Modification of weights for unbalanced dataset

[Eavis]

[Eavis]

I have modified the ADE20K scene parsing benchmark produced by MIT. The original dataset contains 150 categoried(151 include background) and set background=0 to be ignored.
I have chosen 17 categories and set the previous background(label as 0) to 255, and set the other categories as 0(background). Thus ignore_class = 255

I count the pixels in the 20210 images after the modification. <category idx, number of pixel counts>
0,1637454744
1,737138625
2,503599787
3,414602893
4,290185267
5,211497624
6,93447406
7,84733791
8,78265150
9,74475972
10,70227430
11,55492152
12,28967419
13,28361240
14,24627334
15,9994821
16,7816482
17,3719357
255,411637498

The background is 0.6 of all the other categories in total(1-17)
However, the background is 440 times the 17th category. (1637454744 / 3719357= 440.3)
(16th: 1637454744 / 7816482 = 209.5)
Does anyone have any suggestion on how to set the weight for these categories?
Should I adjust the weight for only a few categories or all of the 17 categories?

weights = tf.to_float(tf.equal(scaled_labels, 0)) * label0_weight + tf.to_float(tf.equal(scaled_labels, 1)) * label1_weight + tf.to_float(tf.equal(scaled_labels, ignore_label)) * 0.0
    where you need to tune the label0_weight and label1_weight (e.g., set label0_weight=1 and increase label1_weight).

Thanks!