forked from andy-yun/pytorch-0.4-yolov3
-
Notifications
You must be signed in to change notification settings - Fork 0
/
region_layer.py
188 lines (162 loc) · 8.77 KB
/
region_layer.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
import math
import numpy as np
import sys
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
from utils import bbox_iou, multi_bbox_ious, convert2cpu
class RegionLayer(nn.Module):
def __init__(self, num_classes=0, anchors=[1.0], num_anchors=1, use_cuda=None):
super(RegionLayer, self).__init__()
use_cuda = torch.cuda.is_available() and (True if use_cuda is None else use_cuda)
self.device = torch.device("cuda" if use_cuda else "cpu")
self.num_classes = num_classes
self.num_anchors = num_anchors
self.anchor_step = len(anchors)//num_anchors
#self.anchors = torch.stack(torch.FloatTensor(anchors).split(self.anchor_step)).to(self.device)
self.anchors = torch.FloatTensor(anchors).view(self.num_anchors, self.anchor_step).to(self.device)
self.rescore = 1
self.coord_scale = 1
self.noobject_scale = 1
self.object_scale = 5
self.class_scale = 1
self.thresh = 0.6
self.seen = 0
def build_targets(self, pred_boxes, target, nH, nW):
nB = target.size(0)
nA = self.num_anchors
noobj_mask = torch.ones (nB, nA, nH, nW)
obj_mask = torch.zeros(nB, nA, nH, nW)
coord_mask = torch.zeros(nB, nA, nH, nW)
tcoord = torch.zeros( 4, nB, nA, nH, nW)
tconf = torch.zeros(nB, nA, nH, nW)
tcls = torch.zeros(nB, nA, nH, nW)
nAnchors = nA*nH*nW
nPixels = nH*nW
nGT = 0 # number of ground truth
nRecall = 0
# it works faster on CPU than on GPU.
anchors = self.anchors.to("cpu")
if self.seen < 12800:
tcoord[0].fill_(0.5)
tcoord[1].fill_(0.5)
coord_mask.fill_(0.01)
# initial w, h == 0 means log(1)==0, s.t, anchor is equal to ground truth.
for b in range(nB):
cur_pred_boxes = pred_boxes[b*nAnchors:(b+1)*nAnchors].t()
cur_ious = torch.zeros(nAnchors)
tbox = target[b].view(-1,5).to("cpu")
for t in range(50):
if tbox[t][1] == 0:
break
gx, gw = [ i * nW for i in (tbox[t][1], tbox[t][3]) ]
gy, gh = [ i * nH for i in (tbox[t][2], tbox[t][4]) ]
cur_gt_boxes = torch.FloatTensor([gx, gy, gw, gh]).repeat(nAnchors,1).t()
cur_ious = torch.max(cur_ious, multi_bbox_ious(cur_pred_boxes, cur_gt_boxes, x1y1x2y2=False))
ignore_ix = (cur_ious>self.thresh).view(nA,nH,nW)
noobj_mask[b][ignore_ix] = 0
for t in range(50):
if tbox[t][1] == 0:
break
nGT += 1
gx, gw = [ i * nW for i in (tbox[t][1], tbox[t][3]) ]
gy, gh = [ i * nH for i in (tbox[t][2], tbox[t][4]) ]
gw, gh = gw.float(), gh.float()
gi, gj = int(gx), int(gy)
tmp_gt_boxes = torch.FloatTensor([0, 0, gw, gh]).repeat(nA,1).t()
anchor_boxes = torch.cat((torch.zeros(nA, 2), anchors),1).t()
tmp_ious = multi_bbox_ious(anchor_boxes, tmp_gt_boxes, x1y1x2y2=False)
best_iou, best_n = torch.max(tmp_ious, 0)
if self.anchor_step == 4: # this part is not tested.
tmp_ious_mask = (tmp_ious==best_iou)
if tmp_ious_mask.sum() > 0:
gt_pos = torch.FloatTensor([gi, gj, gx, gy]).repeat(nA,1).t()
an_pos = anchor_boxes[4:6] # anchor_boxes are consisted of [0 0 aw ah ax ay]
dist = pow(((gt_pos[0]+an_pos[0])-gt_pos[2]),2) + pow(((gt_pos[1]+an_pos[1])-gt_pos[3]),2)
dist[1-tmp_ious_mask]=10000 # set the large number for the small ious
_, best_n = torch.min(dist,0)
gt_box = torch.FloatTensor([gx, gy, gw, gh])
pred_box = pred_boxes[b*nAnchors+best_n*nPixels+gj*nW+gi]
iou = bbox_iou(gt_box, pred_box, x1y1x2y2=False)
obj_mask [b][best_n][gj][gi] = 1
noobj_mask[b][best_n][gj][gi] = 0
coord_mask[b][best_n][gj][gi] = 2. - tbox[t][3]*tbox[t][4]
tcoord [0][b][best_n][gj][gi] = gx - gi
tcoord [1][b][best_n][gj][gi] = gy - gj
tcoord [2][b][best_n][gj][gi] = math.log(gw/anchors[best_n][0])
tcoord [3][b][best_n][gj][gi] = math.log(gh/anchors[best_n][1])
tcls [b][best_n][gj][gi] = tbox[t][0]
tconf [b][best_n][gj][gi] = iou if self.rescore else 1.
