forked from facebookresearch/luckmatters
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_gen.py
317 lines (267 loc) · 9.88 KB
/
model_gen.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
# Copyright (c) Facebook, Inc. and its affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import sys
import os
import torch
import torch.nn as nn
import random
from theory_utils import haar_measure, init_separate_w
# Generate random orth matrix.
import numpy as np
import math
def get_aug_w(w):
# w: [output_d, input_d]
# aug_w: [output_d + 1, input_d + 1]
output_d, input_d = w.weight.size()
aug_w = torch.zeros( (output_d + 1, input_d + 1), dtype = w.weight.dtype, device = w.weight.device)
aug_w[:output_d, :input_d] = w.weight.data
aug_w[:output_d, input_d] = w.bias.data
aug_w[output_d, input_d] = 1
return aug_w
def set_orth(layer):
w = layer.weight
orth = haar_measure(w.size(1))
w.data = torch.from_numpy(orth[:w.size(0), :w.size(1)].astype('f4')).cuda()
def set_add_noise(layer, teacher_layer, perturb):
layer.weight.data[:] = teacher_layer.weight.data[:] + torch.randn(teacher_layer.weight.size()).cuda() * perturb
layer.bias.data[:] = teacher_layer.bias.data[:] + torch.randn(teacher_layer.bias.size()).cuda() * perturb
def set_same_dir(layer, teacher_layer):
norm = layer.weight.data.norm()
r = norm / teacher_layer.weight.data.norm()
layer.weight.data[:] = teacher_layer.weight.data * r
layer.bias.data[:] = teacher_layer.bias.data * r
def set_same_sign(layer, teacher_layer):
sel = (teacher_layer.weight.data > 0) * (layer.weight.data < 0) + (teacher_layer.weight.data < 0) * (layer.weight.data > 0)
layer.weight.data[sel] *= -1.0
sel = (teacher_layer.bias.data > 0) * (layer.bias.data < 0) + (teacher_layer.bias.data < 0) * (layer.bias.data > 0)
layer.bias.data[sel] *= -1.0
def normalize_layer(layer):
# [output, input]
w = layer.weight.data
for i in range(w.size(0)):
norm = w[i].pow(2).sum().sqrt() + 1e-5
w[i] /= norm
if layer.bias is not None:
layer.bias.data[i] /= norm
def init_w(layer, use_sep=True):
sz = layer.weight.size()
output_d = sz[0]
input_d = 1
for s in sz[1:]:
input_d *= s
if use_sep:
choices = [-0.5, -0.25, 0, 0.25, 0.5]
layer.weight.data[:] = torch.from_numpy(init_separate_w(output_d, input_d, choices)).view(*sz).cuda()
if layer.bias is not None:
layer.bias.data.uniform_(-.5, 0.5)
def init_w2(w, multiplier=5):
w.weight.data *= multiplier
w.bias.data.normal_(0, std=1)
# w.bias.data *= 5
for i, ww in enumerate(w.weight.data):
pos_ratio = (ww > 0.0).sum().item() / w.weight.size(1) - 0.5
w.bias.data[i] -= pos_ratio
class Model(nn.Module):
def __init__(self, d, ks, d_output, multi=1, has_bn=True, has_bn_affine=True, has_bias=True, bn_before_relu=False):
super(Model, self).__init__()
self.d = d
self.ks = ks
self.has_bn = has_bn
self.ws_linear = nn.ModuleList()
self.ws_bn = nn.ModuleList()
self.bn_before_relu = bn_before_relu
last_k = d
self.sizes = [d]
for k in ks:
k *= multi
self.ws_linear.append(nn.Linear(last_k, k, bias=has_bias))
if has_bn:
self.ws_bn.append(nn.BatchNorm1d(k, affine=has_bn_affine))
self.sizes.append(k)
last_k = k
self.final_w = nn.Linear(last_k, d_output, bias=has_bias)
self.relu = nn.ReLU()
self.sizes.append(d_output)
def init_orth(self):
for w in self.ws:
set_orth(w)
set_orth(self.final_w)
def set_teacher(self, teacher, perturb):
for w_s, w_t in zip(self.ws, teacher.ws):
set_add_noise(w_s, w_t, perturb)
set_add_noise(self.final_w, teacher.final_w, perturb)
def set_teacher_dir(self, teacher):
for w_s, w_t in zip(self.ws, teacher.ws):
set_same_dir(w_s, w_t)
set_same_dir(self.final_w, teacher.final_w)
def set_teacher_sign(self, teacher):
for w_s, w_t in zip(self.ws, teacher.ws):
set_same_sign(w_s, w_t)
set_same_sign(self.final_w, teacher.final_w)
def forward(self, x):
hs = []
pre_bns = []
#bns = []
h = x
for i in range(len(self.ws_linear)):
w = self.ws_linear[i]
h = w(h)
if self.bn_before_relu:
pre_bns.append(h)
if len(self.ws_bn) > 0:
bn = self.ws_bn[i]
h = bn(h)
h = self.relu(h)
else:
h = self.relu(h)
