-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtest.py
214 lines (212 loc) · 9.68 KB
/
test.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
import torch
import numpy as np
import os
from utils.utils import nms,poly_iou,xyxya2corner,xywh2xyxy,decode_yolox_boxes
from tqdm import tqdm
import json
#计算Frame-mAP
def test(net,
dataloader,
class_names,
epoch=0,
critical_iou=0.5,
conf_thres=0.05,
nms_thres=0.4):
bounding_boxes = []
devices = torch.device("cuda" if torch.cuda.is_available() else "cpu")
net = net.to(devices)
sep = dataloader.dataset.patch
model_image_size = dataloader.dataset.shape
for iteration,batch in enumerate(tqdm(dataloader)):
idx, images,image_shapes = batch[0], batch[1],batch[2]
with torch.no_grad():
images = torch.from_numpy(images).float().to(devices)
# labels = [torch.from_numpy(ann).float() for ann in labels]
outputs = net(images)
# loss_item = compute_loss(outputs, labels, np.reshape(anchors, [-1, 2]), len(class_names),label_smooth=0)
# val_loss+=loss_item[0].item()
output = decode_yolox_boxes(outputs,input_shape=model_image_size)
#相对于608x608
# [bs*num_img,anchors,17]
output = xywh2xyxy(output)
# [bs,num_img,anchors,17]
output = output.reshape(-1,sep*sep,output.shape[1],output.shape[2])
bscale =[]
boffset = []
for i in range(output.shape[0]):
stride_w = image_shapes[i][0] // sep
stride_h = image_shapes[i][1] // sep
_scalew = stride_w / model_image_size[0]
_scaleh = stride_h / model_image_size[1]
offset = [[j%sep*stride_w,j//sep*stride_h]*2 for j in range(sep*sep)]
bscale.append([_scalew, _scaleh] * 2)
boffset.append(offset)
bscale = torch.tensor(bscale).unsqueeze(1).unsqueeze(1).to(devices)
boffset = torch.tensor(boffset).unsqueeze(2).to(devices)
output[...,:4] = output[...,:4]*bscale+boffset
#[bs,num_img*anchors,17]
output = output.reshape(output.shape[0],-1,output.shape[3])
# 非极大值抑制
batch_detections = nms(output,conf_thres=conf_thres,nms_thres=nms_thres,only_objection=False,nms_link_classes=True,fast=False)
# batch_detections = non_max_suppression(output,conf_thres=conf_thres,nms_thres=nms_thres,only_objection=False)
for i, o in enumerate(batch_detections):
#存在预测结果就保存，否则continue
if o is None:
continue
o = o.data.cpu()
top_conf = np.array(o[:, 5])
top_label = np.array(o[:, -1], np.int32)
top_bboxes = np.array(o[:, :5])
#截断，取整
# top_bboxes = clip_coords(top_bboxes, image_shape)
#boxes=np.round(boxes).astype('int32')
for c, l, b in zip(top_conf, top_label, top_bboxes):
b = np.around(b, decimals=4)
bounding_boxes.append({
"conf": "%.6f" % c,
"class": int(l),
"bbox": b.tolist(),
"gt_path":idx[i]
})
print("all boxes catched...")
# 将所有结果按类别存成.json文件格式
# [{"confident":0-1,"gt_box":坐标,"truthbox":txt_path,"class":0 or 1 or 2},...]
current_dir = 'logs/detections/detections_' + str(epoch)
if not os.path.exists('logs'):
os.mkdir('logs')
if not os.path.exists('logs/detections'):
os.mkdir('logs/detections')
if not os.path.exists(current_dir):
os.mkdir(current_dir)
for c in range(len(class_names)):
#筛选出其中一类
bounding_box = list(filter(lambda x: int(x["class"]) == c, bounding_boxes))
if len(bounding_box) == 0:
#如果未检测到该类，就从列表中剔除
# count_class.remove(c)
continue
#按置信度排序，从大到小
bounding_box.sort(key=lambda x: float(x['conf']), reverse=True)
with open(current_dir +"/" +class_names[c] + "_dr.json", 'w') as outfile:
json.dump(bounding_box, outfile,indent=1)
print("all classes separated...")
