我正在实现一个自我定义的损失YOLO网络。
假设有两个张量,GT和PD(地面真值和预测),都是4x4的2维矩阵。
假设GT为:

0,0,0,0
0,1,0,0
0,0,1,0
0,0,0,0

PD的大小与一些随机数相同。
这里我需要分别计算均方误差。
计算GT中有1的MSE和计算GT中有0的MSE。
我更喜欢使用一个掩码来覆盖不相关的元素,所以计算时只计算相关的元素。我已经在numpy中实现了这个,但是不知道如何使用tf(v1.14)实现它
import numpy as np
import numpy.ma as ma
conf = y_true[...,0]
conf = np.expand_dims(conf,-1)

conf_pred = y_pred[...,0]
conf_pred = np.expand_dims(conf_pred,-1)

noobj_conf = ma.masked_equal(conf,1)   #cover grid with objects
obj_conf = ma.masked_equal(conf,0)     #cover grid without objects

loss_obj = np.sum(np.square(obj_conf - conf_pred))
loss_noobj = np.sum(np.square(noobj_conf - conf_pred))

关于如何在tensorflow中实现这一点有什么建议吗?

最佳答案

如果我理解正确,你需要分别计算0和1的均方误差。
您可以执行以下操作:

y_true = tf.constant([[0,0,0,0],[0,1,0,0],[0,0,1,0],[0,0,0,0]], dtype=tf.float32)
y_pred = tf.random.uniform([4, 4], minval=0, maxval=1)

# find indices where 0 is present in y_true
indices0 = tf.where(tf.equal(y_true, tf.zeros([1.])))
# find indices where 1 is present in y_true
indices1 = tf.where(tf.equal(y_true, tf.ones([1.])))

# find all values in y_pred which are present at indices0
y_pred_indices0 = tf.gather_nd(y_pred, indices0)
# find all values in y_pred which are present at indices1
y_pred_indices1 = tf.gather_nd(y_pred, indices1)

# mse loss calculations
mse0 = tf.losses.mean_squared_error(labels=tf.gather_nd(y_true, indices0), predictions=y_pred_indices0)
mse1 = tf.losses.mean_squared_error(labels=tf.gather_nd(y_true, indices1), predictions=y_pred_indices1)

# mse0 = tf.reduce_sum(tf.squared_difference(tf.gather_nd(y_true, indices0), y_pred_indices0))
# mse1 = tf.reduce_sum(tf.squared_difference(tf.gather_nd(y_true, indices1), y_pred_indices1))

with tf.Session() as sess:
    y_, loss0, loss1 = sess.run([y_pred, mse0, mse1])
    print(y_)
    print(loss0, loss1)

输出:
[[0.12770343 0.43467927 0.9362457  0.09105921]
 [0.46243036 0.8838414  0.92655015 0.9347118 ]
 [0.14018488 0.14527774 0.8395766  0.14391887]
 [0.1209656  0.7793218  0.70543754 0.749542  ]]
0.341359 0.019614244

关于python - 如何在Tensorflow中用蒙板减去两个张量?,我们在Stack Overflow上找到一个类似的问题:https://stackoverflow.com/questions/57240208/

10-12 21:15