-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathrewire_tools.py
239 lines (211 loc) · 8.88 KB
/
rewire_tools.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
import copy
import torch
import torch.nn as nn
import numpy as np
def sample_gumbel(shape, eps=1e-20):
U = torch.rand(shape)
return -torch.log(-torch.log(U + eps) + eps)
def softsort(
scores,
tau=1.0,
beta=0.0
):
scores = scores + sample_gumbel(scores.shape).to(scores.device) * beta
scores = scores.unsqueeze(-1)
sorted = scores.sort(dim=-2)[0]
pairwise_diff = (scores.transpose(-2, -1) - sorted).abs().neg()
soft = torch.softmax(pairwise_diff / tau, dim=-1)
hard = torch.zeros_like(soft).scatter_(-1, soft.argmax(dim=-1, keepdim=True), 1)
return hard + soft - soft.detach()
def init_sort(v=None, n=None, k=None, scale=1.):
if v is None:
if k is None:
k = 1
v = torch.stack([torch.arange(n) for _ in range(k)]).float()
else:
v = v.argsort(dim=1).argsort(dim=1).float()
return (v - v.min(dim=1)[0][:, np.newaxis]) / (v.max(dim=1)[0] - v.min(dim=1)[0])[:, np.newaxis] * scale
def mul_grad(x, k):
return x * k - x.detach() * (k - 1)
class LinearRewire(nn.Module):
def __init__(self, in_channels, out_channels, bias=True, tau=1.0, beta=1.0, k=3, tau2=1.0, beta2=1.0, cycle=-1, note=3):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.is_bias = bias
self.linear = nn.Linear(in_channels, out_channels, bias=self.is_bias)
self.tau = tau
self.beta = beta
self.k = k
self.tau2 = tau2
self.beta2 = beta2
self.cycle = cycle
self.note = note # 1 post, 2 pre, 3 both
if self.note != 2:
self.v = nn.Parameter(init_sort(n=out_channels, k=k))
if self.beta == 0:
self.dropout = nn.Dropout(0.1) # alternative way to introduce stochasticity
self.cnt = 0
self.t = None
def forward(self, x, is_train=False, k=0):
out = self.linear(x)
if self.beta == 0:
out = self.dropout(out)
if self.note == 2:
return out
elif self.t is None:
if is_train:
p = softsort(mul_grad(self.v[k], self.k), tau=self.tau, beta=self.beta)
return out @ p.T
else:
return out[..., self.v[k].argsort()]
else:
if self.cycle == -1:
t = min(self.t, self.cnt - 1)
else:
t = min(self.t % self.cycle, self.cnt - 1)
return out[..., self.get_buffer(f"v{t}v")]
def set_task(self, task_id=None, k=0):
if self.note == 2:
if task_id == -1:
self.cnt += 1
elif task_id == -1:
cnt = self.cnt
if self.cycle != -1:
cnt = cnt % self.cycle
self.register_buffer(f"v{cnt}v", self.v[k].argsort())
self.v.data = torch.stack([self.v[k].data.clone() for _ in range(self.k)])
self.cnt += 1
else:
self.t = task_id
if task_id is None:
if (self.cycle != -1) and (self.cnt >= self.cycle):
last_v = self.get_buffer(f"v{self.cnt % self.cycle}v").argsort()
self.v.data = torch.stack([last_v.clone() for _ in range(self.k)])
self.v.data = init_sort(v=self.v.data)
def pre_register_and_consolidate(self):
av, zv = None, None
if self.cnt > 0:
if self.note // 2 == 1:
av = self.av.argsort().argsort()
if self.note % 2 == 1:
zv = self.zv.argsort()
return av, zv
def register_and_consolidate(self, zv, next_av):
# update self.vnv
if (self.cnt > 0) and (self.note != 2):
cnt = self.cnt
if self.cycle != -1:
cnt = min(cnt, self.cycle)
for t in range(cnt):
if next_av is not None:
self.register_buffer(f"v{t}v", zv[self.get_buffer(f"v{t}v")][next_av])
else:
self.register_buffer(f"v{t}v", zv[self.get_buffer(f"v{t}v")])
# register mean
for name, param in self.named_parameters():
if name.endswith('v'): continue
name = name.replace('.', '_')
self.register_buffer(f"{name}_mean", param.data.clone())
