Skip to content

Latest commit

 

History

History
46 lines (37 loc) · 2.94 KB

results.md

File metadata and controls

46 lines (37 loc) · 2.94 KB
Raw

Visualization of Time-Domain Augmentations (UCIHAR dataset)

time_aug_heatmap

Visualization of Frequency-Domain Augmentations (SHAR dataset)

freq_aug_heatmap

Augmentation Results on BYOL

UCIHAR SHAR
byol_aug_ucihar byol_aug_shar_fcn

Augmentation Results on SimSiam

UCIHAR SHAR
simsiam_aug_ucihar simsiam_aug_shar_fcn

Augmentation Results on SimCLR

UCIHAR SHAR
simclr_aug_ucihar simclr_aug_shar_fcn

Augmentation Results on NNCLR

UCIHAR SHAR
nnclr_aug_ucihar nnclr_aug_shar_fcn

Augmentation Results on TS-TCC

UCIHAR SHAR
tstcc_aug_ucihar nnclr_aug_shar_fcn

The settings listed in the table below achieve reasonable performance. However, the actual hyper-parameter setup should be subject to the changes of backbone networks and augmentations.

Dataset Model Learning rate Batch Size Weight Decay Temperature EMA_decay Memory Bank Size Epoch
UCIHAR BYOL 5e-4 128 1.5e-6 - 0.996 - 60
SimSiam 5e-4 128 1e-4 - - - 60
SimCLR 3e-3 256 1e-6 0.1 - - 120
NNCLR 3e-3 256 1e-6 0.1 - 1024 120
TS-TCC 3e-4 128 3e-4 0.2 - - 40
SHAR BYOL 1e-3 64 1.5e-6 - 0.996 - 60
SimSiam 3e-4 256 1e-4 - - - 60
SimCLR 2.5e-3 256 1e-6 0.1 - - 120
NNCLR 2e-3 256 1e-6 0.1 - 1024 120
TS-TCC 3e-4 128 3e-4 0.2 - - 40