.3f}'
msg += f"{m}: {v:05.3f}; "
if m == 'spearman' and max_valid_corr < v:
max_valid_corr = v
save_dict = {'epi': i_epi,
'model': self.model.cpu().state_dict()}
if self.args.z_on:
save_dict['inference_network'] = self.inference_network.cpu().state_dict()
self.inference_network.cuda()
if self.args.alpha_on:
save_dict['task_lr'] = {k: v.cpu() for k, v in self.task_lr.items()}
for k, v in self.task_lr.items():
self.task_lr[k].cuda()
save_path = os.path.join(self.save_path, 'checkpoint', f'help_max_corr.pt')
torch.save(save_dict, save_path)
print(f'==> save {save_path}')
self.model.cuda()
logging.info(msg)
t.set_postfix(postfix)
print('\n')
t.update()
self.log['meta_train'].save()
self.log['meta_valid'].save()
print('==> Training done')
def test_predictor(self):
loaded = torch.load(self.load_path)
print(f'==> load {self.load_path}')
if 'epi' in loaded.keys():
epi = loaded['epi']
print(f'==> load {epi} model..')
self.model.load_state_dict(loaded['model'])
if self.z_on:
self.inference_network.load_state_dict(loaded['inference_network'])
if self.alpha_on:
for (k, v), (lk, lv) in zip(self.task_lr.items(), loaded['task_lr'].items()):
self.task_lr[k] = lv.cuda()
self._test_predictor('meta_test', None)
def _test_predictor(self, split, i_epi=None):
save_file_path = os.path.join(self.save_path, f'test_log.txt')
f = open(save_file_path, 'a+')
if self.z_on:
self.inference_network.eval()
avg_metrics = {m: 0.0 for m in self.metrics}
avg_metrics['mse_loss'] = 0.0
tasks = self.data.generate_test_tasks(split)
for (hw_embed, xs, ys, xq, yq, device) in tasks:
yq_hat_mean = None
for _ in range(self.mc_sampling):
adapted_state_dict, kl_loss = \
self.train_single_task(hw_embed, xs, ys, self.num_eval_updates)
if self.search_space in ['fbnet', 'ofa']:
xq, yq = xq.cuda(), yq.cuda()
elif self.search_space == 'nasbench201':
xq, yq = (xq[0].cuda(), xq[1].cuda()), yq.cuda()
hw_embed = hw_embed.cuda()
yq_hat = self.model(xq, hw_embed, adapted_state_dict)
if yq_hat_mean is None:
yq_hat_mean = yq_hat
else:
yq_hat_mean += yq_hat
yq_hat_mean = yq_hat_mean / self.args.mc_sampling
loss = self.loss_fn(yq_hat_mean, yq)
if i_epi is not None:
for metric in self.metrics:
self.log[split].update(i_epi, metric, device,
val=metrics_fn[metric](yq_hat, yq)[0])
self.log[split].update(i_epi, 'mse_loss', device, val=loss.item())
self.log[split].update(i_epi, 'kl_loss', device, val=kl_loss if isinstance(kl_loss, float) else kl_loss.item())
else:
msg = f'[{split}/{device}] '
for m in self.metrics:
msg += f'{m} {metrics_fn[m](yq_hat, yq)[0]:.3f} '
avg_metrics[m] += metrics_fn[m](yq_hat, yq)[0]
msg += f'MSE {loss.item():.3f}'
avg_metrics['mse_loss'] += loss.item()
f.write(msg+'\n')
print(msg)
if i_epi is None:
nd = len(tasks)
msg = f'[{split}/average] '
for m in self.metrics:
msg += f'{m} {avg_metrics[m]/nd:.3f} '
mse_loss = avg_metrics['mse_loss']
msg += f'MSE {mse_loss/nd:.3f} ({nd} devices)'
f.write(msg+'\n')
print(msg)
f.close()
def _denormalization(self, task, yq_hat, adapted_state_dict):
hw_embed, xs, ys, xq, yq, device, ys_gt, yq_gt = task
xs = (xs[0].cuda(), xs[1].cuda())
ys_gt, yq_gt = ys_gt.cuda(), yq_gt.cuda()
ys_hat = self.model(xs, hw_embed.cuda(), adapted_state_dict)
ysh_min = min(ys_hat)
ysh_max = max(ys_hat)
denorm_yq_hat = denorm((yq_hat-ysh_min)/(ysh_max-ysh_min), max(ys_gt), min(ys_gt))
denorm_mse = self.loss_fn(denorm_yq_hat.cuda(), yq_gt)
return denorm_yq_hat, denorm_mse
def load_model(self):
loaded = torch.load(os.path.join(self.load_path))
self.model.load_state_dict(loaded['model'])
self.model.eval()
self.model.cuda()
if self.alpha_on:
self.task_lr = {k: v.cuda() for k, v in loaded['task_lr'].items()}
if self.z_on:
self.inference_network.load_state_dict(loaded['inference_network'])
self.inference_network.eval()
self.inference_network.cuda()
def nas(self):
