-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset.py
182 lines (132 loc) · 7.27 KB
/
dataset.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
import tensorflow as tf
from packaging import version
if version.parse(tf.__version__) < version.parse("2.6"):
AUTOTUNE = tf.data.experimental.AUTOTUNE
else:
AUTOTUNE = tf.data.AUTOTUNE
import tensorflow_datasets as tfds
import numpy as np
from utils import *
import cv2
from opts import opts
import math
import functools
opt = opts()
def preprocess(img, points, pt_labels,aug):
height, width = img.shape[:2]
c = np.array([width / 2., height / 2.], dtype=np.float32)
s = max(height, width) * 1.0
input_h, input_w = opt.net_height, opt.net_width # 网络输入的宽高
trans_input = get_affine_transform(c, s, 0, [input_w, input_h])
# inp = cv2.warpAffine(np.array(img), trans_input, (input_w, input_h), flags=cv2.INTER_LINEAR)
inp = cv2.warpAffine(img, trans_input, (input_w, input_h), flags=cv2.INTER_LINEAR)
inp = (inp.astype(np.float32) / 255.)
if aug:
_data_rng = np.random.RandomState(123)
_eig_val = np.array([0.2141788, 0.01817699, 0.00341571], dtype=np.float32)
_eig_vec = np.array([
[-0.58752847, -0.69563484, 0.41340352],
[-0.5832747, 0.00994535, -0.81221408],
[-0.56089297, 0.71832671, 0.41158938]
], dtype=np.float32)
# 数据增强
color_aug(_data_rng, inp, _eig_val, _eig_vec)
# float32 -> float64
inp = (inp - np.array([0.408, 0.447, 0.47])) / np.array([0.289, 0.274, 0.278])
inp = inp.astype(np.float32)
# '为输出做仿射变换做准备'
output_h = input_h // opt.down_ratio # heatmap的高 128
output_w = input_w // opt.down_ratio # heatmap的宽 128
num_classes = opt.num_classes # 类别数
trans_output = get_affine_transform(c, s, 0, [output_w, output_h])
# 定义所需张量
hm = np.zeros((output_h, output_w, num_classes), dtype=np.float32) # (128, 128, cls_num) heatmap
embedding = np.zeros((opt.max_objs, 2, 2), dtype=np.float32) # 点对的embedding
# embedding = np.zeros((opt.max_objs // 2, 2, 2), dtype=np.float32) # 点对的embedding
reg = np.zeros((opt.max_objs, 2), dtype=np.float32) # 点的误差offset
ind = np.zeros((opt.max_objs), dtype=np.int64)
reg_mask = np.zeros((opt.max_objs), dtype=np.uint8)
draw_gaussian = draw_msra_gaussian if opt.mse_loss else draw_umich_gaussian
# '准备数据集返回变量,主要是真值标签,如热力图、embeddings 和 中心点偏移量
num_points = points.shape[0]
step = int(num_points / 2)
max_distance = 0
# obj_index = np.arange(num_points).reshape(2, int(num_points / 2)).T
for i in range(step):
dis_y = abs(points[i][1] - points[i + step][1])
if dis_y > max_distance:
max_distance = dis_y
for k in range(step): # '按实际有多少个目标对 进行循环'
point = points[k] # gt point: (x,y)
point_em = points[k + step]
cls_id = pt_labels[k] # gt cls id
cls_id_em = pt_labels[k + step] # gt cls id
# '输出的仿射变换,这些仿射变换由相应的函数完成,在实际编写中只需确定诸如中心点c、长边长度s等参数,便于在自己的数据集中使用
# 由于没有gt的w,h 只需要对gt point的x,y进行变换
point = affine_transform(point, trans_output) # 中心点仿射变换
point[0] = np.clip(point[0], 0, output_w - 1)
point[1] = np.clip(point[1], 0, output_h - 1)
point_em = affine_transform(point_em, trans_output) # 中心点仿射变换
point_em[0] = np.clip(point_em[0], 0, output_w - 1)
