-
Notifications
You must be signed in to change notification settings - Fork 8
/
hamburger.py
244 lines (195 loc) · 7.26 KB
/
hamburger.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
'''
_..----.._
.' o '.
/ o o \
|o o o|
/'-.._o __.-'\
\ ````` /
|``--........--'`|
\ /
`'----------'`
'''
#%%
import yaml
with open('config.yaml') as fh:
config = yaml.load(fh, Loader=yaml.FullLoader)
import math
import torch
from functools import partial
from torch import nn
import torch.nn.functional as F
from bricks import NormLayer
'''
Get Bread
'''
from bricks import ConvBNRelu
'''
Get Patty
'''
class _MatrixDecomposition2DBase(nn.Module):
'''
Base class for furhter implementing the NMF, VQ or CD as in paper
https://arxiv.org/pdf/2109.04553.pdf
this script only has NMF as it has best performance for semantic segmentation
as mentioned in paper
D (dictionery) in paper is bases
C (codes) in paper is coef here
'''
def __init__(self, config):
super().__init__()
self.spatial = config['SPATIAL']
self.S = config['MD_S']
self.D = config['MD_D']
self.R = config['MD_R']
self.train_steps = config['TRAIN_STEPS']
self.eval_steps = config['EVAL_STEPS']
self.inv_t = config['INV_T']
self.eta = config['Eta']
self.rand_init = config['RAND_INIT']
print(30*'=')
print('spatial: ', self.spatial)
print('S: ', self.S)
print('D: ', self.D)
print('R: ', self.R)
print('train_steps: ', self.train_steps)
print('eval_steps: ', self.eval_steps)
print('inv_t: ', self.inv_t)
print('eta: ', self.eta)
print('rand_init: ', self.rand_init)
print(30*'=')
def _bild_bases(self, B,S,D,R):
raise NotImplementedError
def local_setp(self, x, bases, coef):
raise NotImplementedError
def compute_coef(self, x, bases, coef):
raise NotImplementedError
@torch.no_grad()
def local_inference(self, x, bases):
# (B * S, D, N)^T @ (B * S, D, R) -> (B * S, N, R)
# here: N = HW and D = C in case of spatial attention
coef = torch.bmm(x.transpose(1,2), bases)
# column wise softmax ignore batch dim, i.e, on HW dim
coef = F.softmax(self.inv_t*coef, dim=-1)
steps = self.train_steps if self.training else self.eval_steps
for _ in range(steps):
bases, coef = self.local_setp(x, bases, coef)
return bases, coef
@torch.no_grad()
def online_update(self, bases):
update = bases.mean(dim=0)
self.bases += self.eta * (update - self.bases)
# column wise normalization i.e. HW dim
self.bases = F.normalize(self.bases, dim=1)
return None
def forward(self, x, return_bases=False):
B, C, H, W = x.shape
if self.spatial:
# spatial attention k
D = C // self.S # reduce channels
N = H * W
x = x.view(B * self.S, D, N)
else:
D = H * W
N = C // self.S
x = x.view(B * self.S, N, D).transpose(1, 2)
if not self.rand_init and not hasattr(self, 'bases'):
bases = self._build_bases(1, self.S, D, self.R)
self.register_buffer('bases', bases)
# (S, D, R) -> (B * S, D, R)
if self.rand_init:
bases = self._build_bases(B, self.S, D, self.R)
else:
bases = self.bases.repeat(B,1,1)
bases, coef = self.local_inference(x, bases)
# (B * S, N, R)
coef = self.compute_coef(x, bases, coef)
# (B * S, D, R) @ (B * S, N, R)^T -> (B * S, D, N)
x = torch.bmm(bases, coef.transpose(1,2))
# (B * S, D, N) -> (B, C, H, W)
if self.spatial:
x = x.view(B, C, H, W)
else:
x = x.transpose(1,2).view(B, C, H, W)
bases = bases.view(B, self.S, D, self.R)
if not self.rand_init and not self.training and not return_bases:
self.online_update(bases)
return x
class NMF2D(_MatrixDecomposition2DBase):
def __init__(self, config):
super().__init__(config)
self.inv_t = 1
def _build_bases(self, B, S, D, R):
bases = torch.rand((B*S, D, R)).to('cuda' if torch.cuda.is_available() else 'cpu')
bases = F.normalize(bases, dim=1) # column wise normalization i.e HW dim
return bases
@torch.no_grad()
def local_setp(self, x, bases, coef):
'''
Algorithm 2 in paper
NMF with multiliplicative update.
'''
# coef (C/codes)update
# (B*S, D, N)T @ (B*S, D, R) -> (B*S, N, R)
numerator = torch.bmm(x.transpose(1,2), bases) # D^T @ X
# (BS, N, R) @ [(BS, D, R)T @ (BS, D, R)] -> (BS, N, R)
denominator = coef.bmm(bases.transpose(1,2).bmm(bases)) # D^T @ D @ C
# Multiplicative update
coef = coef * (numerator / (denominator + 1e-7)) # updated C
# bases (D/dict) update
# (BS, D, N) @ (BS, N, R) -> (BS, D, R)
numerator = torch.bmm(x, coef) # X @ C^T
# (BS, D, R) @ [(BS, D, R)T @ (BS, D, R)] -> (BS, D, R)
denominator = bases.bmm(coef.transpose(1,2).bmm(coef)) # D @ D @ C^T
# Multiplicative update
bases = bases * (numerator / (denominator + 1e-7)) # updated D
return bases, coef
def compute_coef(self, x, bases, coef):
# (B*S, D, N)T @ (B*S, D, R) -> (B*S, N, R)
numerator = torch.bmm(x.transpose(1,2), bases) # D^T @ X
# (BS, N, R) @ [(BS, D, R)T @ (BS, D, R)] -> (BS, N, R)
denominator = coef.bmm(bases.transpose(1,2).bmm(bases)) # D^T @ D @ C
# Multiplicative update
coef = coef * (numerator / (denominator + 1e-7))
return coef
'''
Make Burger
'''
class HamBurger(nn.Module):
def __init__(self, inChannels, config):
super().__init__()
self.put_cheese = config['put_cheese']
C = config["MD_D"]
# add Relu at end as NMF works of non-negative only
self.lower_bread = nn.Sequential(nn.Conv2d(inChannels, C, 1),
nn.ReLU(inplace=True)
)
self.ham = NMF2D(config)
self.cheese = ConvBNRelu(C, C)
self.upper_bread = nn.Conv2d(C, inChannels, 1, bias=False)
# self.init_weights()
# def init_weights(self):
# for m in self.modules():
# if isinstance(m, nn.Conv2d):
# fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
# fan_out //= m.groups
# nn.init.normal_(m.weight, std=math.sqrt(2.0/fan_out), mean=0)
def forward(self, x):
skip = x.clone()
x = self.lower_bread(x)
x = self.ham(x)
if self.put_cheese:
x = self.cheese(x)
x = self.upper_bread(x)
x = F.relu(x + skip, inplace=True)
return x
def online_update(self, bases):
if hasattr(self.ham, 'online_update'):
self.ham.online_update(bases)
#%%
# from torchsummary import summary
# model = HamBurger(inChannels=512, config=config)
# summary(model, (512,256,256))
# model = model.to('cuda' if torch.cuda.is_available() else 'cpu')
# y = torch.randn((6,512,32,32)).to('cuda' if torch.cuda.is_available() else 'cpu')
# x = model.forward(y)
# print(x.shape)