Questions about efficientdet D0 network ouput

[minjung-mmc]

Hi, I am following efficientdet network algorithm, and I have questions about efficientdet D0 output size.
I got tensor size of [[1, 810, 64, 64], [1, 810, 32, 32], [1, 810, 16, 16], [1, 810, 8, 8], [1, 810, 4, 4]] for x_class output, and
[[1, 36, 64, 64], [1, 36, 32, 32], [1, 36, 16, 16], [1, 36, 8, 8], [1, 36, 4, 4]] for x_box output.
Can someone tell me the meaning of these outputs?
I am assuming the last two sizes are about number of grids, but I can't figure out what 810 and 36 means.
Thanks for the help.

[rwightman]

@minjung-mmc those outputs need post processing, which is done by the bench wrappers, as is those outputs are relative to the anchors

https://github.com/rwightman/efficientdet-pytorch/blob/master/effdet/bench.py#L20
https://github.com/rwightman/efficientdet-pytorch/blob/master/effdet/anchors.py

You should only use the models without the bench if you plan to write your own post-processing code that's more efficient, for embeded use, higher inference throughput, etc.

[minjung-mmc]

Thank you for your answer