if iou > 0.5:
nRecall += 1
return nGT, nRecall, obj_mask, noobj_mask, coord_mask, tcoord, tconf, tcls
def get_mask_boxes(self, output):
if not isinstance(self.anchors, torch.Tensor):
self.anchors = torch.FloatTensor(self.anchors).view(self.num_anchors, self.anchor_step).to(self.device)
masked_anchors = self.anchors.view(-1)
num_anchors = torch.IntTensor([self.num_anchors]).to(self.device)
return {'x':output, 'a':masked_anchors, 'n':num_anchors}
def forward(self, output, target):
#output : BxAs*(4+1+num_classes)*H*W
t0 = time.time()
nB = output.data.size(0) # batch size
nA = self.num_anchors
nC = self.num_classes
nH = output.data.size(2)
nW = output.data.size(3)
cls_anchor_dim = nB*nA*nH*nW
if not isinstance(self.anchors, torch.Tensor):
self.anchors = torch.FloatTensor(self.anchors).view(self.num_anchors, self.anchor_step).to(self.device)
output = output.view(nB, nA, (5+nC), nH, nW).to(self.device)
cls_grid = torch.linspace(5,5+nC-1,nC).long().to(self.device)
ix = torch.LongTensor(range(0,5)).to(self.device)
pred_boxes = torch.FloatTensor(4, cls_anchor_dim).to(self.device)
coord = output.index_select(2, ix[0:4]).view(nB*nA, -1, nH*nW).transpose(0,1).contiguous().view(-1,cls_anchor_dim) # x, y, w, h
coord[0:2] = coord[0:2].sigmoid()
conf = output.index_select(2, ix[4]).view(cls_anchor_dim).sigmoid()
cls = output.index_select(2, cls_grid)
cls = cls.view(nB*nA, nC, nH*nW).transpose(1,2).contiguous().view(cls_anchor_dim, nC)
t1 = time.time()
grid_x = torch.linspace(0, nW-1, nW).repeat(nB*nA, nH, 1).view(cls_anchor_dim).to(self.device)
grid_y = torch.linspace(0, nH-1, nH).repeat(nW,1).t().repeat(nB*nA, 1, 1).view(cls_anchor_dim).to(self.device)
anchor_w = self.anchors.index_select(1, ix[0]).repeat(nB, nH*nW).view(cls_anchor_dim)
anchor_h = self.anchors.index_select(1, ix[1]).repeat(nB, nH*nW).view(cls_anchor_dim)
pred_boxes[0] = coord[0] + grid_x
pred_boxes[1] = coord[1] + grid_y
pred_boxes[2] = coord[2].exp() * anchor_w
pred_boxes[3] = coord[3].exp() * anchor_h
# for build_targets. it works faster on CPU than on GPU
pred_boxes = convert2cpu(pred_boxes.transpose(0,1).contiguous().view(-1,4)).detach()
t2 = time.time()
nGT, nRecall, obj_mask, noobj_mask, coord_mask, tcoord, tconf, tcls = \
self.build_targets(pred_boxes, target.detach(), nH, nW)
cls_mask = (obj_mask == 1)
tcls = tcls[cls_mask].long().view(-1).to(self.device)
cls_mask = cls_mask.view(-1, 1).repeat(1,nC).to(self.device)
cls = cls[cls_mask].view(-1, nC)
nProposals = int((conf > 0.25).sum())
tcoord = tcoord.view(4, cls_anchor_dim).to(self.device)
tconf = tconf.view(cls_anchor_dim).to(self.device)
conf_mask = (self.object_scale * obj_mask + self.noobject_scale * noobj_mask).view(cls_anchor_dim).to(self.device)
obj_mask = obj_mask.view(cls_anchor_dim).to(self.device)
coord_mask = coord_mask.view(cls_anchor_dim).to(self.device)
t3 = time.time()
loss_coord = self.coord_scale * nn.MSELoss(reduction='sum')(coord*coord_mask, tcoord*coord_mask)/nB
loss_conf = nn.MSELoss(reduction='sum')(conf*conf_mask, tconf*conf_mask)/nB
loss_cls = self.class_scale * nn.CrossEntropyLoss(reduction='sum')(cls, tcls)/nB
loss = loss_coord + loss_conf + loss_cls
t4 = time.time()
if False:
print('-'*30)
print(' activation : %f' % (t1 - t0))
print(' create pred_boxes : %f' % (t2 - t1))
print(' build targets : %f' % (t3 - t2))
print(' create loss : %f' % (t4 - t3))
print(' total : %f' % (t4 - t0))
print('%d: nGT %3d, nRC %3d, nPP %3d, loss: box %6.3f, conf %6.3f, class %6.3f, total %7.3f'
% (self.seen, nGT, nRecall, nProposals, loss_coord, loss_conf, loss_cls, loss))
if math.isnan(loss.item()):
print(conf, tconf)
sys.exit(0)
return loss