point_em[1] = np.clip(point_em[1], 0, output_h - 1)
h, w = int(max_distance / 4), int(max_distance / 4) # 决定高斯核的宽高
if h > 0 and w > 0:
radius = gaussian_radius((math.ceil(h), math.ceil(w)))
radius = max(0, int(radius))
# radius = [opt.]hm_gauss if opt.mse_loss else radius
ct = np.array(point, dtype=np.float32) # 中心点
ct_int = ct.astype(np.int32)
ct_em = np.array(point_em, dtype=np.float32) # 中心点
ct_int_em = ct_em.astype(np.int32)
draw_gaussian(hm[..., cls_id], ct_int, radius) # 绘制热力图,绘制在其所属类别的通道上
draw_gaussian(hm[..., cls_id_em], ct_int_em, radius) # 绘制热力图,绘制在其所属类别的通道上
# k对 2(类) 2(idx, 1)
# embedding[k][cls_id], embedding[k][cls_id_em] = (ct_int[1] * output_w + ct_int[0], 1), (
# (output_h * output_w) + ct_int_em[1] * output_w + ct_int_em[0], 1)
embedding[k][cls_id], embedding[k][cls_id_em] = (ct_int[1] * output_w * 2 + ct_int[0] * 2 + 0, 1), \
(ct_int_em[1] * output_w * 2 + ct_int_em[0] * 2 + 1, 1)
# print(ct_int[1], ct_int[0])
# print(ct_int_em[1], ct_int_em[0])
# print("______")
# print(ct[1] * output_h + ct[0], 1), ((output_h * output_w) + ct_em[1] * output_h + ct_em[0], 1)
ind[k] = ct_int[1] * output_w + ct_int[0] # 中心点的位置,用一维表示h*W+w
ind[k + step] = ct_int_em[1] * output_w + ct_int_em[0] # 中心点的位置,用一维表示h*W+w
reg[k] = ct - ct_int # 由取整引起的误差 中心点误差offset的gt
reg[k + step] = ct_em - ct_int_em
reg_mask[k] = 1 # 设为1, 表示该位置有目标存在
reg_mask[k + step] = 1
# if opt.dense_wh:
# draw_dense_reg(dense_wh, hm.max(axis=0), ct_int, embedding[k], radius)
# ret = {'input': inp, 'hm': hm, 'reg_mask': reg_mask, 'reg': reg, 'ind': ind, 'embedding': embedding}
return inp, hm, reg_mask, reg, ind, embedding
def preprocess_function(data,aug):
inp, hm, reg_mask, reg, ind, embedding = tf.numpy_function(func=functools.partial(preprocess,aug=aug), inp=[data['image'], data['points']['point'], data['points']['category']],
Tout=[tf.float32, tf.float32, tf.uint8, tf.float32, tf.int64, tf.float32])
# return inp, (hm, reg_mask, reg, ind, embedding)
return inp, {"hm":hm, "reg_mask":reg_mask, "reg": reg, "ind":ind, "embedding": embedding}
def prepare(ds, batch_size, aug=True, shuffle=False):
if shuffle:
ds = ds.shuffle(1000)
ds = ds.map(functools.partial(preprocess_function,aug=aug), num_parallel_calls=AUTOTUNE)
# Batch all datasets.
ds = ds.batch(batch_size)
# Use buffered prefetching on all datasets.
return ds.prefetch(buffer_size=AUTOTUNE)
if __name__ == "__main__":
# ds, info = tfds.load("foot_robot", with_info=True)
# print(tfds.benchmark(ds))
# print(info.splits['train'].num_examples)
# data = next(iter(ds['train']))
# preprocess(data['image'], data['points']['point'], data['points']['category'])
ds, info = tfds.load("foot_robot", with_info=True)
ds = ds['train'].take(100)
for data in ds:
preprocess(data['image'].numpy(), data['points']['point'].numpy(), data['points']['category'].numpy(), aug=True)
# # data = next(iter(ds))
# # preprocess(data['image'], data['points']['point'], data['points']['category'], aug=True)
#
# ds = prepare(ds['train'], 8, shuffle=True)
# print(tfds.benchmark(ds, batch_size=8))
# data = next(iter(ds))
#
# print("asdf")