-
Notifications
You must be signed in to change notification settings - Fork 7
/
unas.py
307 lines (255 loc) · 14.1 KB
/
unas.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
# ---------------------------------------------------------------
# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
#
# This work is licensed under the NVIDIA Source Code License
# for UNAS. To view a copy of this license, see the LICENSE file.
# ---------------------------------------------------------------
import time
import os
import numpy as np
import torch
import torch.nn as nn
from apex.fp16_utils import to_python_float
from time import time
from util import utils
from surrogate import SurrogateLinear
class UNAS(object):
def __init__(self, model, alpha, args, writer, logging):
self.args = args
self.model = model
self.alpha = alpha
self.logging = logging
self.arch_optimizer = torch.optim.Adam(self.alpha.parameters(), lr=args.arch_learning_rate,
weight_decay=args.arch_weight_decay)
self.arch_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
self.arch_optimizer, float(args.epochs), eta_min=args.arch_learning_rate_min)
self.latency_cost = args.target_latency > 0.
self.target_latency = args.target_latency
if self.latency_cost or self.args.meta_loss == 'relax':
assert args.meta_loss in {'relax', 'rebar', 'reinforce'}, 'this is only implemented for rebar and reinforce'
normal_size, reduce_size = self.alpha.module.alphas_size()
alpha_size = normal_size + reduce_size
self.surrogate = SurrogateLinear(alpha_size, self.logging).cuda()
self.latency_pred_loss = utils.AverageMeter()
self.latency_value = utils.AverageMeter()
self.latency_coeff = args.latency_coeff
self.latency_coeff_curr = None
self.num_repeat = 10
self.latency_batch_size = 24
assert self.latency_batch_size <= args.batch_size
self.num_arch_samples = 10000
# print('***************** change the number of samples *******')
# self.num_arch_samples = 200
self.surrogate_not_train = True
self.latency_actual = []
self.latency_estimate = []
# Extra layers, if any.
self.meta_loss = args.meta_loss
# weights generalization error
self.gen_error_alpha = args.gen_error_alpha
self.gen_error_alpha_lambda = args.gen_error_alpha_lambda
# Get the meta learning criterion.
if self.meta_loss in ['default', 'rebar', 'reinforce']:
self.criterion = nn.CrossEntropyLoss(reduction='none')
self.criterion = self.criterion.cuda()
if self.meta_loss == 'reinforce':
self.exp_avg1 = utils.ExpMovingAvgrageMeter()
self.exp_avg2 = utils.ExpMovingAvgrageMeter()
self.alpha_loss = args.alpha_loss
# Housekeeping.
self.loss = None
self.accuracy = None
self.count = None
self.loss_diff_sign = None
self.reset_counter()
self.report_freq = args.report_freq
self.writer = writer
def reset_counter(self):
"""Resets counters."""
self.count = 0
self.loss = utils.AverageMeter()
self.accuracy = utils.AverageMeter()
self.loss_diff_sign = utils.AverageMeter()
if self.latency_cost:
self.latency_pred_loss = utils.AverageMeter()
self.latency_value = utils.AverageMeter()
if self.meta_loss == 'relax':
self.relax_pred_loss = utils.AverageMeter()
def mean_accuracy(self):
"""Return mean accuracy."""
