up
100
作者 blueardour 2019-10-28 12:38:11
Wrote 0 BlogsTotally 0 words
PopcountNet

Networks Accuracy trained

Paper Dataset Network Bit(A/W/G) Paper report My Accuracy Comment
- imagenet resnet18 32/32/32 - 70.090 bacs-0
- imagenet resnet18 t/ t/32 - 62.138 sgd-1, all quant
- imagenet resnet18 t/ t/32 - 63.700 sgd-1, force-fp
- imagenet resnet18 t/ t/32 - 58.066 sgd-3, all quant 0.250
- imagenet resnet18 t/ t/32 - 61.986 sgd-3, all quant 0.500
- imagenet resnet18 t/ t/32 - 62.760 sgd-3, all quant 0.618
- imagenet resnet18 t/ t/32 - 62.9?? sgd-1, all quant 0.625
- imagenet resnet18 t/ t/32 - 61.986 sgd-3, all quant 0.750
- imagenet resnet18 t/ t/32 - 62.176 sgd-3, all quant 0.625, wt adaptive: mean-var
- imagenet resnet18 t/ t/32 - 62.968 sgd-1, all quant 0.625, wt adaptive: mean-var
- <-> - - - - -
- imagenet bireal18 32/32/32 - 68.892 cbsa-1
- imagenet bireal18 t/ t/32 - 63.146 sgd-3, force-fp
- imagenet bireal18 t/ t/32 - 61.816 sgd-3, all quant
- <-> - - - - -
- imagenet resnet34 32/32/32 - 73.328 bacs-1
- imagenet resnet34 t/ t/32 - 67.442 sgd-1, all quant
- imagenet resnet34 t/ t/32 - 68.990 sgd-1, force-fp
- imagenet resnet34 t/ t/32 - 68.020 sgd-1, all quant 0.625
- <-> - - - - -
- imagenet bireal34 32/32/32 - 67.610 cbsa-1, custom-step
- imagenet bireal34 32/32/32 - 69.390 cbsa-1, sgdr
- imagenet bireal34 t/ t/32 - 66.058 sgd-3, force-fp
- imagenet bireal34 t/ t/32 - xx.xxx sgd-3, all quant
- <-> - - - - -
- imagenet resnet50 32/32/32 - xx.xxx bacs-1
- imagenet resnet50 t/ t/32 - xx.xxx sgd-1, all quant
- imagenet resnet50 t/ t/32 - 71.1?? sgd-1, all quant 0.625
- <-> - - - - -
- imagenet det-rs18 t/ t/32 - 62.576 sgd-3, all quant 0.7
- imagenet det-rs18 t/ t/32 - 62.720 sgd-3, all quant 0.7, wt adaptive: mean-var
- imagenet det-rs18 t/ t/32 - 62.746 sgd-3, all quant 0.625, wt adaptive: mean-var
- imagenet det-rs34 t/ t/32 - 67.970 sgd-3, all quant 0.625, wt adaptive: mean-var

Tips:
update the threshold according to
‘Ternary neural networks with fine-grained quantization’ and ‘TWN’?

-->