Why training starts with extremely high accuracy?

2019-05-17 13:18:50,651 - INFO - Epoch [1][50/32058]	lr: 0.00199, eta: 13 days, 10:51:24, time: 0.755, data_time: 0.018, memory: 5706, loss_rpn_cls: 0.4750, loss_rpn_reg: 0.0318, loss_cls: 0.1646, acc: 96.1250, loss_reg: 0.0419, loss_mask: 0.8765, loss: 1.5897
2019-05-17 13:19:29,251 - INFO - Epoch [1][100/32058]	lr: 0.00233, eta: 13 days, 14:24:12, time: 0.772, data_time: 0.015, memory: 5706, loss_rpn_cls: 0.1524, loss_rpn_reg: 0.0362, loss_cls: 0.1496, acc: 96.0703, loss_reg: 0.0888, loss_mask: 0.6807, loss: 1.1078

As you can see, the accuracy is almost 100% at the start. Why?

[panjianning]

Most are negative samples.

Maybe I have a misunderstanding.

But I think the ratio of positive and negative samples should be 50:50？

The MaxIoUAssigner will assigned bboxes as postive, negitive or ignored samples according to the iou setting. Then, the RandomSampler will sample some bboxes for training according to the assign result.

Num=256 means it will sample 256 samples, pos_fracion=0.5 means it will TRY to make pos:neg=1:1. But more often than not, there are not enough positive bboxes. (here, we need 128), so the actual ratio MAY BE something like 2:254.

[hellock]

Most are negative samples.