-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathpoint_rend.py
175 lines (154 loc) · 7.64 KB
/
point_rend.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import time
class MLP(nn.Module):
def __init__(self, nIn, num_class):
super(MLP, self).__init__()
self.bn0 = nn.BatchNorm2d(54)
self.conv1 = nn.Conv2d(nIn, 256, kernel_size=1, padding=0, bias=False)
self.bn1 = nn.BatchNorm2d(259)
self.conv2 = nn.Conv2d(259, 128, kernel_size=1, padding=0, bias=False)
self.bn2 = nn.BatchNorm2d(131)
self.conv3 = nn.Conv2d(131, num_class, kernel_size=1, padding=0, bias=False)
#self.ReLU = nn.ReLU(inplace=True)
self.ReLU = nn.PReLU()
self._init_weight()
def _init_weight(self):
for m in self.modules():
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
nn.init.xavier_normal_(m.weight)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def forward(self, x):
#x:batch,N,channel
#model_input:batch,channel,N,1
model_input = x.unsqueeze(2).permute([0,3,1,2])
#model_input = self.bn0(model_input_pre)
print('model_input:')
print(model_input[0,:,0,0])
print(model_input.shape)
layer1 = self.conv1(model_input)
print('layer1')
print(layer1.shape)
#layer1_output = self.ReLU(self.bn1(torch.cat((layer1, model_input[:,-3:,:,:]),1)))
#layer2 = self.conv2(layer1_output)
#layer2_output = self.ReLU(self.bn2(torch.cat((layer2, model_input[:,-3:,:,:]),1)))
#batch,num_class,N
#last_conv = torch.squeeze(self.conv3(layer2_output),3)
layer1_output = self.ReLU(layer1)
layer2 = self.conv2(torch.cat((layer1_output, model_input[:,-3:,:,:]),1))
layer2_output = self.ReLU(layer2)
last_conv = torch.squeeze(self.conv3(torch.cat((layer2_output, model_input[:,-3:,:,:]),1)),3)
print('layer1_output:')
print(layer1_output[0,:,0,0])
print(layer1_output.shape)
print('layer2_output:')
print(layer2_output[0,:,0,0])
print(layer2_output.shape)
#batch,num_class,N
return last_conv
class PointRend(nn.Module):
def __init__(self, num_classes,N,coarse_size,fine_size,img_size,fine_channels,is_training=True,k=3,belta=0.75):
super(PointRend, self).__init__()
self.num_classes = num_classes
self.N = N
self.is_training = is_training
self.coarse_size = coarse_size
self.img_size = img_size
self.fine_size = fine_size
self.k = k
self.belta = belta
self.mlp = MLP(fine_channels+3, num_classes)
if self.is_training:
self.up1 = nn.Upsample(scale_factor=img_size[0]/coarse_size[0], mode='bilinear')
self.up2 = nn.Upsample(scale_factor=img_size[0]/fine_size[0], mode='bilinear')
else:
self.up = nn.Upsample(scale_factor=2, mode='bilinear')
def forward(self, fine_grained, coarse_pre):
if self.is_training:
#batch,c,h,w
up_coarse_o = self.up1(coarse_pre)
up_coarse_softmax = F.softmax(up_coarse_o, dim=1)
up_fine_o = self.up2(fine_grained)
#batch,h,w,c
up_coarse = up_coarse_softmax.permute([0,2,3,1])
up_coarse_ori_fea = up_coarse_o.permute([0,2,3,1])
up_fine = up_fine_o.permute([0,2,3,1])
#选点方法一
num_over = self.k * self.N
step1_n = int(self.N * self.belta)
step2_n = self.N - step1_n
#torch产生整形随机数作为坐标，组合新的feature map，计算度量距离，计算topk的index，采样剩余index，组合特征向量
