/
loader.py
254 lines (216 loc) · 11.5 KB
/
loader.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
####################################################################################################
# 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 random
import numpy as np
import torch
from utils import *
class Data:
def __init__(self, mode,
data_path,
search_space,
meta_train_devices,
meta_valid_devices,
meta_test_devices,
num_inner_tasks,
num_meta_train_sample,
num_sample,
num_query,
sampled_arch_path,
num_query_meta_train_task=200,
remove_outlier=True):
self.mode = mode
self.data_path = data_path
self.search_space = search_space
self.meta_train_devices = meta_train_devices
self.meta_valid_devices = meta_valid_devices
self.meta_test_devices = meta_test_devices
self.num_inner_tasks = num_inner_tasks
self.num_meta_train_sample = num_meta_train_sample
self.num_sample = num_sample
self.num_query = num_query
self.num_query_meta_train_task = num_query_meta_train_task
self.remove_outlier = remove_outlier
# Latency-constrainted NAS
self.sampled_arch_path = sampled_arch_path
self.load_archs()
self.train_idx ={}
self.valid_idx = {}
self.latency = {}
self.norm_latency = {}
nts = self.num_meta_train_sample
for device in meta_train_devices + meta_valid_devices + meta_test_devices:
self.latency[device] = torch.FloatTensor(
torch.load(os.path.join(data_path, 'latency', f'{device}.pt')))
train_idx = torch.arange(len(self.archs))[:nts]
valid_idx = torch.arange(len(self.archs))[nts:nts+self.num_query]
if self.remove_outlier:
self.train_idx[device] = train_idx[
np.argsort(self.latency[device][train_idx])[
int(len(train_idx)*0.1):int(len(train_idx)*0.9)]]
self.valid_idx[device] = valid_idx[
np.argsort(self.latency[device][valid_idx])[
int(len(valid_idx)*0.1):int(len(valid_idx)*0.9)]]
self.norm_latency[device] = normalization(
latency=self.latency[device],
index = self.train_idx[device]
)
# load index set of reference architectures
self.hw_emb_idx = torch.load(
os.path.join(data_path, 'hardware_embedding_index.pt'))
if self.mode == 'nas':
self.max_lat_idx, self.min_lat_idx = get_minmax_latency_index(
meta_train_devices + meta_valid_devices, self.train_idx, self.latency)
self.nas_norm_latency = {}
for device in meta_train_devices + meta_valid_devices + meta_test_devices:
self.nas_norm_latency[device] = normalization(
latency=self.latency[device],
index = torch.tensor([self.max_lat_idx, self.min_lat_idx] + self.hw_emb_idx))
if self.search_space == 'nasbench201':
self._load_arch_candidates()
self._load_arch_str2idx()
print('==> load data ...')
def load_archs(self):
if self.search_space == 'nasbench201':
# operations and adjacency matrix of neural architectures in NAS-Bench-201
self.archs = [[add_global_node(_['operation'], ifAdj=False),
add_global_node(_['adjacency_matrix'],ifAdj=True)]
for _ in torch.load(os.path.join(self.data_path, 'architecture.pt'))]
elif self.search_space == 'fbnet':
self.archs = [arch_enc(_['op_idx_list']) for _ in
torch.load(os.path.join(self.data_path, 'metainfo.pt'))['arch']]
elif self.search_space == 'ofa':
self.archs = [arch_encoding_ofa(arch) for arch in
torch.load(os.path.join(self.data_path, 'ofa_archs.pt'))['arch']]
def generate_episode(self):
# metabatch
episode = []
# meta-batch
rand_device_idx = torch.randperm(
len(self.meta_train_devices))[:self.num_inner_tasks]
for t in rand_device_idx:
# sample devices
device = self.meta_train_devices[t]
# hardware embedding
latency = self.latency[device]
hw_embed = latency[self.hw_emb_idx]
hw_embed = normalization(hw_embed, portion=1.0)
# samples for finetuning & test (query)
rand_idx = self.train_idx[device][torch.randperm(len(self.train_idx[device]))]
finetune_idx = rand_idx[:self.num_sample]
qry_idx = rand_idx[self.num_sample:self.num_sample+self.num_query_meta_train_task]
if self.search_space in ['fbnet', 'ofa']:
x_finetune = torch.stack([self.archs[_] for _ in finetune_idx])
x_qry = torch.stack([self.archs[_] for _ in qry_idx])
elif self.search_space == 'nasbench201':
x_finetune = [torch.stack([self.archs[_][0] for _ in finetune_idx]),
torch.stack([self.archs[_][1] for _ in finetune_idx])]
x_qry = [torch.stack([self.archs[_][0] for _ in qry_idx]),
torch.stack([self.archs[_][1] for _ in qry_idx])]
y_finetune = self.norm_latency[device][finetune_idx].view(-1, 1)
y_qry = self.norm_latency[device][qry_idx].view(-1, 1)
episode.append((hw_embed, x_finetune, y_finetune, x_qry, y_qry, device))
return episode
