-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest.py
108 lines (94 loc) · 3.86 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
from yolov4 import yolov4_net
from tensorflow.keras import Input
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
import cv2 as cv
import tensorflow.keras.backend as K
import tensorflow as tf
#inputs = Input([416,416,3])
#model = yolov4_net(inputs,3,80)
#model.summary()
#for i,layer in enumerate(model.layers):
# print(i,layer.name)
grid_shape = tf.constant([13,13])
num_anchors = 3
print([K.arange(0,stop=grid_shape[0])])
print(tf.shape([K.arange(0,stop=grid_shape[0])] * grid_shape[1] * num_anchors))
#print(tf.shape(tf.reshape([K.arange(0,stop=grid_shape[0])] * grid_shape[1] * num_anchors,
# (num_anchors,grid_shape[0],grid_shape[1]))))
"""
grid_shape = tf.constant([13,13])
num_anchors = 3
print([K.arange(0,stop=grid_shape[0])])
print(tf.shape(tf.tile([K.arange(0,stop=grid_shape[0])],[grid_shape[1] * num_anchors,1])))
#print(tf.shape([K.arange(0,stop=grid_shape[0])] * grid_shape[1] * num_anchors))
print(tf.shape(tf.reshape(tf.tile([K.arange(0,stop=grid_shape[0])],[grid_shape[1] * num_anchors,1]),
(num_anchors,grid_shape[0],grid_shape[1]))))
"""
x_origin = tf.cast(tf.expand_dims(tf.transpose(tf.reshape(tf.tile(
[K.arange(0,stop=grid_shape[0])],[grid_shape[1] * num_anchors,1]),
(num_anchors,grid_shape[0],grid_shape[1])),(1,2,0)),axis=0),dtype=tf.float32)
y_origin = tf.transpose(x_origin,(0,2,1,3))
#print(x_origin)
#print(y_origin)
'''
def get_anchors(file_path):
with open(file_path) as f:
anchors = f.readline() # 只有一行
anchors = [float(x) for x in anchors.split(',')]
return np.array(anchors).reshape(-1,2)
line = './VOCdevkit/VOC2007/JPEGImages/000162.jpg 306,227,380,299,19 196,143,309,369,14'.split()
bboxes = np.array([np.array(list(map(int,box.split(',')))) for box in line[1:]])
#print(line)
print(bboxes.shape)
#print(bboxes[:,0:2][bboxes[:,0:2] > 0])
bboxes_w = bboxes[:,2] - bboxes[:,0]
bboxes_h = bboxes[:,3] - bboxes[:,1]
bboxes = bboxes[np.logical_and(bboxes_w > 416,bboxes_h < 416)]
#print(np.logical_and(bboxes_w > 416,bboxes_h < 416))
# 原理是这样的，首先bboxes_w和bboxes_h维度要一样，然后对每个一个单体进行逻辑与比较，比如
# bboxes_w[0] > 416 & bboxes_h[0] < 416,因为他有两个w和h，所以还有
# bboxes_w[0] > 416 & bboxes_h[0] < 416，就会有两个结果，[False,False]
# 然后就是与的语义，> 1的意思是所有bbox宽高都要大于1，不然就是废的
#print(bboxes)
def rand(a:float=0.0,b:float=1.0):
# np.random.rand()产生一个服从0,1均匀分布的值，# 这样写跟原来没区别
return np.random.rand() * (b - a) + a
w,h = 416,416
jitter = 0.3
#fig,ax = plt.subplots(3,3)
for i in range(9):
new_ar = w/h * rand(1-jitter,1+jitter)/rand(1-jitter,1+jitter)
scale = rand(.25, 2)
if new_ar < 1:
nh = int(scale*h)
nw = int(nh*new_ar)
#print(nh,nw)
else:
nw = int(scale*w)
nh = int(nw/new_ar)
#print(nh,nw)
dx = int(rand(0, w-nw))
dy = int(rand(0, h-nh))
#print(dx,dy)
image = Image.open('./VOCdevkit/VOC2007/JPEGImages/000162.jpg')
image = image.resize((nw, nh), Image.BICUBIC)
new_image = Image.new('RGB', (w, h), (128,128,128))
new_image.paste(image, (dx, dy))
#plt.figure()
#plt.imshow(new_image)
#plt.show()
image = Image.open('./VOCdevkit/VOC2007/JPEGImages/000162.jpg')
print(np.array(image).shape)
image_hsv = cv.cvtColor(np.array(image,np.float32) / 255.,cv.COLOR_RGB2HSV)
print(image_hsv.shape)
input_shape = np.array((416,416),dtype='float32')
grid_shapes = [input_shape // {0:32,1:16,2:8}[layer] for layer in range(3)]
print(grid_shapes)
import os
print(os.path.exists(path='./model_data/yolov4_weight.h5'))
'''
input_shape = tf.constant([512,512])
image_shape = tf.constant([720,1280])
print(tf.reduce_min(input_shape / image_shape))