return self.accuracy.avg
def compute_latency(self, train_batch, discrete_weight):
discrete_indices = self.model.module.get_indices(discrete_weight)
self.model.eval()
train_batch = train_batch[:self.latency_batch_size]
elapsed_time = np.zeros(self.num_repeat)
with torch.no_grad():
for i in range(self.num_repeat):
torch.cuda.synchronize()
start = time()
tmp = self.model.module.fast_forward(train_batch, discrete_indices)
torch.cuda.synchronize()
end = time()
elapsed_time[i] = (end - start)
self.model.train()
return np.median(elapsed_time) * 1000
def train_surrogate(self, train_batch):
# measure
self.logging.info('collecting latency samples')
past_alphas = []
past_target = []
for i in range(self.num_arch_samples):
with torch.no_grad():
weights = self.alpha(1)
disc_weights = self.alpha.module.discretize(weights)
latency = self.compute_latency(train_batch, disc_weights)
alphas = self.alpha.module.get_arch_sample(disc_weights)
past_alphas.append(alphas.cpu())
past_target.append(torch.FloatTensor([latency]))
if i % 100 == 0:
self.logging.info('collected %d samples' % i)
all_alphas = torch.cat(past_alphas, dim=0)
all_target = torch.cat(past_target, dim=0)
num_train = int(0.8 * self.num_arch_samples)
test_alphas = all_alphas[num_train:]
test_target = all_target[num_train:]
train_alphas = all_alphas[:num_train]
train_target = all_target[:num_train]
self.surrogate.learn(train_alphas, train_target, test_alphas, test_target)
self.surrogate.eval()
self.surrogate_not_train = False
def training_obj(self, train, train_target, weights, model_opt, val, val_target, global_step):
if not self.gen_error_alpha:
logits = self.model(train, weights)
loss = self.criterion(logits, train_target)
accuracy = utils.accuracy(logits, train_target)[0]
loss1, loss2 = loss, torch.zeros_like(loss)
else:
logits_train = self.model(train, weights)
loss_train = self.criterion(logits_train, train_target)
logits_val = self.model(val, weights)
loss_val = self.criterion(logits_val, val_target)
loss2 = torch.abs(loss_val - loss_train)
self.loss_diff_sign.update(torch.mean(((loss_val - loss_train) > 0).float()).data)
loss1 = loss_train
loss = loss1 + self.gen_error_alpha_lambda * loss2
accuracy = utils.accuracy(logits_train, train_target)[0]
if self.alpha_loss:
alpha_loss = self.alpha.module.alpha_loss(weights)
loss += self.args.alpha_loss_lambda * alpha_loss
if self.count % self.report_freq == 0:
self.writer.add_scalar(
'meta/alpha_loss', torch.mean(alpha_loss), global_step)
return loss, accuracy, loss1, loss2
def step(self, input_valid, target_valid, global_step, weights, input_valid2=None, target_valid2=None, model_opt=None):
"""Optimizer for the architecture params."""
self.arch_optimizer.zero_grad()
if self.meta_loss == 'default':
loss, accuracy, loss1, loss2 = self.training_obj(input_valid, target_valid, weights, model_opt,
input_valid2, target_valid2, global_step)
loss, loss1, loss2 = torch.mean(loss), torch.mean(loss1), torch.mean(loss2)
elif self.meta_loss == 'rebar':
# compute loss with discrete weights
with torch.no_grad():
disc_weights = {
'normal': weights['dis_normal'], 'reduce': weights['dis_reduce']}
loss_disc, accuracy, loss1, loss2 = self.training_obj(input_valid, target_valid, disc_weights,
model_opt, input_valid2, target_valid2, global_step)
# compute baseline
loss_cont, _, _, _ = self.training_obj(input_valid, target_valid, weights,
model_opt, input_valid2, target_valid2, global_step)
reward = (loss_disc - loss_cont).detach()
log_q_d = self.alpha.module.log_prob(weights)
loss = torch.mean(log_q_d * reward) + torch.mean(loss_cont)
loss1, loss2 = torch.mean(loss1), torch.mean(loss2)
if self.latency_cost:
# train the surrogate function initially.
if self.surrogate_not_train:
self.train_surrogate(input_valid)
# sample a single architecture sample
weight_lat = self.alpha(1)
disc_weights_lat = {'normal': weight_lat['dis_normal'], 'reduce': weight_lat['dis_reduce']}
# compute latency for the discrete weights.
elapsed_time = self.compute_latency(input_valid, disc_weights_lat)
# latency prediction for continuous weights
self.surrogate.eval()
alphas = self.alpha.module.get_arch_sample(weight_lat)
latency_cont = self.surrogate(alphas)
# latency prediction for discrete weights
alphas = self.alpha.module.get_arch_sample(disc_weights_lat)
latency_discrete = self.surrogate(alphas)
surrogate_loss = torch.mean(torch.abs(elapsed_time - latency_discrete.squeeze(1)))
self.latency_coeff_curr = self.latency_coeff * max(min(global_step / self.args.latency_iter, 1.0), 0.)
