yolo loss 将图像标注的真实事坐标转换到anchor坐标

最近在看yolov3 的源码，在看yolo_loss的时候遇到了一个卡点，就是将真是标注的box终点坐标转换

到anchor点的坐标

        true_xy = true_xy * tf.cast(grid_size, tf.float32) - tf.cast(grid, tf.float32)

        raw_true_xy = y_true[l][..., :2] * grid_shapes[l][:] - grid

import tensorflow as tf


import numpy as np



#anchor box=13*13
grid_size=13
grid = tf.meshgrid(tf.range(grid_size), tf.range(grid_size))
grid = tf.expand_dims(tf.stack(grid, axis=-1), axis=2)


#batch_size=8 box=13*13 ,每一种规格的anchor box 对应3个box ,中点坐标是２维
T_xy=np.zeros([8, 13, 13, 3, 2])*1.0

T_xy[6,4,2]=[0.309,0.46212122]


T_xy=tf.constant(T_xy,dtype=tf.float32)


true_xy = T_xy * tf.cast(grid_size, tf.float32) - tf.cast(grid, tf.float32)

print(true_xy)

[[[[  0.   0.]
    [  0.   0.]
    [  0.   0.]]

   [[ -1.   0.]
    [ -1.   0.]
    [ -1.   0.]]

   [[ -2.   0.]
    [ -2.   0.]
    [ -2.   0.]]

   ...


   ...

   [[-10. -12.]
    [-10. -12.]
    [-10. -12.]]

   [[-11. -12.]
    [-11. -12.]
    [-11. -12.]]

   [[-12. -12.]
    [-12. -12.]
    [-12. -12.]]]]], shape=(8, 13, 13, 3, 2), dtype=float32)

        xy_loss = obj_mask * box_loss_scale * \
            tf.reduce_sum(tf.square(true_xy - pred_xy), axis=-1)

obj_mask=np.zeros([8, 13, 13, 3])

obj_mask[0,6,4,2]=1

obj_mask=tf.constant(obj_mask,dtype=tf.float32)

发现这样处理后计算loss，有大量常量数值１，…11,12等，感觉会有问题因为用了tf.reduce_sum(tf.square(true_xy - pred_xy)

实际上没关系，因为在前面乘以了obj_mask ,对于yolo_loss中，中点坐标的loss ，只会计算有标记的对应anchor 的loss其余点loss会设置为0