up
100
作者 blueardour 2020-01-06 12:38:11
Wrote 0 Blogs， Totally 0 words，
 « Homepage » 
COCO Detection

Networks Accuracy trained

Paper Dataset Network Bit(A/W/G) Paper report My Accuracy Comment
Retina-Net COCO MSRA-R50 32/32/32 - 36.465 1x
Retina-Net COCO det-R50 32/32/32 - 35.070 1x round2
Retina-Net COCO det-R50 32/32/32 - 36.564 1x round3
Retina-Net COCO det-R50 32/32/32 - 37.765 1x-Full-BN syncbn bs=8 double-iteration lr=0.005
Retina-Net COCO det-R34 32/32/32 - 31.661 1x
Retina-Net COCO det-R34 32/32/32 - 30.134 1x round2
Retina-Net COCO det-R34 32/32/32 - 33.240 1x round3
Retina-Net COCO det-R34 32/32/32 - 35.168 1x round4
Retina-Net COCO det-R18 32/32/32 - 27.488 1x
Retina-Net COCO det-R18 32/32/32 - 31.523 1x round2
Retina-Net COCO det-R18 32/32/32 - 08.491 1x fp16
Retina-Net COCO det-R18 32/32/32 - 32.681 1x FPN-BN
Retina-Net COCO det-R18 32/32/32 - 32.983 1x Full-BN
FCOS COCO MSRA-R50 32/32/32 - 36.002 1x bs=8
FCOS COCO MSRA-R50 32/32/32 - 38.561 1x
FCOS COCO MSRA-R50 32/32/32 - 09.990 1x fp16
FCOS COCO det-R50 32/32/32 - 38.349 1x
FCOS COCO det-R50 32/32/32 - 35.146 1x fp16, FPN-BN, bs=8
FCOS COCO det-R50 32/32/32 - 38.451 1x FPN-BN
FCOS COCO det-R50 32/32/32 - 38.894 1x Full-BN
FCOS COCO det-R34 32/32/32 - 37.313 1x
FCOS COCO det-R18 32/32/32 - 32.157 1x, bias on
FCOS COCO det-R18 32/32/32 - 31.819 1x, bias off
FCOS COCO det-R18 32/32/32 - 33.350 1x FPN-BN
FCOS COCO det-R18 32/32/32 - 33.909 1x Full-BN
FCOS COCO det-R18 32/32/32 - 23.713 1x Full-BN round2 bs=12
FCOS COCO det-R18 32/32/32 - 33.759 1x Full-BN round3
FCOS COCO det-R18 32/32/32 - 34.271 1x Full-BN round4 syncbn
FCOS COCO det-R18 32/32/32 - 33.504 1x fp16
FCOS COCO det-R18 32/32/32 - 33.865 1x FPN-BN-fp16
FCOS COCO det-R18 32/32/32 - 33.488 1x FPN-BN-fp16 round2
FCOS COCO det-R18 32/32/32 - 32.901 1x FPN-BN-fp16 round3 RELU-INPLACE-False
FCOS COCO det-R18 32/32/32 - 32.786 1x FPN-BN-fp16-input
FCOS COCO det-R18 32/32/32 - 18.593 1x Full-BN-fp16 round1
FCOS COCO det-R18 32/32/32 - 26.671 1x Full-BN-fp16 round2
FCOS COCO det-R18 32/32/32 - 26.149 1x Full-BN-fp16 round3 single-GPU
FCOS COCO det-R18 2/2/32 - 32.549 1x FPN-BN fp16 dorefa-tet backbone-only
FCOS COCO det-R18 2/2/32 - 33.384 1x FPN-BN fp16 dorefa-tet backbone-only round2
FCOS COCO det-R18 2/2/32 - 31.786 1x FPN-BN fp16 dorefa-tet backbone-fpn
FCOS COCO det-R18 2/2/32 - 30.333 1x FPN-BN fp16 dorefa-tet whole
FCOS COCO det-R18 2/2/32 - 26.334 1x FPN-BN fp16 dorefa-tet whole lr=2.5e-3
FCOS COCO det-R18 2/2/32 - 31.981 1x FPN-BN fp16 dorefa-tet whole wtet
FCOS COCO det-R18 ter/ter/32 - 26.494 1x, bias on , pocount backbone only
FCOS COCO det-R18 ter/ter/32 - 26.149 1x, bias off, pocount backbone only
FCOS COCO det-R18 ter/ter/32 - 19.657 1x, bias off, pocount backbone only, lr=1e-3
FCOS COCO det-R18 ter/ter/32 - 31.355 1x FPN-BN fp16 popcount backbone-only
FCOS COCO det-R18 ter/ter/32 - 32.579 1x FPN-BN fp16 popcount backbone-only bs=32
FCOS COCO det-R18 ter/ter/32 - 26.171 1x FPN-BN fp16 popcount whole
  1. export PYTHONWARNINGS=’ignore:semaphore_tracker:UserWarning’

© Blueardour' Technic Blog

TAGS: experiments training-trick COCO detection result
 « Homepage » 
-->