positive_loss is much larger than negative_loss

[heiyuxiaokai]

Train log: 2019-10-11 16:29:21,673 maskrcnn_benchmark.trainer INFO: eta: 1 day, 0:40:30 iter: 260 loss: 3.8217 (3.8392) negative_loss: 0.0326 (0.0354) positive_loss: 3.7731 (3.8038) time: 1.3759 (1.4870) data: 0.0050 (0.0068) lr: 0.000680 max mem: 7173 2019-10-11 16:29:48,522 maskrcnn_benchmark.trainer INFO: eta: 1 day, 0:29:44 iter: 280 loss: 3.7343 (3.8322) negative_loss: 0.0513 (0.0364) positive_loss: 3.6920 (3.7958) time: 1.2558 (1.4766) data: 0.0049 (0.0067) lr: 0.000707 max mem: 7173 2019-10-11 16:30:17,056 maskrcnn_benchmark.trainer INFO: eta: 1 day, 0:25:56 iter: 300 loss: 3.5909 (3.8169) negative_loss: 0.0517 (0.0395) positive_loss: 3.5172 (3.7775) time: 1.1965 (1.4733) data: 0.0047 (0.0066) lr: 0.000733 max mem: 7173_ It's normal?

[zhangxiaosong18]

RetinaNet and FreeAnchor initialize the classifier bias to make it predict lower scores, so the negative loss is very small.