lipond

I use the default parameters set in original code, and got the following training process. The "train_decode" and "val_decode" both are null till "epoch:0, batch:1700", but the loss is still...

Thanks.

Meet some problems when training on my own data. See below.

3

When the negative proposals is 0, the training ends up with the following complaints. Anyone met similar problems? ``` rpn: num_positive 128 rpn: num_negative 0 rpn: num_positive avg 128 rpn:...

test accuracy so low

1

Hi, After running all 500000 steps, the test accuracy is only 0.4375, is it right? 

How about the similar idea applied to object detection task?

1

I want to know it. Thanks a lot!