关亚勇

thanks for your answer. yeah, the label is in the format N x 20, and each line value is [0, 1, 0, 1, ..., 1, 0], length is 20, where...

 I use the SigmoidCrossEntropyLoss, so that is why i doubt！i did not use your create_data_lmdb.py, instead of caffe own convert_imageset, is that influences？？

thanks, i will try it again!

i got why, the accuracy style is wrong  you konw how to test the accuracy when train like above?

@zengarden, yeah, how to apply FPN to light head rcnn ?