Neural Backed Decision Trees Versions Save

Making decision trees competitive with neural networks on CIFAR10, CIFAR100, TinyImagenet200, Imagenet

0.0.4

4 years ago

0.0.2

4 years ago
  • Code release on PyPi for testing
  • Includes 3 console scripts: nbdt, nbdt-hierarchy, nbdt-wnids
  • A number of default hierarchies and WordNet IDs for CIFAR10, CIFAR100, TinyImagenet, Imagenet

0.0.1

4 years ago

The following checkpoints were produced by training with this repository.

Includes checkpoints for the following neural networks:

  • WideResNet28x10 on TinyImagenet200
  • ResNet18 on CIFAR10, CIFAR100, TinyImagenet200
  • ResNet10 on CIFAR10, CIFAR100

Includes checkpoints for the following NBDTs, trained with induced hierarchies:

  • WideResNet28x10 on CIFAR10, CIFAR100, TinyImagenet200
  • ResNet18 on CIFAR10, CIFAR100, TinyImagenet200

Includes checkpoints for the following NBDTs, trained with the wordnet hierarchy (not currently load-able using --pretrained. Use --path-resume instead, after downloading manually):

  • WideResNet28x10 on CIFAR10

Note that we initialized with pytorchcv's pretrained WideResNet28x10 when training NBDTs on CIFAR10 and CIFAR100.