gaussclb comments

Results 5 comments of


                                            gaussclb

Model detail: Why does "attention_heads" exist in "modeling.py"?

@jacobdevlin-google, why do you maintain the `attention_heads`?

The result on pascal_voc 2007 validation set is bad.

![test](https://user-images.githubusercontent.com/11674304/32947035-ef80b8a2-cbd4-11e7-85d5-1528a0f68276.png) ![test](https://user-images.githubusercontent.com/11674304/32947183-9027b044-cbd5-11e7-8f81-093ffda0229f.png) ![test](https://user-images.githubusercontent.com/11674304/32947101-3e2ad582-cbd5-11e7-8f25-e2be41bfa473.png)

The result on pascal_voc 2007 validation set is bad.

I'm sorry, I think it is too difficult to reproduce the MAP of yolo. Maybe the author use many tricks. And I'm also wonder whether the difference of `tensorflow` and...

help，thanks！

Because the test code only output the top_1 confidence of object, if you want to output many objects, you may set the threshold of confidence and use `NMS`.

mAP

Does anyone try to train yolov1 on pascal voc(2007+2012 trainval) and surpass mAP by 60% on 2007 test?