#所有真实框个数初始化
label_dict={}
t_class_list=[]
#获得真实框
lines = dataloader.dataset.lines
for line in lines:
li = line.rstrip().split()
gt_box = np.array([np.array(list(map(float, box.split(',')))) for box in li[1:]])
#特别注意
# gt_box[:, 0] = gt_box[:, 0] - 1
cur_class = list(map(int,gt_box[:,0].tolist()))
t_class_list.extend(cur_class)
label_dict[li[0]]=gt_box.tolist()
#所有真实框个数
truth_num_box = len(t_class_list)
#所有预测框个数
pre_num_box = len(bounding_boxes)
print("num_truthbox: ",truth_num_box)
print("num_prebox: ",pre_num_box)
#真实框类别
count_class = np.unique(t_class_list)
#真阳性样本初始化
tp = 0
del bounding_boxes
#计算ap
Maplist = []
for c in count_class:
t_box_length = t_class_list.count(c)
gt_class_path = current_dir +'/'+ class_names[c] + "_dr.json"
if not os.path.exists(gt_class_path):
Maplist.append(0.0)
continue
#打开刚刚保存的json,循环取出每一个框
bounding_box = json.load(open(gt_class_path))
#初始化真阳性序列
pred_match = np.zeros(len(bounding_box))
for i, obj in enumerate(bounding_box):
#取出一个预测框
pre_box = np.array(obj["bbox"])
#读取真实框
gt_box = np.array(label_dict[obj["gt_path"]])
if len(gt_box) == 0:
continue
#------------------待更新-------------------------#
# gt_box_t = torch.from_numpy(gt_box[:, 1:5]).float()
# pre_box = pre_box.expand_as(gt_box_t)
# overlaps = bbox_iou(pre_box, gt_box_t).numpy()
#xywha2corner转换为角点
#poly_iou计算overlaps
#计算预测框与所有真实框iou
bs = gt_box.shape[0]
gt_box_t = xyxya2corner(gt_box[:,1:])
pre_box = xyxya2corner(np.expand_dims(pre_box,0))
overlaps = np.zeros(bs)
for k in range(bs):
overlaps[k] = poly_iou(gt_box_t[k],pre_box[0])
#------------------------------------------------#
#iou按从大到小排序
sorted_ixs = np.argsort(-overlaps)
for s in sorted_ixs:
#依次判断iou是否大于0.5，小于则说明是假阳性样本，直接退出
if overlaps[s] < critical_iou:
break
#大于就判断预测类别和真实框类别是否一致，一致则真阳性，将序列位置置一，直接退出循环
if obj["class"] == int(gt_box[s, 0]):
tp += 1
pred_match[i] = 1
#匹配到便去除真实框
label_dict[obj["gt_path"]].pop(s)
break
#累加
precisions = np.cumsum(pred_match) / (np.arange(len(pred_match)) + 1)
#此时召回率逐渐上升
recalls = np.cumsum(pred_match).astype(np.float32) / t_box_length
# Pad with start and end values to simplify the math
precisions = np.concatenate([[0], precisions, [0]])
recalls = np.concatenate([[0], recalls, [1]])
# Ensure precision values decrease but don't increase. This way, the
# precision value at each recall threshold is the maximum it can be
# for all following recall thresholds, as specified by the VOC paper.
#保证准确率取每个召回率最大值
for i in range(len(precisions) - 2, -1, -1):
precisions[i] = np.maximum(precisions[i], precisions[i + 1])
# Compute mean AP over recall range
#取出召回率发生变化的点
#为了求面积
indices = np.where(recalls[:-1] != recalls[1:])[0] + 1
ap = np.sum((recalls[indices] - recalls[indices - 1]) * precisions[indices])
Maplist.append(ap)
aps={class_names[c]:round(Maplist[i],3) for i, c in enumerate(count_class)}
mAP = sum(Maplist) / len(Maplist)
recall = tp / truth_num_box
precision = tp / pre_num_box if pre_num_box != 0 else 0.0
return aps, mAP, recall, precision
if __name__ == "__main__":
from utils.dataset import listDataset, dataset_collate_val
from torch.utils.data import DataLoader
from utils.utils import get_classes, get_anchors
from yolox.yolox import YoloX
import torch.nn as nn
model_image_size = (640,640)
batch_size = 1
classes_path = 'model_data/yolo_classes.txt'
model_path = 'logs/yolox_s_0.pt'
phi = 's'
#加载类别,先验框,模型,多GPU
class_names = get_classes(classes_path)
test_dataset = listDataset('dataset/vallist.txt',patch=4,shape = model_image_size,train=False)
test_dataloader = DataLoader(test_dataset,batch_size=batch_size,
num_workers=2,pin_memory=True,
drop_last=False,collate_fn=dataset_collate_val)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
net = nn.DataParallel(YoloX(len(class_names), phi=phi))
net.load_state_dict(torch.load(model_path, map_location=device)["model"])
net.to(device).eval()
print("load model done!")
APs, mAP, recall, precision = test(net,test_dataloader,class_names,epoch=0)
print("each class ap:")
print(str(APs)[1:-1])
print("recall:{:.3f} precision:{:.3f} mAP:{:.3f}".format(recall, precision, mAP))