up
100
作者 blueardour 2020-03-19 12:38:11
Wrote 0 Blogs， Totally 0 words，
 « Homepage » 
Dorefa-net Training

Networks Accuracy trained

  • Network A/W/G Top1 Comment
    det-rs18 2/2/32 66.300 wt-none, sgd-1, FP16, tet + qfn
    det-rs18 2/2/32 65.504 wt-none, sgd-9, FP16, tet + qfn
    det-rs18 2/2/32 65.272 wt-none, sgd-9, FP16, tet + qfn
    det-rs18 2/2/32 65.268 wt-var-mean, sgd-9, FP16, tet + qfn
    det-rs18 2/2/32 65.176 wt-var-mean, sgd-9, FP16, tet + learned-step
    toch-r18 2/2/32 66.804 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, bs=256, decay-small, no stable
    det-rs18 2/2/32 66.752 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, bs=256, decay-small, no stable
    det-rs18 2/2/32 66.320 learned-clip, wt-var-mean, sgd-1, FP16, wtg=1, no stable
    det-rs18 2/2/32 66.222 learned-clip, wt-var-mean, sgd-4, FP16, wtg=1
    det-rs18 2/2/32 65.878 learned-clip, wt-var-mean, sgd-4, FP16, wtg=1, no stable, decay-small, bs=256
    resnet18 2/2/32 64.968 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, bacs
    det-rs18 2/2/32 65.804 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1
    det-rs18 2/2/32 65.692 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, no stable
    det-rs18 2/2/32 65.584 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, no stable, decay-small
    toch-r18 2/2/32 66.876 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1
    toch-r18 2/2/32 66.532 learned-clip, wt-var-mean, sgd-8, FP16, wtg=1, bs=256, init-jing extra-epoch=15
    toch-r18 2/2/32 66.784 learned-step, wt-var-mean, sgd-0, FP16, wtg=1, bs=256, no stable, decay-small
    toch-r18 2/2/32 66.690 learned-step, wt-var-mean, sgd-0, FP16, wtg=1, bs=256
    resnet18 2/2/32 64.984 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, bacs
    det-rs18 2/2/32 65.392 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, fm using tet
    det-rs18 2/2/32 65.666 learned-step, wt-var-mean, sgd-5, FP16, wtg=1
    det-rs18 2/2/32 64.992 learned-step, wt-var-mean, sgd-5, FP16
    det-rs18 2/2/32 41.386 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, no stable, sigmoid-fix4.0
    det-rs18 2/2/32 66.414 learned-step, wt-var-mean, sgd-6, FP16, wtg=1
    toch-r18 2/2/32 66.778 learned-step, wt-var-mean, sgd-6, FP16, wtg=1
    det-rs18 2/2/32 65.978 learned-step, wt-var-mean, sgd-7, FP16, wtg=1
    toch-r18 2/2/32 66.468 learned-step, wt-var-mean, sgd-7, FP16, wtg=1, init-jing
    det-rs18 2/2/32 65.820 learned-step, wt-var-mean, sgd-8, FP16, wtg=1
    det-rs18 2/2/32 66.106 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform + learned-step
    toch-r18 2/2/32 66.504 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform + learned-step
    det-rs18 2/2/32 66.112 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform + learned-clip
    toch-r18 2/2/32 66.032 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform + learned-clip
    bcas-r18 t/t/32 54.052 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, real-skip, full-range, no-stable
    bcas-r18 t/t/32 53.424 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, real-skip, full-range, no-stable, fmv
    toch-r18 t/t/32 64.122 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, init-jing, real-skip fm-max r1
    toch-r18 t/t/32 64.446 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, init-jing, real-skip
    toch-r18 t/t/32 65.056 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip
    toch-r18 t/t/32 65.142 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, no-stable decay-small
    toch-r18 t/t/32 65.492 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, init-jing, real-skip, no-stable decay-small r3
    toch-r18 t/t/32 65.444 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, init-jing, real-skip, no-stable decay-small r2
    toch-r18 t/t/32 64.806 non-uniform-both, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, no-stable, scale=0.8
    toch-r18 t/t/32 65.054 non-uniform-both, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, no-stable, scale=0.8 decay-small
    toch-r18 t/t/32 64.470 non-uniform-both, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, no-stable, scale=1.0 decay-small bug-B
    toch-r18 t/t/32 65.372 non-uniform-both, wt-var-mean, sgd-0, FP16, wtg=1, init-jing, real-skip, no-stable, scale=0.8 decay-small r3 (*)
    toch-r18 t/t/32 65.378 non-uniform-both, wt-var-mean, sgd-0, FP16, wtg=1, init-jing, real-skip, no-stable, scale=0.8 decay-small wd1e-5
    toch-r18 t/t/32 64.930 non-uniform-both, wt-var-mean, sgd-1, FP16, wtg=1, init-jing, real-skip, no-stable, scale=0.8 decay-small wd1e-5
    toch-r18 t/t/32 64.710 non-uniform-both, wt-var-mean, sgd-2, FP16, wtg=1, init-jing, real-skip, no-stable, scale=0.8 decay-small
