-
Notifications
You must be signed in to change notification settings - Fork 268
/
relationnet.py
154 lines (121 loc) · 6.29 KB
/
relationnet.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
# This code is modified from https://github.com/floodsung/LearningToCompare_FSL
import backbone
import torch
import torch.nn as nn
from torch.autograd import Variable
import numpy as np
import torch.nn.functional as F
from methods.meta_template import MetaTemplate
import utils
class RelationNet(MetaTemplate):
def __init__(self, model_func, n_way, n_support, loss_type = 'mse'):
super(RelationNet, self).__init__(model_func, n_way, n_support)
self.loss_type = loss_type #'softmax'# 'mse'
self.relation_module = RelationModule( self.feat_dim , 8, self.loss_type ) #relation net features are not pooled, so self.feat_dim is [dim, w, h]
if self.loss_type == 'mse':
self.loss_fn = nn.MSELoss()
else:
self.loss_fn = nn.CrossEntropyLoss()
def set_forward(self,x,is_feature = False):
z_support, z_query = self.parse_feature(x,is_feature)
z_support = z_support.contiguous()
z_proto = z_support.view( self.n_way, self.n_support, *self.feat_dim ).mean(1)
z_query = z_query.contiguous().view( self.n_way* self.n_query, *self.feat_dim )
z_proto_ext = z_proto.unsqueeze(0).repeat(self.n_query* self.n_way,1,1,1,1)
z_query_ext = z_query.unsqueeze(0).repeat( self.n_way,1,1,1,1)
z_query_ext = torch.transpose(z_query_ext,0,1)
extend_final_feat_dim = self.feat_dim.copy()
extend_final_feat_dim[0] *= 2
relation_pairs = torch.cat((z_proto_ext,z_query_ext),2).view(-1, *extend_final_feat_dim)
relations = self.relation_module(relation_pairs).view(-1, self.n_way)
return relations
def set_forward_adaptation(self,x,is_feature = True): #overwrite parent function
assert is_feature == True, 'Finetune only support fixed feature'
full_n_support = self.n_support
full_n_query = self.n_query
relation_module_clone = RelationModule( self.feat_dim , 8, self.loss_type )
relation_module_clone.load_state_dict(self.relation_module.state_dict())
z_support, z_query = self.parse_feature(x,is_feature)
z_support = z_support.contiguous()
set_optimizer = torch.optim.SGD(self.relation_module.parameters(), lr = 0.01, momentum=0.9, dampening=0.9, weight_decay=0.001)
self.n_support = 3
self.n_query = 2
z_support_cpu = z_support.data.cpu().numpy()
for epoch in range(100):
perm_id = np.random.permutation(full_n_support).tolist()
sub_x = np.array([z_support_cpu[i,perm_id,:,:,:] for i in range(z_support.size(0))])
sub_x = torch.Tensor(sub_x).cuda()
if self.change_way:
self.n_way = sub_x.size(0)
set_optimizer.zero_grad()
y = torch.from_numpy(np.repeat(range( self.n_way ), self.n_query ))
scores = self.set_forward(sub_x, is_feature = True)
if self.loss_type == 'mse':
y_oh = utils.one_hot(y, self.n_way)
y_oh = Variable(y_oh.cuda())
loss = self.loss_fn(scores, y_oh )
else:
y = Variable(y.cuda())
loss = self.loss_fn(scores, y )
loss.backward()
set_optimizer.step()
self.n_support = full_n_support
self.n_query = full_n_query
z_proto = z_support.view( self.n_way, self.n_support, *self.feat_dim ).mean(1)
z_query = z_query.contiguous().view( self.n_way* self.n_query, *self.feat_dim )
z_proto_ext = z_proto.unsqueeze(0).repeat(self.n_query* self.n_way,1,1,1,1)
z_query_ext = z_query.unsqueeze(0).repeat( self.n_way,1,1,1,1)
z_query_ext = torch.transpose(z_query_ext,0,1)
extend_final_feat_dim = self.feat_dim.copy()
extend_final_feat_dim[0] *= 2
relation_pairs = torch.cat((z_proto_ext,z_query_ext),2).view(-1, *extend_final_feat_dim)
relations = self.relation_module(relation_pairs).view(-1, self.n_way)
self.relation_module.load_state_dict(relation_module_clone.state_dict())
return relations
def set_forward_loss(self, x):
y = torch.from_numpy(np.repeat(range( self.n_way ), self.n_query ))
scores = self.set_forward(x)
if self.loss_type == 'mse':
y_oh = utils.one_hot(y, self.n_way)
y_oh = Variable(y_oh.cuda())
return self.loss_fn(scores, y_oh )
else:
y = Variable(y.cuda())
return self.loss_fn(scores, y )
class RelationConvBlock(nn.Module):
def __init__(self, indim, outdim, padding = 0):
super(RelationConvBlock, self).__init__()
self.indim = indim
self.outdim = outdim
self.C = nn.Conv2d(indim, outdim, 3, padding = padding )
self.BN = nn.BatchNorm2d(outdim, momentum=1, affine=True)
self.relu = nn.ReLU()
self.pool = nn.MaxPool2d(2)
self.parametrized_layers = [self.C, self.BN, self.relu, self.pool]
for layer in self.parametrized_layers:
backbone.init_layer(layer)
self.trunk = nn.Sequential(*self.parametrized_layers)
def forward(self,x):
out = self.trunk(x)
return out
class RelationModule(nn.Module):
"""docstring for RelationNetwork"""
def __init__(self,input_size,hidden_size, loss_type = 'mse'):
super(RelationModule, self).__init__()
self.loss_type = loss_type
padding = 1 if ( input_size[1] <10 ) and ( input_size[2] <10 ) else 0 # when using Resnet, conv map without avgpooling is 7x7, need padding in block to do pooling
self.layer1 = RelationConvBlock(input_size[0]*2, input_size[0], padding = padding )
self.layer2 = RelationConvBlock(input_size[0], input_size[0], padding = padding )
shrink_s = lambda s: int((int((s- 2 + 2*padding)/2)-2 + 2*padding)/2)
self.fc1 = nn.Linear( input_size[0]* shrink_s(input_size[1]) * shrink_s(input_size[2]), hidden_size )
self.fc2 = nn.Linear( hidden_size,1)
def forward(self,x):
out = self.layer1(x)
out = self.layer2(out)
out = out.view(out.size(0),-1)
out = F.relu(self.fc1(out))
if self.loss_type == 'mse':
out = F.sigmoid(self.fc2(out))
elif self.loss_type == 'softmax':
out = self.fc2(out)
return out