up
100
作者 blueardour 2019-07-17 12:38:11
Wrote 0 Blogs， Totally 0 words，
 « Homepage » 
ALQnet Training History

Networks Accuracy trained

Paper Dataset Network Bit(A/W/G) Paper report My Accuracy Comment
Dorefa imagenet resnet18 32/32/32 - 70.09 bacs, without fix
  imagenet resnet34 32/32/32 - 73.328 bacs, without fix
  imagenet resnet50 32/32/32 - 76.03 bacs, fix
LQ-net imagenet resnet18 32/2/32 68.0 68.70 finetune-0
LQ-net imagenet resnet18 32/3/32 69.3 69.54 finetune-0
LQ-net imagenet resnet18 32/4/32 70.0 - finetune-0
ALQ-net imagenet resnet18 32/2/32 - 68.69 finetune-0
ALQ-net imagenet resnet18 32/3/32 - 69.04 finetune-0
ALQ-net imagenet resnet18 32/3/32 - 69.75 finetune-2
LQ-net imagenet resnet18 1/1/32 - 54.702 finetune-0
LQ-net imagenet resnet18 1/1/32 - 53.632 finetune-1
LQ-net imagenet resnet18 1/1/32 - 54.786 finetune-2
LQ-net imagenet resnet18 2/2/32 64.9 65.000 finetune-0
LQ-net imagenet resnet18 2/2/32 64.9 64.756 finetune-1
LQ-net imagenet resnet18 2/2/32 64.9 65.964 finetune-1, fg=8
LQ-net imagenet resnet18 2/2/32 64.9 66.294 finetune-1, fg1, wt-mean
LQ-net imagenet resnet18 2/2/32 64.9 66.338 finetune-1, fg1, wt-mean-var
LQ-net imagenet resnet18 2/2/32 64.9 60.958 finetune-1, fg1, fm-mean(plus)
LQ-net imagenet resnet18 2/2/32 64.9 65.390 finetune-1, fg1, fm-var (bohan)
LQ-net imagenet resnet18 2/2/32 64.9 64.708 finetune-2
LQ-net imagenet resnet18 3/3/32 68.2 68.020 finetune-0
LQ-net imagenet resnet18 3/3/32 68.2 67.714 finetune-1
LQ-net imagenet resnet18 3/3/32 68.2 68.390 finetune-2
LQ-net imagenet resnet18 4/4/32 - 69.678 finetune-2, adaptive
LQ-net imagenet resnet34 2/2/32 69.8 70.266 finetune-2
LQ-net imagenet resnet34 3/3/32 71.9 72.552 finetune-2
LQ-net imagenet resnet50 2/2/32 71.5 72.028 finetune-2
LQ-net imagenet resnet50 3/3/32 74.2 74.144 finetune-2
on   - - - -  
LQ-net imagenet resnet18 2/2/32 - 65.226 sgd-1, epoch 24, baseline
LQ-net imagenet resnet18 2/2/32 - 65.130 sgd-1, epoch 24, non-uniform-pact
LQ-net imagenet resnet18 2/2/32 - 65.092 sgd-1, epoch 24, non-uniform-gradscale
LQ-net imagenet resnet18 2/2/32 - 64.908 sgd-1, epoch 24, non-uniform-FixWithGrad
LQ-net imagenet resnet18 2/2/32 - 64.930 sgd-4, epoch 30, baseline
LQ-net imagenet resnet18 2/2/32 - 64.546 sgd-4, epoch 30, baseline, loop2
LQ-net imagenet resnet18 2/2/32 - 64.782 sgd-4, epoch 30, non-uniform-gradscale
LQ-net imagenet resnet18 2/2/32 - 64.946 sgd-4, epoch 30, non-uniform-FixWithGrad
LQ-net imagenet resnet18 2/2/32 - 65.122 sgd-4, epoch 30, non-uniform-FixWithGrad, round2
LQ-net imagenet resnet18 2/2/32 - 65.006 sgd-4, epoch 30, non-uniform-FixWithGrad, round3, lr=1
LQ-net imagenet resnet18 2/2/32 - 64.892 sgd-4, epoch 30, non-uniform-FixWithGrad, round4, lr=custom lr
LQ-net imagenet resnet18 2/2/32 - 64.872 sgd-4, epoch 30, non-uniform-FixWithGrad, round5, lr=10
LQ-net imagenet resnet18 2/2/32 - 64.992 sgd-4, epoch 30, non-uniform-FixWithGrad, round6, lr=100
LQ-net imagenet resnet18 2/2/32 - 63.610 sgd-4, epoch 30, non-uniform-GradSlope
LQ-net imagenet resnet18 2/2/32 - 65.082 sgd-4, epoch 30, spatial, quant-group=4, pad-after-quant
LQ-net imagenet resnet18 2/2/32 - 60.724 sgd-4, epoch 30, spatial, quant-group=8, pad-after-quant
LQ-net imagenet resnet18 2/2/32 - 56.950 sgd-4, epoch 30, spatial, quant-group=8, spatial-bn, early try