    toch-r18 t/t/32 62.732 non-uniform-both, wt-var-mean, sgd-8, FP16, wtg=1, init-jing, real-skip, no-stable, scale=0.8 decay-small mixup
    det-rs18 t/t/32 64.502 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip
    det-rs18 t/t/32 64.666 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, no-stable ep100 (*)
    det-rs18 t/t/32 64.664 non-uniform + clip, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip
    det-rs18 t/t/32 64.724 non-uniform-both, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, scale=0.7
    det-rs18 t/t/32 62.570 non-uniform-both, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, re-init-ep84
    toch-r34 t/t/32 69.036 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, init-offi, real-skip, no-stable decay-small (*)
    toch-r34 t/t/32 69.066 learned-clip, wt-var-mean, sgd-1, FP16, wtg=1, init-offi, real-skip, no-stable decay-small
    toch-r34 t/t/32 69.252 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-offi, real-skip, no-stable ep100
    det-rs34 t/t/32 69.570 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-offi, real-skip
    toch-r34 t/t/32 68.946 non-uniform-both, wt-var-mean, sgd-6, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.0
    toch-r34 t/t/32 68.474 non-uniform-both, wt-var-mean, sgd-5, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.0
    toch-r34 t/t/32 68.960 non-uniform-both, wt-var-mean, sgd-0, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.2
    toch-r34 t/t/32 69.464 non-uniform-both, wt-var-mean, sgd-0, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.0 top:88.924
    toch-r34 t/t/32 69.016 non-uniform-both, wt-var-mean, sgd-1, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.0 wd=5e-6
    toch-r34 t/t/32 69.444 non-uniform-both, wt-var-mean, sgd-1, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.0 wd=1e-5
    toch-r34 t/t/32 69.120 non-uniform-both, wt-var-mean, sgd-1, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.0 wd=3e-5
    toch-r34 t/t/32 69.270 non-uniform-both, wt-var-mean, sgd-1, FP16, wtg=1, init-offi, real-skip, no-stable scale=1.0 wd=5e-5
    toch-r50 t/t/32 72.018 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, init-offi, real-skip, no-stable decay-small
    toch-r50 t/t/32 71.768 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.0
    toch-r50 t/t/32 71.470 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.0
    det-rs50 t/t/32 71.786 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, no-stable
    toch-r50 t/t/32 71.514 non-uniform-both, wt-var-mean, sgd-0, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=0.8
    toch-r50 t/t/32 71.456 non-uniform-both, wt-var-mean, sgd-1, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=0.8 w1e-5
    toch-r50 t/t/32 71.418 non-uniform-both, wt-var-mean, sgd-1, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=1.0 w2e-5
    toch-r50 t/t/32 71.648 non-uniform-both, wt-var-mean, sgd-6, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=0.8
    det-rs50 t/t/32 71.466 non-uniform-both, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, no-stable
    tch-r101 t/t/32 73.998 learned-clip, wt-var-mean, sgd-6, FP32, wtg=1, init-offi, real-skip, no-stable decay-small-half ep100
    tch-r101 t/t/32 73.720 learned-clip, wt-var-mean, sgd-0, FP32, wtg=1, init-offi, real-skip, no-stable decay-small
    tch-r101 t/t/32 73.232 non-uniform-both, wt-var-mean, sgd-0, FP32, wtg=1, init-offi, real-skip, no-stable decay-small scale=0.8 w1e-5
    tch-r101 t/t/32 73.322 non-uniform-both, wt-var-mean, sgd-5, FP32, wtg=1, init-offi, real-skip, no-stable decay-small scale=0.8
    tch-r101 t/t/32 74.260 non-uniform-both, wt-var-mean, sgd-6, FP32, wtg=1, init-offi, real-skip, no-stable decay-small scale=0.8 top5:91.820
    tch-r101 t/t/32 74.316 non-uniform-both, wt-var-mean, sgd-0, FP32, wtg=1, init-offi, real-skip, no-stable decay-small scale=0.8 top5:91.728
    det-rs18 2/1/32 64.642 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, no-stable ep100
    toch-r18 2/1/32 64.726 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-jing, real-skip, no-stable ep100
    toch-r18 2/1/32 65.168 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, init-jing, real-skip, no-stable decay-small
    toch-r18 2/1/32 64.892 learned-clip-both, wt-var-mean, sgd-0, FP16, wtg=1, init-jing, real-skip, no-stable decay-small
    toch-r34 2/1/32 69.146 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-offi, real-skip, no-stable ep100
    toch-r34 2/1/32 69.126 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, init-offi, real-skip, no-stable decay-small
    toch-r34 2/1/32 69.120 learned-clip-both, wt-var-mean, sgd-0, FP16, wtg=1, init-offi, real-skip, no-stable decay-small ??
