up
100
作者 blueardour 2019-04-20 12:38:11
Wrote 0 Blogs， Totally 0 words，
 « Homepage » 
Training History

Networks Accuracy trained

Paper Dataset Network Bit(A/W/G) Paper report My Accuracy Comment
Group-net imagenet resnet18 1/1/32 63.x 63.772 without-softgate, sgd-with-decay, small-lr
LQ-net cifar10 vgg-small 32/32/32 93.8 94.21 -
LQ-net cifar10 vgg-small 2/1/32 93.4 - -
LQ-net cifar10 vgg-small 2/2/32 93.5 94.41 -
LQ-net cifar10 vgg-small 3/2/32 93.8 - -
LQ-net cifar10 vgg-small 3/3/32 93.8 - -
LQ-net cifar10 resnet20 32/32/32 92.1 92.86 bacs
LQ-net cifar10 resnet20 32/32/32 92.1 92.36 cbas,proxquant
LQ-net cifar10 resnet20 2/1/32 88.4 88.97 bacs
LQ-net cifar10 resnet20 2/2/32 90.2 90.16 bacs
LQ-net cifar10 resnet20 2/2/32 90.2 90.84 bacs, momentum, lr=0.1, v7-1
MLQ-net cifar10 resnet20 2/2/32 90.2 90.41 bacs, lqnet W + alqnet A
ALQ-net cifar10 resnet20 2/2/32 90.2 90.36 bacs, v0_rand_1
LQ-net cifar10 resnet20 3/2/32 91.1 91.34 bacs
LQ-net cifar10 resnet20 3/3/32 91.6 92.23 bacs
LQ-net cifar10 resnet20 32/2/32 91.8 91.86 bacs, _0
ALQ-net cifar10 resnet20 32/2/32 91.8 91.93 bacs, v2-0
LQ-net cifar10 mobilenetv1 32/32/32 - 90.93 cbas
LQ-net cifar100 resnet20 32/32/32 - 68.75 bacs
LQ-net cifar100 resnet20 32/2/32 - 67.3 bacs
LQ-net cifar100 resnet20 2/2/32 - 63.10 bacs, v7-0.001_v9_off_0
MLQ-net cifar100 resnet20 2/2/32 - 63.55 bacs, av0_wv0-v9_1
NLQ-net cifar100 resnet20 2/2/32 - 63.88 bacs, aoff-v0_wv0_1
NLQ-net cifar100 resnet20 2/2/32 - 63.73 bacs, aoff-v0_wv0_3
ALQ-net cifar100 resnet20 2/2/32 - 63.42 bacs, v0_1
ALQ-net cifar100 resnet20 2/2/32 - 63.7 bacs, v0_nod_1
ALQ-net cifar100 resnet20 2/2/32 - 63.57 bacs, v0_nod_2
ALQ-net cifar100 resnet20 2/2/32 - 63.58 bacs, v0_rand_1
ALQ-net cifar100 resnet20 2/2/32 - 64.47 bacs, v0_nod2_1
LQ-net cifar100 mobilenetv1 32/32/32 65.98 68.23 cbas, weiaicunzai
LQ-net imagenet alexnet 32/32/32 61.8 62.57 acb, dali, server
LQ-net imagenet alexnet 32/32/32 61.8 62.644 acb, phoenix, imagenet
LQ-net imagenet alexnet 32/2/32 60.5 60.418 acb, imagenet
LQ-net imagenet resnet18 32/32/32 69.6 69.7 epochs=100, SGDR, bacs
LQ-net imagenet resnet18 32/32/32 69.6 70.2 epochs=120, custom-step, bacs, imagenet
LQ-net imagenet resnet18 32/32/32 69.6 70.09 epochs=120, custom-step, bacs, dali
LQ-net imagenet resnet18 2/2/32 64.0 64.19 archlab, epoch2=120, custom-step, imagenet
LQ-net imagenet resnet18 2/1/32 62.6 ? server
Dorefa cifar10 resnet20 32/32/32 - 92.86 TTN,bacs
Dorefa cifar10 resnet20 32/1/32 - 90.47 pytorch-dorefa
Dorefa cifar10 resnet20 32/2/32 - 91.7 pytorch-dorefa
Dorefa cifar10 resnet20 32/1/32 - 90.95 my code, bacs
Dorefa cifar10 resnet20 2/2/32 - 89.51 my code, cbas, stratch
Dorefa cifar10 resnet20 2/2/32 - 85.06 my code, cbas, finetune, epoch=100
Dorefa cifar10 resnet20 2/2/32 - 89.65 my code, cbas, finetune, epoch=200
PACT cifar10 resnet20 2/2/32 - 89.36 my code, cbas, stratch
Dorefa cifar10 resnet20 3/3/32 - 90.44 my code, cbas, stratch
Dorefa imagenet alexnet 32/32/32 61.8 61.83 acb, imagenet, sgdr
Dorefa imagenet alexnet 32/32/32 61.8 57.176 acb, imagenet, custom-step
Dorefa imagenet alexnet 2/1/32 53.4 56.524 acb, imagenet, adam, stratch