#index = (torch.rand(num_over, 2)*self.img_size[0]).int()
random_1d_index = torch.randperm(self.img_size[0]*self.img_size[1])
h_index = (random_1d_index/self.img_size[1]).int()
w_index = (random_1d_index%self.img_size[1]).int()
all_index = torch.cat((h_index.reshape(-1,1),w_index.reshape(-1,1)),1)
index = all_index[:num_over,:].cuda()
pre_list = []
for i in range(num_over):
#batch,3
one_shot = up_coarse[:,index[i][0],index[i][1],:]
#batch,1,3
one_shot_ext = one_shot.unsqueeze(1)
pre_list.append(one_shot_ext)
#batch,num_over,3
over_pre = torch.cat(pre_list, 1)
#计算置信度最高的两个值,top2:n,num_over,2
top2,_ = torch.topk(over_pre,2,2)
#batch,num_over
certain_score = top2[:,:,0] - top2[:,:,1]
uncertain_score = 1-certain_score
#batch, step1_n 代表over_pre中step1_n个uncertain点
_,uncertain_index = torch.topk(uncertain_score,step1_n,1)
uncertain_index_reshape = uncertain_index.reshape([-1,1])
uncertain_index_repeat = uncertain_index_reshape.repeat(1,2)
uncertain_points_unbatch = torch.gather(index, dim=0, index=uncertain_index_repeat)
#batch,step1_n,2 uncertain点在原图中的坐标
uncertain_points = uncertain_points_unbatch.reshape([-1,step1_n,2])
#covrage_points = (torch.rand(step2_n, 2)*self.img_size[0]).int()
covrage_points = all_index[-step2_n:,:].unsqueeze(0).repeat([uncertain_points.shape[0],1,1])
#batch,N,2 选出的所有点
all_select_point = torch.cat([uncertain_points.int().cuda(),covrage_points.cuda()],1)
#组合特征向量
batch_id = torch.arange(all_select_point.shape[0])
#1,batch*N
batch_index = batch_id.reshape([-1,1]).repeat([1,self.N]).reshape([1,-1]).int()
#2,batch*N
all_select_point_prtmute = all_select_point.permute([2,0,1]).reshape(2,-1)
#3,batch*N
mask_pre_data = torch.cat((batch_index.cuda(),all_select_point_prtmute.cuda()),0)
val = torch.tensor([1]*mask_pre_data.shape[1]).long()
#batch,h,w
mask = torch.sparse.FloatTensor(mask_pre_data.long().cuda(), val.long().cuda(), torch.Size([all_select_point.shape[0],self.img_size[0],self.img_size[1]])).to_dense()
#mask = torch.where(mask_tmp>1,torch.tensor([1]).int(),mask_tmp)
select_coarse = torch.masked_select(up_coarse_ori_fea,mask[:,:,:,None].byte())
select_fine = torch.masked_select(up_fine,mask[:,:,:,None].byte())
#batch,N,channel+3
select_feature = torch.cat([select_fine, select_coarse], -1).reshape([all_select_point.shape[0],self.N, -1])
#batch,num_class,N
out = self.mlp(select_feature)
pre = F.softmax(out, dim=1)
#debug
print(pre[0])
ori_pre = select_coarse.reshape([all_select_point.shape[0],self.N, -1]).permute([0,2,1])
print(F.softmax(ori_pre, dim=1)[0])
#return select_coarse.reshape([all_select_point.shape[0],self.N, -1]).permute([0,2,1]), mask
return out, mask
if __name__ == '__main__':
net = PointRend(3,1000,[128,128],[128,128],[256,256],128)
net = net.cuda()
for i in range(5):
start = time.time()
coarse_prediction=torch.rand([32, 3, 128, 128]).cuda()
fine_grained = torch.rand([32, 128, 128, 128]).cuda()
out,mask = net(up_fine, up_coarse)
end = time.time()
print(end-start)
print(out.shape)
print(mask.shape)
torch.save(net, 'model.pkl')