pre_bns.append(h)
if len(self.ws_bn) > 0:
bn = self.ws_bn[i]
h = bn(h)
hs.append(h)
#bns.append(h)
y = self.final_w(hs[-1])
return dict(hs=hs, pre_bns=pre_bns, y=y)
def init_w(self, use_sep=True):
for w in self.ws_linear:
init_w(w, use_sep=use_sep)
init_w(self.final_w, use_sep=use_sep)
def reset_parameters(self):
for w in self.ws_linear:
w.reset_parameters()
for w in self.ws_bn:
w.reset_parameters()
self.final_w.reset_parameters()
def normalize(self):
for w in self.ws_linear:
normalize_layer(w)
normalize_layer(self.final_w)
def from_bottom_linear(self, j):
if j < len(self.ws_linear):
return self.ws_linear[j].weight.data
elif j == len(self.ws_linear):
return self.final_w.weight.data
else:
raise RuntimeError("j[%d] is out of bound! should be [0, %d]" % (j, len(self.ws)))
def from_bottom_aug_w(self, j):
if j < len(self.ws_linear):
return get_aug_w(self.ws_linear[j])
elif j == len(self.ws_linear):
return get_aug_w(self.final_w)
else:
raise RuntimeError("j[%d] is out of bound! should be [0, %d]" % (j, len(self.ws)))
def num_layers(self):
return len(self.ws_linear) + 1
def from_bottom_bn(self, j):
assert j < len(self.ws_bn)
return self.ws_bn[j]
class ModelConv(nn.Module):
def __init__(self, input_size, ks, d_output, multi=1, has_bn=True, bn_before_relu=False):
super(ModelConv, self).__init__()
self.ks = ks
self.ws_linear = nn.ModuleList()
self.ws_bn = nn.ModuleList()
self.bn_before_relu = bn_before_relu
init_k, h, w = input_size
last_k = init_k
for k in ks:
k *= multi
self.ws_linear.append(nn.Conv2d(last_k, k, 3))
if has_bn:
self.ws_bn.append(nn.BatchNorm2d(k))
last_k = k
h -= 2
w -= 2
self.final_w = nn.Linear(last_k * h * w, d_output)
self.relu = nn.ReLU()
def forward(self, x):
hs = []
#bns = []
h = x
for i in range(len(self.ws_linear)):
w = self.ws_linear[i]
h = w(h)
if self.bn_before_relu:
if len(self.ws_bn) > 0:
bn = self.ws_bn[i]
h = bn(h)
h = self.relu(h)
else:
h = self.relu(h)
if len(self.ws_bn) > 0:
bn = self.ws_bn[i]
h = bn(h)
hs.append(h)
#bns.append(h)
h = hs[-1].view(h.size(0), -1)
y = self.final_w(h)
return dict(hs=hs, y=y)
def init_w(self, use_sep=True):
for w in self.ws_linear:
init_w(w, use_sep=use_sep)
init_w(self.final_w, use_sep=use_sep)
def normalize(self):
for w in self.ws_linear:
normalize_layer(w)
normalize_layer(self.final_w)
def normalize_last(self):
normalize_layer(self.final_w)
def reset_parameters(self):
for w in self.ws_linear:
w.reset_parameters()
for w in self.ws_bn:
w.reset_parameters()
self.final_w.reset_parameters()
def from_bottom_linear(self, j):
if j < len(self.ws_linear):
return self.ws_linear[j].weight.data
elif j == len(self.ws_linear):
return self.final_w.weight.data
else:
raise RuntimeError("j[%d] is out of bound! should be [0, %d]" % (j, len(self.ws)))
def num_layers(self):
return len(self.ws_linear) + 1
def from_bottom_bn(self, j):
assert j < len(self.ws_bn)
return self.ws_bn[j]
def prune(net, ratios):
# Prune the network and finetune.
n = net.num_layers()
# Compute L1 norm and and prune them globally
masks = []
inactive_nodes = []
for i in range(1, n):
W = net.from_bottom_linear(i)
# Prune all input neurons
input_dim = W.size(1)
fc_to_conv = False
if isinstance(net, ModelConv):
if len(W.size()) == 4:
# W: [output_filter, input_filter, x, y]
w_norms = W.permute(1, 0, 2, 3).contiguous().view(W.size(1), -1).abs().mean(1)
else:
# The final FC layer.
input_dim = net.from_bottom_linear(i - 1).size(0)
W_reshaped = W.view(W.size(0), -1, input_dim)
w_norms = W_reshaped.view(-1, input_dim).abs().mean(0)
fc_to_conv = True
else:
# W: [output_dim, input_dim]
w_norms = W.abs().mean(0)
sorted_w, sorted_indices = w_norms.sort(0)
n_pruned = int(input_dim * ratios[i - 1])
inactive_mask = sorted_indices[:n_pruned]
m = W.clone().fill_(1.0)
if fc_to_conv:
m = m.view(m.size(0), -1, input_dim)
m[:, :, inactive_mask] = 0
m = m.view(W.size(0), W.size(1))
else:
m[:, inactive_mask] = 0
# Set the mask for the lower layer to zero.
inactive_nodes.append(inactive_mask.cpu().tolist())
masks.append(m)
return inactive_nodes, masks