Support in Float16

[davsol]

Hi, I've recently been thinking about ways to make SSD even faster, since mxnet is capable of running in float16.. So, I've tried to modify 'inceptionv3.py' to support float16, as in "https://github.com/apache/incubator-mxnet/blob/master/example/image-classification/symbols/inception-v3.py"

but got this exception: RuntimeError: simple_bind error. Arguments: data: (64, 3L, 300L, 300L) label: (64, 58, 6) Error in operator ch_concat_mixed_2_chconcat_anchors: [18:16:58] include/mxnet/operator.h:227: Check failed: in_type->at(i) == mshadow::default_type_flag || in_type->at(i) == -1 Unsupported data type 2

I guess it is because we have a detection framework and not classification, but it sounds like there's a way.. is there? @zhreshold Thanks, David

[zhreshold]

You need to cast not only data but also label to float16 in symbol. You can follow the instruction on mxnet/example/image-classifcation/symbols/resnet.py Basically it's a mx.sym.cast('float16') for all inputs.

[kaishijeng]

@davsol

Are you able to make float16 working?

Thanks,

[davsol]

I tried but had no luck.. Aside from casting inputs I ran into some errors that I didn't handle

Did you try yourself?

On Feb 21, 2018 6:20 AM, "kaishijeng" [email protected] wrote:

@davsol https://github.com/davsol

Are you able to make float16 working?

Thanks,

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/zhreshold/mxnet-ssd/issues/148#issuecomment-367208973, or mute the thread https://github.com/notifications/unsubscribe-auth/ALzmee9put8VJub_Q3fUKlPCyHzs_yu1ks5tW5mWgaJpZM4QzaO- .

[kaishijeng]

Not yet

Thanks,