wrong loss penalize？

[ideask]

i.e:

200: 姿态(pose) 0

201: 表情(expression) 0

202: 照度(illumination) 0

203: 化妆(make-up) 0

204: 遮挡(occlusion) 0

205: 模糊(blur) 0

the cleaner sample, i.e. without attributes, the loss = 0 ？？？ why???

[ucohen]

it seems like the loss function does not match this dataset, as there are numerous items with no attributes - causing zero effect on the loss. Any suggestions?