Skip to content

v1.0.0

Latest
Compare
Choose a tag to compare
@wgcban wgcban released this 29 Nov 05:26
· 129 commits to main since this release

Initial release of pre-trained models of Mixed Barlow Twins.

ResNet-18 [CIFAR-10, CIFAR-100, TinyImageNet, and STL-10]

Dataset $d$ $\lambda_{BT}$ $\lambda_{reg}$ Download Link to Pretrained Model KNN Acc. Linear Acc.
CIFAR-10 1024 0.0078125 4.0 4wdhbpcf_0.0078125_1024_256_cifar10_model.pth 90.52 92.58
CIFAR-100 1024 0.0078125 4.0 76kk7scz_0.0078125_1024_256_cifar100_model.pth 61.25 69.31
TinyImageNet 1024 0.0009765 4.0 02azq6fs_0.0009765_1024_256_tiny_imagenet_model.pth 38.11 51.67
STL-10 1024 0.0078125 2.0 i7det4xq_0.0078125_1024_256_stl10_model.pth 88.94 91.02

ResNet-50 [CIFAR-10, CIFAR-100, TinyImageNet, and STL-10]

Dataset $d$ $\lambda_{BT}$ $\lambda_{reg}$ Download Link to Pretrained Model KNN Acc. Linear Acc.
CIFAR-10 1024 0.0078125 4.0 v3gwgusq_0.0078125_1024_256_cifar10_model.pth 91.39 93.89
CIFAR-100 1024 0.0078125 4.0 z6ngefw7_0.0078125_1024_256_cifar100_model.pth 64.32 72.51
TinyImageNet 1024 0.0009765 4.0 kxlkigsv_0.0009765_1024_256_tiny_imagenet_model.pth 42.21 51.84
STL-10 1024 0.0078125 2.0 pbknx38b_0.0078125_1024_256_stl10_model.pth 87.79 91.70

ResNet-50 on ImageNet

Setting: epochs = 300, $d$ = 8192, $\lambda_{BT}$ = 0.0051

$\lambda_{reg}$ Download Link to Pretrained Model Train Log Download Link to Linear-Probed Model Val. Log Linear Acc.
0.0 (BT) 3on0l4wl_0.0000_8192_1024_imagenet_resnet50.pth train_log checkpoint_3tb4tcvp.pth val_log 71.3
0.0025 l418b9zw_0.0025_8192_1024_imagenet_resnet50.pth train_log checkpoint_09g7ytcz.pth val_log 70.9
0.1 13awtq23_0.1000_8192_1024_imagenet_resnet50.pth train_log checkpoint_pgawzr4e.pth val_log 71.6
1.0 3fb1op86_1.0000_8192_1024_imagenet_resnet50.pth train_log checkpoint_wvi0hle8.pth val_log 72.2
2.0 5n9yqio0_1.0000_8192_1024_imagenet_resnet50.pth train_log checkpoint_p9aeo8ga.pth val_log 72.1
3.0 q03u2xjz_3.0000_8192_1024_imagenet_resnet50.pth train_log checkpoint_00atvp6x.pth val_log 72.0

Download the pre-trained models and store them in the checkpoints/ folder.