These are chat archives for beniz/deepdetect

5th
Apr 2017
cchadowitz-pf
@cchadowitz-pf
Apr 05 2017 21:49
@beniz I'm trying out the pre-trained openimages-inceptionv3 model (https://deepdetect.com/models/tf/openimages_inception_v3/). Is this simply the model checkpoint discussed here https://github.com/openimages/dataset/wiki/Running-a-pretrained-classifier but adapted for use with DD? If so, and this may be a stretch without knowing if the original checkpoint had been changed between then and now, do you have any ideas as to why the confidences shown here https://github.com/openimages/dataset/issues/3#issuecomment-259552328 are significantly different from those I get from DD? Just trying to determine if somethings off or if this is to be expected.
animal: 0.25574803352355957
fauna: 0.17656546831130981
mammal: 0.12311390042304993
wildlife: 0.12224547564983368
vertebrate: 0.07816784083843231
grass: 0.014900168403983116
pasture: 0.011792576871812344
prairie: 0.011496784165501595
grassland: 0.010585609823465347
meadow: 0.006277795415371656
cchadowitz-pf
@cchadowitz-pf
Apr 05 2017 22:10
I know the OpenImages maintainer mentioned the need to calibrate the output as rare labels or whatnot will have very low confidences - I wasn't sure if the model used in DD was modified at all from the original pre-trained model in the respect.