Skip to content

Latest commit

 

History

History
93 lines (76 loc) · 6.32 KB

Binary-Network-Results.md

File metadata and controls

93 lines (76 loc) · 6.32 KB

Hyperparameters, training graph, tail of training log, and csv with values in the log can be found in details for each model.

ResNet

CIFAR10

Layers Version Scaled? Best Top1 Acc Best Top5 Acc Details
18 v1 no 87.9 99.7 here

ImageNet

Layers Version Scaled? Best Top1 Acc Best Top5 Acc Details Commment
18 v1 no (49.2) (73.3) here no LR update
18 v1 yes 51.9 75.5 here
18 v2 no 53.2 76.5 here
18 v2 yes 49.5 73.4 here

DenseNet

CIFAR10

Layers Growth-Rate Init-Features Reduction Model Size FP SC Relu SC Best Top1 Acc Best Top5 Acc Details
21 128 128 1,1.4,1.4 1.4 MB no no 87.8 99.5 here
21 128 128 2,2,2 yes no 86.4 99.4 here
21 128 128 2,2,2 yes yes 90.3 99.7 here
21 128 128 1,1.4,1.4 no no 87.6 99.5 here

ImageNet

Layers Growth-Rate Init-Features Reduction Model Size FP SC Relu SC Best Top1 Acc Best Top5 Acc Details
13 256 64 1,1.4,1.4 3.31 MB no no 50.2 73.7 here
21 128 64 1,1.4,1.4 3.39 MB no no 52.7 75.7 here
37 64 64 1,1.4,1.4 3.45 MB no no 54.3 77.3 here
13 256 128 1,1.4,1.4 no no 51.9 75.2 here
21 128 128 1,1.4,1.4 no no 54.1 77.1 here
21 128 128 1.5,1.4,1.4(+sc) no no 52.7 76.2 here
37 64 64 2,2,2 3.08 MB yes yes 57.1 80.0 here
21 128 64 2,2,2 3.03 MB yes yes 55.9 78.5 here
21 128 128 2,2,2 3.48 MB yes yes 57.2 79.6 here
21 160 128 2,2,2 4.4 MB yes yes 59.3 81.4 here
21 160 128 2.2,2.2,2.2 3.99 MB yes yes 58.6 81.0 here

ResNetE

CIFAR10

Layers Version Scaled? FP SC Pool SC Approx Sign Best Top1 Acc Best Top5 Acc Details
18 1 no no yes yes 84.1 99.3 here
18 1 no no yes no 85.6 99.3 here
18 1 no yes yes yes 84.9 99.3 here
18 1 no yes yes no 86.1 99.3 here
18 2 no no yes yes 84.9 99.3 here
18 2 no no yes no 87.2 99.5 here
18 2 no yes yes yes 86.1 99.4 here
18 2 no yes yes no 87.6 99.5 here
18 1 yes no yes yes 83.7 99.2 here
18 1 yes no yes no 85.6 99.3 here
18 1 yes yes yes yes 83.8 99.1 here
18 1 yes yes yes no 86.1 99.2 here
18 2 yes no yes yes 84.2 99.2 here
18 2 yes no yes no 83.6 99.2 here
18 2 yes yes yes yes 84.4 99.3 here
18 2 yes yes yes no 84.7 99.2 here
34 1 no yes yes no 86.2 99.3 here

ImageNet

Layers Version Scaled? FP SC Pool SC Approx Sign Best Top1 Acc Best Top5 Acc Details
18 1 no no yes yes 54.3 77.6 here
18 1 no no yes no 54.4 77.5 here
18 1 no yes yes yes 56.6 79.3 here
18 1 no yes yes no 56.7 79.2 here
18 1 yes yes yes yes 55.3 78.3 here
18 1 yes yes yes no 55.6 78.4 here
18 1 yes no yes yes 53.3 76.4 here
18 1 yes no yes no 52.7 76.1 here
18 2 no yes yes no 54.5 77.5 here
18 2 yes yes yes no 52.1 75.4 here
34 1 no yes yes no 59.5 81.5 here
34 1 no no yes no 58.1 80.6 here

Extended trainings of above models:

Layers Version Scaled? FP SC Pool SC Approx Sign Best Top1 Acc Best Top5 Acc Details
18 1 no yes yes no 56.9 79.7 here
34 1 no yes yes no 60.0 82.0 here