Dorefa imagenet resnet18 4/1/32 59.2 65.348 bacs, adam, 90 epochs, finetune
Dorefa imagenet resnet18 4/1/32 59.2 61.156 bacs, sgd, 120 epochs, stratch
Dorefa imagenet resnet18 2/2/32 - 64.206 bacs, sgd-4, 30 epochs, finetune
Dorefa-TET imagenet resnet18 2/2/32 - 66.734 bacs, wt-var, sgd-5,
Dorefa-TET imagenet resnet18 2/2/32 - 67.008 bacs, wt-var, sgd-4, bs=256, imagenet, WD=1e-4
Dorefa-TET imagenet resnet18 2/2/32 - 66.856 bacs, wt-var, sgd-4, FP16
Dorefa-TET imagenet resnet18 2/2/32 - 66.842 bacs, wt-var, sgd-9, FP16, 40EP, n-d-s, WD=2e-5
Dorefa-TET imagenet resnet18 1/1/32 - 51.338 bacs, wt-var, sgd-9, FP16, 40EP, n-d-s, WD=2e-5
Dorefa-TET imagenet resnet18 1/1/32 - ??.??? bacs, wt-var, sgd-9, FP16, 40EP, n-d-s, WD=2e-5, wtet
Dorefa-TET imagenet resnet18 2/2/32 - 65.796 bacs, wt-non, sgd-9, FP16, 40EP, n-d-s, WD=2e-5, wtet
Dorefa-TET imagenet resnet18 2/2/32 - 66.962 bacs, wt-var, sgd-9, FP16, 40EP, n-d-s, WD=2e-5, wtet
Dorefa-TET imagenet resnet18 2/2/32 - 63.132 bacs, wt-var, sgd-9, bs=1024 40EP
Dorefa-TET imagenet resnet18 2/2/32 - 66.788 bacs, wt-var, sgd-9, FP16, 40EP, n-d-s, WD=1e-4
Dorefa-TET imagenet resnet18 2/2/32 - 65.918 bacs, wt-var, sgd-9, FP16, 40EP, n-d-s, WD=1e-4 scale5-fan
Dorefa-TET imagenet resnet18 2/2/32 - diverg bacs, wt-var, sgd-4, FP16, grad-scale:fan-scale2
Dorefa-TET imagenet resnet18 2/2/32 - diverg bacs, wt-var, sgd-4, FP16, grad-scale:fan-scale1
Dorefa-TET imagenet resnet18 2/2/32 - 63.586 bacs, wt-var, sgd-4, FP16, grad-scale:mean-fan-scale2
Dorefa-TET imagenet resnet18 2/2/32 - 63.078 bacs, wt-var, sgd-4, FP16, grad-scale:mean-fan-scale1
Dorefa-TET imagenet resnet18 2/2/32 - 66.032 bacs, wt-var, sgd-8, mixup0.7, 60EP, n-d-s
Dorefa-TET imagenet resnet18 2/2/32 - 66.508 bacs, wt-var, sgd-4, fix-arch
Dorefa-TET imagenet resnet18 2/2/32 - 66.968 bacs, wt-var, sgd-4, fix-arch, small WD
Dorefa-TET imagenet resnet18 2/2/32 - 66.442 bacs, wt-var, sgd-7, fix-arch, mixup0.7, 90EP, n-d-s
Dorefa-TET imagenet resnet18 2/2/32 - 66.802 bacs, wt-var, sgd-4, fix-arch2 singleconv
Dorefa-TET imagenet det-r18 2/2/32 - 65.504 wt-mean-var {1}, sgd-9, FP16 O1 wd2e-5
Dorefa-TET imagenet det-r34 2/2/32 - 69.858 wt-mean-var {1}, sgd-9, FP16 O1 wd2e-5
Dorefa-TET imagenet det-r18 2/2/32 - 67.306 wt-var, sgd-9, FP16 O1 wd2e-5, wtet
Dorefa-TET imagenet det-r34 2/2/32 - 71.122 wt-var, sgd-9, FP16 O1 wd2e-5, wtet
Dorefa-TET imagenet det-r50 2/2/32 - on-progress wt-var, sgd-9, FP16 O1 wd2e-5, wtet
Dorefa-TET imagenet det-r18 3/3/32 - on-progress wt-var, sgd-9, FP16 O1 wd2e-5, wtet
Dorefa-TET imagenet det-r50 3/3/32 - on-progress wt-var, sgd-9, FP16 O1 wd2e-5, wtet
Fixup imagenet resnet18 32/32/32 68.776 68.956 cbsa, mixup0.7, 120 epochs, stratch
Fixup imagenet resnet18 32/32/32 68.776 68.776 cbsa, no mixup, 120 epochs, stratch
HORQ++ imagenet resnet18 32/32/32 - 67.902 bacs, PReLU
HORQ++ imagenet resnet18 32/32/32 - 68.282 bcas

{1}: has no effect (no weight normalization), as it uses the dorefa.qfn/tet-wt function which supports var only

© Blueardour' Technic Blog

TAGS: experiments training-trick quantization
 « Homepage » 
-->