-
Notifications
You must be signed in to change notification settings - Fork 1
/
distill_higher.py
188 lines (163 loc) · 6.94 KB
/
distill_higher.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as init
import torch.optim as optim
import torch.multiprocessing as mp
import higher
from copy import deepcopy
import numpy as np
import random
import time
from framework.config import get_arch
def _weights_init(m):
reset_parameters = getattr(m, "reset_parameters", None)
if callable(reset_parameters):
init.kaiming_normal_(m.weight)
class Distill(nn.Module):
def __init__(self, x_init, y_init, arch, window, lr, num_train_eval, img_pc, batch_pc, num_classes=2, task_sampler_nc=2, train_y=False,
channel=3, im_size=(32, 32), inner_optim='SGD', syn_intervention=None, real_intervention=None, cctype=0):
super(Distill, self).__init__()
self.data = nn.Embedding(img_pc*num_classes, int(channel*np.prod(im_size)))
self.train_y = train_y
if train_y:
self.label = nn.Embedding(img_pc*num_classes, num_classes)
self.label.weight.data = y_init.float().cuda()
else:
self.label = y_init
self.num_classes = num_classes
self.channel = channel
self.im_size = im_size
self.net = get_arch(arch, self.num_classes, self.channel, self.im_size)
self.img_pc = img_pc
self.batch_pc = batch_pc
self.arch = arch
self.lr = lr
self.window = window
self.criterion = nn.CrossEntropyLoss(reduction='mean')
self.num_train_eval = num_train_eval
self.curriculum = window
self.inner_optim = inner_optim
self.batch_id = 0
self.syn_intervention = syn_intervention
self.real_intervention = real_intervention
self.task_sampler_nc = task_sampler_nc
self.cctype = cctype
# shuffle the data
def shuffle(self):
#True
self.order_list = torch.randperm(self.img_pc)
if self.img_pc >= self.batch_pc:
self.order_list = torch.cat([self.order_list, self.order_list], dim=0)
# randomly sample label sets from the full label set
def get_task_indices(self):
task_indices = list(range(self.num_classes))
if self.task_sampler_nc < self.num_classes:
random.shuffle(task_indices)
task_indices = task_indices[:self.task_sampler_nc]
task_indices.sort()
return task_indices
def subsample(self):
indices = []
if self.task_sampler_nc == self.num_classes:
for i in range(self.num_classes):
ind = torch.randperm(self.img_pc)[:self.batch_pc].sort()[0] + self.img_pc * i
indices.append(ind)
else:
task_indices = self.get_task_indices()
for i in task_indices:
ind = torch.randperm(self.img_pc)[:self.batch_pc].sort()[0] + self.img_pc * i
indices.append(ind)
indices = torch.cat(indices).cuda()
imgs = self.data(indices)
imgs = imgs.view(
self.task_sampler_nc * min(self.img_pc, self.batch_pc),
self.channel,
self.im_size[0],
self.im_size[1]
).contiguous()
if self.train_y:
labels = self.label(indices)
labels = labels.view(
self.task_sampler_nc * min(self.img_pc, self.batch_pc),
self.num_classes
).contiguous()
else:
labels = self.label[indices]
return imgs, labels
def forward(self, x):
self.net = get_arch(self.arch, self.num_classes, self.channel, self.im_size).cuda()
self.net.train()
if self.inner_optim == 'SGD':
self.optimizer = optim.SGD(self.net.parameters(), lr=self.lr, momentum=0.9, weight_decay=5e-4)
# TODO: add decay rules for SGD
elif self.inner_optim == 'Adam':
self.optimizer = optim.Adam(self.net.parameters(), lr=self.lr)
if self.dd_type not in ['curriculum', 'standard']:
print('The dataset distillation method is not implemented!')
raise NotImplementedError()
if self.dd_type == 'curriculum':
for i in range(self.curriculum):
self.optimizer.zero_grad()
imgs, label = self.subsample()
imgs = self.syn_intervention(imgs, dtype='syn', seed=random.randint(0, 10000))
ratio = 0
out, pres = self.net(imgs)
loss = self.criterion(out, label)
loss.backward()
self.optimizer.step()
with higher.innerloop_ctx(
self.net, self.optimizer, copy_initial_weights=True
) as (fnet, diffopt):
for i in range(self.window):
imgs, label = self.subsample()
imgs = self.syn_intervention(imgs, dtype='syn', seed=random.randint(0, 10000))
ratio = 0
out, pres = fnet(imgs)
loss = self.criterion(out, label)
diffopt.step(loss)
x = self.real_intervention(x, dtype='real', seed=random.randint(0, 10000))
return fnet(x)
def init_train(self, epoch, init=False, lim=True):
if init:
self.net = get_arch(self.arch, self.num_classes, self.channel, self.im_size).cuda()
if self.inner_optim == 'SGD':
self.optimizer = optim.SGD(self.net.parameters(), lr=self.lr, momentum=0.9, weight_decay=5e-4)
elif self.inner_optim == 'Adam':
self.optimizer = optim.Adam(self.net.parameters(), lr=self.lr)
#self.shuffle()
for i in range(epoch):
self.optimizer.zero_grad()
imgs, label = self.subsample()
imgs = self.syn_intervention(imgs, dtype='syn', seed=random.randint(0, 10000))
out, pres = self.net(imgs)
loss = self.criterion(out, label)
loss.backward()
self.optimizer.step()
# initialize the EMA with the currect data value
def ema_init(self, ema_coef):
self.shadow = -1e5
self.ema_coef = ema_coef
# update the EMA value
def ema_update(self, grad_norm):
if self.shadow == -1e5:
self.shadow = grad_norm
else:
self.shadow -= (1 - self.ema_coef) * (self.shadow - grad_norm)
return self.shadow
def test(self, x):
with torch.no_grad():
out = self.net(x)
return out
def random_indices(y, nclass=10, intraclass=False, device='cuda'):
n = len(y)
if intraclass:
index = torch.arange(n).to(device)
for c in range(nclass):
index_c = index[y == c]
if len(index_c) > 0:
randidx = torch.randperm(len(index_c))
index[y == c] = index_c[randidx]
else:
index = torch.randperm(n).to(device)
return index