    toch-r50 2/1/32 72.360 learned-clip, wt-var-mean, sgd-6, FP16, wtg=1, init-offi, real-skip, no-stable ep100
    toch-r50 2/1/32 72.690 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, init-offi, real-skip, no-stable decay-small
    toch-r50 2/1/32 72.064 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, init-offi, real-skip, no-stable decay-small scale=0.8
    det-rs50 2/1/32 71.020 learned-clip, wt-var-mean, sgd-5, FP16, wtg=1, init-offi, real-skip, no-stable decay-small
    tch-r101 2/1/32 74.856 learned-clip, wt-var-mean, sgd-0, FP32, wtg=1, init-offi, real-skip, no-stable decay-small
    tch-r101 2/1/32 74.678 learned-clip, wt-var-mean, sgd-6, FP32, wtg=1, init-offi, real-skip, no-stable half-decay-small top5:92.096
    resnet18 32/32/32 68.282 bcas
    toch-r18 32/32/32 65.160 PReLU, sgd-4
    toch-r18 32/32/32 67.976 extra-preBN, sgd-1
    toch-r18 32/32/32 66.582 extra-preBN, sgd-5
    toch-r18 32/32/32 67.828 expand-block=1.5, sgd-4
    toch-r18 32/32/32 68.080 SE-Module=16, sgd-4
    toch-r34 32/32/32 70.956 expand-block=1.5, sgd-4
    toch-r18 t/t/32 63.160 learned-step, wt-var-mean, sgd-0, FP16, wtg=1, bs=256 init-jing decay-small no-stable
    det-rs18 t/t/32 63.200 learned-step, wt-var-mean, sgd-1, FP16, wtg=1
    det-rs18 t/t/32 57.727 learned-step, wt-var-mean, sgd-2, FP16, wtg=1
    det-rs18 t/t/32 63.014 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, no stable
    det-rs18 t/t/32 29.692 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, no stable, sigmoid-fix4.0
    det-rs18 t/t/32 62.580 learned-step, wt-var-mean, sgd-5, FP16, no stable
    det-rs18 t/t/32 64.229 learned-step, wt-var-mean, sgd-6, FP16, wtg=1
    toch-r18 t/t/32 63.966 learned-step, wt-var-mean, sgd-6, FP16, wtg=1, init-jing
    det-rs18 t/t/32 64.039 learned-step, wt-var-mean, sgd-7, FP16, wtg=1
    det-rs18 t/t/32 63.523 learned-step, wt-var-mean, sgd-8, FP16, wtg=1
    det-rs18 t/t/32 63.302 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform(fm-only), no stable, re-init
    det-rs18 t/t/32 63.160 learned-step, wt-var-mean, adam-1, FP16, wtg=1, non-uniform(fm-only), no stable, re-init
    det-rs18 t/t/32 62.986 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform-both
    det-rs18 t/t/32 61.146 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform-both r2 scale=2.5
    det-rs18 t/t/32 62.800 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform(fm-only)
    det-rs18 t/t/32 63.230 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform(fm-only) r2
    toch-r18 t/t/32 63.570 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform-both r1-lcl scale=0.8
    toch-r18 t/t/32 61.606 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform-both r2 scale=2.5
    toch-r18 t/t/32 63.154 learned-step, wt-var-mean, sgd-5, FP16, wtg=1, non-uniform(fm-only) r2
    det-rs18 t/t/32 46.190 learned-step-ter, wt-var-mean, sgd-4, FP16, wtg=1, no-stable, train-stratch, PReLU, scale=1
    det-rs18 t/t/32 59.392 learned-step-ter, wt-var-mean, adam-1, FP16, wtg=1, no-stable, train-stratch, PReLU, scale=1
    resnet18 32/32/32 68.732 TResnet, sgd-1, FP16
    resnet18 32/32/32 68.448 TResnet, sgd-1, FP16, sk=3
    resnet18 32/32/32 67.222 TResnet, sgd-2, FP16
    toch-r18 b/b/32 57.700 learned-clip-both, wt-var-mean, sgd-1, FP16, wtg=1, init-jing, real-skip, no-stable, scale=1.0 decay-small
    toch-r18 b/b/32 56.114 learned-clip-both, wt-var-mean, sgd-2, FP16, wtg=1, init-jing, real-skip, no-stable, scale=1.0 decay-small
    toch-r18 b/b/32 58.488 learned-clip, wt-var-mean, sgd-0, FP16, wtg=1, init-jing, real-skip, no-stable, scale=1.0 decay-small
    toch-r18 b/b/32 57.922 learned-clip, wt-var-mean, sgd-1, FP16, wtg=1, init-jing, real-skip, no-stable, scale=1.0 decay-small
    toch-r18 b/b/32 56.468 learned-clip, wt-var-mean, sgd-2, FP16, wtg=1, init-jing, real-skip, no-stable, scale=1.0 decay-small
    toch-r18 b/b/32 58.120 learned-clip, wt-var-mean, sgd-1, FP16, wtg=1, init-jing, real-skip, no-stable, scale=1.0 decay-small scale-skip

© Blueardour' Technic Blog

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