if self.search_space == 'ofa':
self._nas_ofa()
elif self.search_space == 'nasbench201':
self._nas_metad2a()
def _nas_metad2a(self):
save_file_path = os.path.join(self.save_path, f'nas_results_{self.nas_target_device}.txt')
f = open(save_file_path, 'a+')
self.load_model()
search_results = {}
task = self.data.get_nas_task(self.nas_target_device)
hw_embed, xs, ys, xq, yq, device, ys_gt, yq_gt = task
yq_hat_mean = None
for _ in range(self.mc_sampling):
adapted_state_dict, kl_loss = \
self.train_single_task(hw_embed, xs, ys, self.num_eval_updates)
xq, yq = (xq[0].cuda(), xq[1].cuda()), yq.cuda()
hw_embed = hw_embed.cuda()
yq_hat = self.model(xq, hw_embed, adapted_state_dict)
if yq_hat_mean is None:
yq_hat_mean = yq_hat
else:
yq_hat_mean += yq_hat
yq_hat_mean = yq_hat_mean / self.args.mc_sampling
loss = self.loss_fn(yq_hat_mean, yq)
# Denormalization
denorm_yq_hat, denorm_mse = self._denormalization(task, yq_hat_mean, adapted_state_dict)
search_results = []
top = 3
true_acc = self.data.arch_candidates['true_acc']
arch_str = self.data.arch_candidates['arch']
const = float(self.latency_constraint)
for dyq_hat, yq_, acc_, arch_ in \
zip(denorm_yq_hat, yq_gt, true_acc, arch_str):
if dyq_hat.item() <= const:
if len(search_results) < top:
search_results.append({
'yq': yq_,
'acc': acc_,
'arch_str': arch_
})
if len(search_results) >= top:
break
max_acc_result = search_results[0]
for result in search_results:
if result['acc'] > max_acc_result['acc']:
max_acc_result = result
lat = max_acc_result['yq'].item()
acc = float(max_acc_result['acc'])
arch = max_acc_result['arch_str']
msg = f'[NAS Result] Target Device {self.nas_target_device} Constraint {const} '
msg += f'| Latency {lat:.1f} | Accuracy {acc:.1f} | Neural Architecture {arch}'
print(msg)
f.write(msg+'\n')
f.close()
def _nas_ofa(self):
from ofa.tutorial.accuracy_predictor import AccuracyPredictor
from ofa.finder import EvolutionFinder
# load HELP
self.load_model()
task = self.data.get_nas_task(self.nas_target_device)
#hw_embed, xs, ys, ys_gt = task
#import pdb; pdb.set_trace()
hw_embed, xs, ys, ys_gt = [_.cuda() for _ in task]
ys_hat_mean = None
for _ in range(self.mc_sampling):
adapted_state_dict, kl_loss = \
self.train_single_task(hw_embed, xs, ys, self.num_eval_updates)
ys_hat = self.model(xs, hw_embed, adapted_state_dict)
if ys_hat_mean is None:
ys_hat_mean = ys_hat
else:
ys_hat_mean += ys_hat
ys_hat = ys_hat_mean / self.args.mc_sampling
latency_constraint = data_norm(self.latency_constraint, ys_gt, ys_hat).item()
# load accuracy predictor of once-for-all
acc_predictor = AccuracyPredictor(pretrained=True)
params = {
'constraint_type': self.nas_target_device,
'efficiency_constraint': latency_constraint,
'hardware_embedding': hw_embed,
'adapted_state_dict': adapted_state_dict,
'mutate_prob': 0.1, # The probability of mutation in evolutionary search
'mutation_ratio': 0.5, # The ratio of networks that are generated through mutation in generation n >= 2.
'efficiency_predictor': self.model , # To use a predefined efficiency predictor.
'accuracy_predictor': acc_predictor, # To use a predefined accuracy_predictor predictor.
'ys_gt' : ys_gt,
'ys_hat': ys_hat,
'population_size': 100,
'max_time_budget': 500,
'parent_ratio': 0.25,
}
finder = EvolutionFinder(**params)
best_valids, best_info, top_k = finder.run_evolution_search()
pred_acc = best_info[0]
arch_config = best_info[1]
pred_lat = data_norm(best_info[2], ys_hat, ys_gt).item()
msg = f'[NAS Result] Target Device {self.nas_target_device} '
msg += f'Constraint {self.latency_constraint} '
msg += f'Neural Architecture Config {arch_config}'
print(msg)
save_file_path = os.path.join(self.save_path, f'nas_results_{self.nas_target_device}.json')
print(f'save path is {save_file_path}')
json.dump(arch_config, open(save_file_path, 'w'), indent=4)