-
Notifications
You must be signed in to change notification settings - Fork 53
/
Loss.py
82 lines (72 loc) · 4.19 KB
/
Loss.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
# -*- coding: utf-8 -*-
# --------------------------------------
# @Time : 2018/5/12$ 17:49$
# @Author : KOD Chen
# @Email : 821237536@qq.com
# @File : Loss$.py
# Description :Yolo_v2 Loss损失函数.
# --------------------------------------
import numpy as np
import tensorflow as tf
def compute_loss(predictions,targets,anchors,scales,num_classes=20,output_size=(13,13)):
W,H = output_size
C = num_classes
B = len(anchors)
anchors = tf.constant(anchors,dtype=tf.float32)
anchors = tf.reshape(anchors,[1,1,B,2]) # 存放输入的anchors的wh
# 【1】ground truth：期望值、真实值
sprob,sconf,snoob,scoor = scales # loss不同部分的前面系数
_coords = targets["coords"] # ground truth [-1, H*W, B, 4]，真实坐标xywh
_probs = targets["probs"] # class probability [-1, H*W, B, C] ，类别概率——one hot形式，C维
_confs = targets["confs"] # 1 for object, 0 for background, [-1, H*W, B]，置信度，每个边界框一个
# ground truth计算IOU-->_up_left, _down_right
_wh = tf.pow(_coords[:, :, :, 2:4], 2) * np.reshape([W, H], [1, 1, 1, 2])
_areas = _wh[:, :, :, 0] * _wh[:, :, :, 1]
_centers = _coords[:, :, :, 0:2]
_up_left, _down_right = _centers - (_wh * 0.5), _centers + (_wh * 0.5)
# ground truth汇总
truths = tf.concat([_coords, tf.expand_dims(_confs, -1), _probs], 3)
# 【2】decode the net prediction：预测值、网络输出值
predictions = tf.reshape(predictions,[-1,H,W,B,(5+C)])
# t_x, t_y, t_w, t_h
coords = tf.reshape(predictions[:,:,:,:,0:4],[-1,H*W,B,4])
coords_xy = tf.nn.sigmoid(coords[:,:,:,0:2]) # 0-1，xy是相对于cell左上角的偏移量
coords_wh = tf.sqrt(tf.exp(coords[:,:,:,2:4])*anchors/np.reshape([W,H],[1,1,1,2])) # 0-1，除以特征图的尺寸13，解码成相对于整张图片的wh
coords = tf.concat([coords_xy,coords_wh],axis=3) # [batch_size, H*W, B, 4]
# 置信度
confs = tf.nn.sigmoid(predictions[:,:,:,:,4])
confs = tf.reshape(confs,[-1,H*W,B,1]) # 每个边界框一个置信度，每个cell有B个边界框
# 类别概率
probs = tf.nn.softmax(predictions[:,:,:,:,5:]) # 网络最后输出是"得分"，通过softmax变成概率
probs = tf.reshape(probs,[-1,H*W,B,C])
# prediction汇总
preds = tf.concat([coords,confs,probs],axis=3) # [-1, H*W, B, (4+1+C)]
# prediction计算IOU-->up_left, down_right
wh = tf.pow(coords[:, :, :, 2:4], 2) * np.reshape([W, H], [1, 1, 1, 2])
areas = wh[:, :, :, 0] * wh[:, :, :, 1]
centers = coords[:, :, :, 0:2]
up_left, down_right = centers - (wh * 0.5), centers + (wh * 0.5)
# ※※※【3】计算ground truth和anchor的IOU：※※※
# 计算IOU只考虑形状，先将anchor与ground truth的中心点都偏移到同一位置（cell左上角），然后计算出对应的IOU值。
# ①IOU值最大的那个anchor与ground truth匹配，对应的预测框用来预测这个ground truth：计算xywh、置信度c(目标值为1)、类别概率p误差。
# ②IOU小于某阈值的anchor对应的预测框：只计算置信度c(目标值为0)误差。
# ③剩下IOU大于某阈值但不是max的anchor对应的预测框：丢弃，不计算任何误差。
inter_upleft = tf.maximum(up_left, _up_left)
inter_downright = tf.minimum(down_right, _down_right)
inter_wh = tf.maximum(inter_downright - inter_upleft, 0.0)
intersects = inter_wh[:, :, :, 0] * inter_wh[:, :, :, 1]
ious = tf.truediv(intersects, areas + _areas - intersects)
best_iou_mask = tf.equal(ious, tf.reduce_max(ious, axis=2, keep_dims=True))
best_iou_mask = tf.cast(best_iou_mask, tf.float32)
mask = best_iou_mask * _confs # [-1, H*W, B]
mask = tf.expand_dims(mask, -1) # [-1, H*W, B, 1]
# 【4】计算各项损失所占的比例权重weight
confs_w = snoob * (1 - mask) + sconf * mask
coords_w = scoor * mask
probs_w = sprob * mask
weights = tf.concat([coords_w, confs_w, probs_w], axis=3)
# 【5】计算loss：ground truth汇总和prediction汇总均方差损失函数，再乘以相应的比例权重
loss = tf.pow(preds - truths, 2) * weights
loss = tf.reduce_sum(loss, axis=[1, 2, 3])
loss = 0.5 * tf.reduce_mean(loss)
return loss