def generate_test_tasks(self, split=None):
if split == 'meta_train':
device_list = self.meta_train_devices
elif split == 'meta_valid':
device_list = self.meta_valid_devices
elif split == 'meta_test':
device_list = self.meta_test_devices
else: NotImplementedError
tasks = []
for device in device_list:
tasks.append(self.get_task(device))
return tasks
def get_task(self, device=None, num_sample=None):
if num_sample == None:
num_sample = self.num_sample
latency = self.latency[device]
# hardware embedding
hw_embed = latency[self.hw_emb_idx]
hw_embed = normalization(hw_embed, portion=1.0)
# samples for finetuing & test (query)
rand_idx = self.train_idx[device][torch.randperm(len(self.train_idx[device]))]
finetune_idx = rand_idx[:num_sample]
if self.search_space in ['fbnet', 'ofa']:
x_finetune = torch.stack([self.archs[_] for _ in finetune_idx])
x_qry = torch.stack([self.archs[_] for _ in self.valid_idx[device]])
elif self.search_space == 'nasbench201':
x_finetune = [torch.stack([self.archs[_][0] for _ in finetune_idx]),
torch.stack([self.archs[_][1] for _ in finetune_idx])]
x_qry = [torch.stack([self.archs[_][0] for _ in self.valid_idx[device]]),
torch.stack([self.archs[_][1] for _ in self.valid_idx[device]])]
y_finetune = self.norm_latency[device][finetune_idx].view(-1, 1)
y_qry = self.norm_latency[device][self.valid_idx[device]].view(-1, 1)
return hw_embed, x_finetune, y_finetune, x_qry, y_qry, device
def _load_arch_candidates(self):
# architecture candidates obtained by MetaD2A
loaded = open(self.sampled_arch_path, 'r')
self.arch_candidates = {
'arch': [],
'true_acc': []
}
for line in loaded.readlines()[1:]:
arch, true_acc, _ = line.split(',')
self.arch_candidates['arch'].append(arch)
self.arch_candidates['true_acc'].append(true_acc)
def _load_arch_str2idx(self):
self.arch_str2idx = torch.load(os.path.join(self.data_path, 'str_arch2idx.pt'))
def get_nas_task(self, device=None):
num_sample = self.num_sample
latency = self.latency[device]
# hardware embedding
hw_embed = latency[self.hw_emb_idx]
hw_embed = normalization(hw_embed, portion=1.0)
if self.search_space in 'ofa':
# samples for finetuning & test (query)
finetune_idx = self.hw_emb_idx
norm_latency = self.nas_norm_latency[device]
x_finetune = torch.stack([self.archs[_] for _ in finetune_idx])
elif self.search_space == 'nasbench201':
# samples for finetuning & test (query)
rand_idx = torch.randperm(len(self.train_idx[device]))
finetune_idx = self.train_idx[device][rand_idx[:num_sample]]
norm_latency = self.nas_norm_latency[device]
x_finetune = [torch.stack([self.archs[_][0] for _ in finetune_idx]),
torch.stack([self.archs[_][1] for _ in finetune_idx])]
y_finetune = norm_latency[finetune_idx].view(-1, 1)
y_finetune_gt = latency[finetune_idx].view(-1, 1)
if self.search_space == 'nasbench201':
# architecture candidates obtained by MetaD2A
metad2a_idx = [self.arch_str2idx[_] for _ in self.arch_candidates['arch']]
x_qry = [torch.stack([self.archs[_][0] for _ in metad2a_idx]),
torch.stack([self.archs[_][1] for _ in metad2a_idx])]
y_qry = norm_latency[metad2a_idx].view(-1, 1)
y_qry_gt = latency[metad2a_idx].view(-1, 1)
return hw_embed, x_finetune, y_finetune, x_qry, y_qry, device, y_finetune_gt, y_qry_gt
elif self.search_space == 'ofa':
return hw_embed, x_finetune, y_finetune, y_finetune_gt
# def get_nas_task(self, device=None):
# num_sample = self.num_sample
# latency = self.latency[device]
# # hardware embedding
# hw_embed = latency[self.hw_emb_idx]
# hw_embed = normalization(hw_embed, portion=1.0)
# # samples for finetuning & test (query)
# rand_idx = torch.randperm(len(self.train_idx[device]))
# finetune_idx = self.train_idx[device][rand_idx[:num_sample]]
# norm_latency = self.nas_norm_latency[device]
# x_finetune = [torch.stack([self.archs[_][0] for _ in finetune_idx]),
# torch.stack([self.archs[_][1] for _ in finetune_idx])]
# y_finetune = norm_latency[finetune_idx].view(-1, 1)
# y_finetune_gt = latency[finetune_idx].view(-1, 1)
# if self.search_space == 'nasbench201':
# # architecture candidates obtained by MetaD2A
# metad2a_idx = [self.arch_str2idx[_] for _ in self.arch_candidates['arch']]
# x_qry = [torch.stack([self.archs[_][0] for _ in metad2a_idx]),
# torch.stack([self.archs[_][1] for _ in metad2a_idx])]
# y_qry = norm_latency[metad2a_idx].view(-1, 1)
# y_qry_gt = latency[metad2a_idx].view(-1, 1)
# else:
# x_qry, y_qry, y_qry_gt = None, None, None
# return hw_embed, x_finetune, y_finetune, x_qry, y_qry, device, y_finetune_gt, y_qry_gt