About Loss Function

[ssdutHB]

Hi, everyone. I read the code in data_parser.py, found that there is an option for input labels, that is "target regression". If we choose this option, the loaded ground truth will be a matrix consisting of real numbers from 0 to 1 rather than a binary matrix. After that, I checked losses.py and found that there are two lines of codes:
"count_neg = tf.reduce_sum(1. - y)"
"count_pos = tf.reduce_sum(y)"

there two lines of codes seems work well for a binary ground truth, but if they work for a ground truth that consists of real numbers from 0 to 1? I am looking forward to your answers.

[CangHaiQingYue]

I had the same question as you. I used

y_0 = tf.zeros_like(y)
beta = tf.reduce_mean(tf.cast(tf.equal(y, y_0), tf.float32))

to solve.
But during evalution,
"count_neg = tf.reduce_sum(1. - y)" "count_pos = tf.reduce_sum(y)"
works better, I dont know why