loss_disc_lat = self.latency_coeff_curr * torch.relu(torch.Tensor([elapsed_time]).cuda() - self.target_latency)
loss_cont_lat = self.latency_coeff_curr * torch.relu(latency_cont[0] - self.target_latency)
# collect latency information
self.latency_pred_loss.update(utils.reduce_tensor(surrogate_loss.data, self.args.world_size))
self.latency_value.update(elapsed_time)
self.latency_actual.append(elapsed_time)
self.latency_estimate.append(latency_discrete.squeeze(1).data.cpu().numpy()[0])
if global_step % 50 == 0:
self.logging.info('latency_pred_loss %f' % np.mean(np.abs(np.array(self.latency_actual)[-50:] - np.array(self.latency_estimate)[-50:])))
# saving some latency info
if global_step % 1000 == 100 and self.args.local_rank == 0:
import pickle
print('saving')
with open(os.path.join(self.args.save, 'latency.pkl'), 'wb') as f:
pickle.dump([self.latency_actual, self.latency_estimate, global_step], f)
reward = (loss_disc_lat - loss_cont_lat).detach()
log_q_d = self.alpha.module.log_prob(weight_lat)
loss = loss + torch.mean(log_q_d * reward) + torch.mean(loss_cont_lat)
elif self.meta_loss == 'reinforce':
# compute loss with discrete weights
with torch.no_grad():
disc_weights = self.alpha.module.discretize(weights)
loss_disc, accuracy, loss1, loss2 = self.training_obj(input_valid, target_valid, disc_weights,
model_opt, input_valid2, target_valid2,
global_step)
reduce_loss_disc = utils.reduce_tensor(loss_disc.data, self.args.world_size)
avg = torch.mean(reduce_loss_disc).detach()
baseline = self.exp_avg1.avg
# update the moving average
self.exp_avg1.update(avg)
reward = (loss_disc - baseline).detach()
log_q_d = self.alpha.module.log_prob(weights)
loss = torch.mean(log_q_d * reward) + baseline
loss1, loss2 = torch.mean(loss1), torch.mean(loss2)
if self.latency_cost:
weight_lat = self.alpha(1)
disc_weights_lat = self.alpha.module.discretize(weights)
elapsed_time = self.compute_latency(input_valid, disc_weights_lat)
self.latency_coeff_curr = self.latency_coeff * min(global_step / self.args.latency_iter, 1.0)
loss_disc_lat = self.latency_coeff_curr * elapsed_time
self.latency_value.update(elapsed_time)
baseline = self.exp_avg2.avg
# update the moving average
self.exp_avg2.update(float(loss_disc_lat))
reward = loss_disc_lat - baseline
log_q_d = self.alpha.module.log_prob(weight_lat)
loss = loss + torch.mean(log_q_d * reward) + baseline
loss1, loss2 = torch.mean(loss1), torch.mean(loss2)
entropy_loss = self.alpha.module.entropy_loss(weights)
# Backward pass and update.
loss.backward()
self.arch_optimizer.step()
# Logging.
reduced_loss = utils.reduce_tensor(loss.data, self.args.world_size)
accuracy = utils.reduce_tensor(accuracy, self.args.world_size)
self.loss.update(to_python_float(reduced_loss), 1)
self.accuracy.update(to_python_float(accuracy), 1)
self.count += 1
if self.count % self.report_freq == 0:
self.logging.info('Meta Loss:%s %03d %e %f', self.meta_loss,
self.count, self.loss.avg, self.accuracy.avg)
self.writer.add_scalar('meta/loss', self.loss.avg, global_step)
self.writer.add_scalar('meta/acc', self.accuracy.avg, global_step)
self.writer.add_scalar('meta/lr', self.arch_optimizer.state_dict()['param_groups'][0]['lr'], global_step)
self.writer.add_scalar('meta/entropy', entropy_loss, global_step)
if self.gen_error_alpha:
self.writer.add_scalar('meta/loss_val', loss1, global_step)
self.writer.add_scalar('meta/loss_cov', loss2, global_step)
self.writer.add_scalar('meta/loss_diff_sign', self.loss_diff_sign.avg, global_step)
if self.latency_cost:
self.writer.add_scalar('meta/latency_time', self.latency_value.avg, global_step)
self.writer.add_scalar('meta/latency_prediction_loss', self.latency_pred_loss.avg, global_step)
self.writer.add_scalar('meta/latency_coeff', self.latency_coeff_curr, global_step)