Open gilsm opened 4 months ago
你设置了 reduce_zero_label=True, 在推理的时候。他还会把这个0类加回来的,所以如果原来标签中 1类 是1,及时reduce_zero_label=True,最后推理出来的结果还会是1. 所以你应该在可视化的时候,把背景类,黑色加上,才能正确的可视化。
What you are saying is what I thought should happen but it diesn't . I till get wrong classifications
I am trying to use the SAN model I started with san-vit-l14_coco-stuff164k-640x640.py I have a set of my own with 9 classes in which 1st class '0' is black_pixels which I want to ignore training on. I tried learning with setting class '0' as 'black_pixels' or 'background' In any case there is always a miss by 1 with the class label in the images saved for showing . meaning class 2 for ex is displayed with color of class 1 with its wrong label. The clases in the out directory are numbered correctly
I have set: The classes # to 9 reduce_zero_label=True I removed class_weight because it required me to put 10 weights for it to run which I did not understand