# initialize self.vv
if self.note // 2 == 1:
self.register_parameter('av', nn.Parameter(init_sort(n=self.in_channels)[0]))
if self.note % 2 == 1:
self.register_parameter('zv', nn.Parameter(init_sort(n=self.out_channels)[0]))
def add_regularizer(self):
losses = []
if self.note // 2 == 1:
ap = softsort(self.av, tau=self.tau2, beta=self.beta2)
if self.note % 2 == 1:
zp = softsort(self.zv, tau=self.tau2, beta=self.beta2)
for name, param in self.named_parameters():
if name.endswith('v'): continue
name = name.replace('.', '_')
mean = self.get_buffer(f"{name}_mean")
if (self.note == 1) or ((self.note == 3) and name.endswith('bias')):
losses.append(- 2 * ((zp.T @ mean) * param).sum() + (param ** 2).sum())
elif (self.note == 2) and name.endswith('weight'):
losses.append(- 2 * ((mean @ ap) * param).sum() + (param ** 2).sum())
elif (self.note == 3) and name.endswith('weight'):
losses.append(- 2 * ((zp.T @ mean @ ap) * param).sum() + (param ** 2).sum())
else: # self.note == 2 and name.endswith('bias')
losses.append(((mean - param) ** 2).sum())
return losses
def roll_back(self):
for name, param in self.named_parameters():
if name.endswith('v'): continue
name = name.replace('.', '_')
param.data = self.get_buffer(f"{name}_mean")
if (self.cycle != -1) and (self.cnt >= self.cycle):
last_cnt = self.cnt % self.cycle
else:
last_cnt = self.cnt - 1
last_v = self.get_buffer(f"v{last_cnt}v").argsort()
self.v.data = torch.stack([last_v.clone() for _ in range(self.k)])
self.v.data = init_sort(v=self.v.data)
if self.note // 2 == 1:
self.register_parameter('av', nn.Parameter(init_sort(n=self.in_channels)[0]))
if self.note % 2 == 1:
self.register_parameter('zv', nn.Parameter(init_sort(n=self.out_channels)[0]))
class LinearExpand(nn.Module):
def __init__(self, in_channels, out_channels, bias=True, cycle=-1):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.is_bias = bias
self.cycle = cycle
self.linears = nn.ModuleList([])
self.cnt = 0
self.t = None
def forward(self, x):
if self.t is None:
t = self.cnt
if self.cycle != -1:
t = t % self.cycle
else:
if self.cycle == -1:
t = min(self.t, self.cnt - 1)
else:
t = min(self.t % self.cycle, self.cnt - 1)
return self.linears[t](x)
def set_task(self, task_id=None):
if task_id == -1:
self.cnt += 1
else:
self.t = task_id
if task_id is None:
if self.cnt == 0:
self.linears.append(nn.Linear(self.in_channels, self.out_channels, bias=self.is_bias).cuda())
else:
if (self.cycle == -1) or (self.cnt < self.cycle):
self.linears.append(copy.deepcopy(self.linears[-1]))
cnt = self.cnt
last_cnt = self.cnt - 1
else:
cnt = self.cnt % self.cycle
last_cnt = (self.cnt - 1) % self.cycle
for name, param in self.linears[cnt].named_parameters():
param.requires_grad = True
name = name.replace('.', '_')
self.register_buffer(f"{name}_mean", param.data.clone())
for param in self.linears[last_cnt].parameters():
param.requires_grad = False
def roll_back(self):
if (self.cycle == -1) or (self.cnt < self.cycle):
cnt = self.cnt
else:
cnt = self.cnt % self.cycle
for name, param in self.linears[cnt].named_parameters():
name = name.replace('.', '_')
param.data = self.get_buffer(f"{name}_mean")
class SequentialRewire(nn.Sequential):
def forward(self, input, is_train=False, k=0):
for module in self:
input = module(input, is_train, k) if isinstance(module, LinearRewire) else module(input)
return input
def set_task(self, task_id=None, k=0):
for module in self:
if isinstance(module, LinearRewire):
module.set_task(task_id, k)
elif isinstance(module, LinearExpand):
module.set_task(task_id)