LQ-net imagenet resnet18 32/32/32 - 69.324 sgd-1, epoch 120, spatial, quant-group=8, spatial-bn
LQ-net imagenet resnet18 2/2/32 - 60.296 sgd-4, epoch 30, spatial, quant-group=8, spatial-bn
off   - - - -  
ALQ-net imagenet resnet18 1/1/32 - 53.856 finetune-0 poly adam
ALQ-net imagenet resnet18 1/1/32 - 52.948 finetune-1 sgdr sgd
ALQ-net imagenet resnet18 t/t/32 - 63.140 finetune-0
ALQ-net imagenet resnet18 t/t/32 - 63.438 finetune-1
ALQ-net imagenet resnet18 t/t/32 - 62.316 finetune-2
ALQ-net imagenet resnet18 2/2/32 - 65.698 finetune-0
ALQ-net imagenet resnet18 2/2/32 - 66.276 finetune-1
ALQ-net imagenet resnet18 2/2/32 - 65.944 finetune-2
ALQ-net imagenet resnet18 3/3/32 - 65.758 finetune-0
ALQ-net imagenet resnet18 3/3/32 - 68.148 finetune-1
ALQ-net imagenet resnet18 3/3/32 - 68.670 finetune-2
ALQ-net imagenet resnet18 3/3/32 - 61.114 finetune-3
ALQ-net imagenet resnet18 3/3/32 - 67.116 finetune-4, adam
ALQ-net imagenet resnet34 1/1/32 - 58.862 finetune-2
ALQ-net imagenet resnet34 t/t/32 - 67.772 finetune-1
ALQ-net imagenet resnet34 2/2/32 - 70.656 finetune-1
ALQ-net imagenet resnet34 2/2/32 - 69.944 finetune-2
ALQ-net imagenet resnet34 3/3/32 - 72.374 finetune-2
ALQ-net imagenet resnet50 2/2/32 - 73.320 finetune-2
ALQ-net imagenet resnet50 3/3/32 - 74.240 finetune-2
<> - - - - - -
LQ-net cifar100 resnet20 32/32/32 - 68.75 bacs
LQ-net cifar100 resnet20 2/2/32 - 63.50 bacs, v0-0
LQ-net cifar100 resnet20 2/2/32 - 63.59 bacs, v0-1
MLQ-net cifar100 resnet20 2/2/32 - 63.35 bacs, v0-0
MLQ-net cifar100 resnet20 2/2/32 - 62.95 bacs, v0-1
NLQ-net cifar100 resnet20 2/2/32 - 63.35 bacs, v0-0
NLQ-net cifar100 resnet20 2/2/32 - 63.48 bacs, v0-1
ALQ-net cifar100 resnet20 2/2/32 - 63.66 bacs, v0-0
ALQ-net cifar100 resnet20 2/2/32 - 63.71 bacs, v0-1
LQ-net cifar100 resnet20 2/2/32 - 63.12 bacs, custom-update-0
LQ-net cifar100 resnet20 2/2/32 - 63.34 bacs, custom-update-1
MLQ-net cifar100 resnet20 2/2/32 - 63.47 bacs, custom-update-0
MLQ-net cifar100 resnet20 2/2/32 - 63.50 bacs, custom-update-1
NLQ-net cifar100 resnet20 2/2/32 - 63.20 bacs, custom-update-0
NLQ-net cifar100 resnet20 2/2/32 - 63.21 bacs, custom-update-1
ALQ-net cifar100 resnet20 2/2/32 - 63.40 bacs, custom-update-0
ALQ-net cifar100 resnet20 2/2/32 - 63.83 bacs, custom-update-1
LQ-net cifar10 resnet20 2/2/32 - 89.85 bacs, v0-0
LQ-net cifar10 resnet20 2/2/32 - 90.02 bacs, v0-1
LQ-net cifar10 resnet20 2/2/32 - 89.77 bacs, v1-0
LQ-net cifar10 resnet20 2/2/32 - 89.60 bacs, v1-1
MLQ-net cifar10 resnet20 2/2/32 - 89.96 bacs, v0-0
MLQ-net cifar10 resnet20 2/2/32 - 89.72 bacs, v0-1
NLQ-net cifar10 resnet20 2/2/32 - 89.97 bacs, v0-0
NLQ-net cifar10 resnet20 2/2/32 - 89.99 bacs, v0-1
ALQ-net cifar10 resnet20 2/2/32 - 90.01 bacs, v0-0
ALQ-net cifar10 resnet20 2/2/32 - 90.18 bacs, v0-1
ALQ-net cifar10 resnet20 2/2/32 - 89.47 bacs, v1-0
ALQ-net cifar10 resnet20 2/2/32 - 90.01 bacs, v1-1
LQ-net cifar100 resnet20 2/2/32 - 63.99 bacs, custom-update-2
LQ-net cifar100 resnet20 2/2/32 - 64.22 bacs, custom-update-3
ALQ-net cifar100 resnet20 2/2/32 - 64.32 bacs, custom-update-2
ALQ-net cifar100 resnet20 2/2/32 - 63.75 bacs, custom-update-3

© Blueardour' Technic Blog

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