CIFAR weights about isonet HOT 2 CLOSED

Hi @juancprzs,

Thanks for your interest.

That part is directly adopted from the original pycls code (acknowledged in the file header). For ISONet on CIFAR, we didn't carefully explore the architecture & hyperparameter settings and do not report experiments in our paper.

Currently, I don't have a plan to carefully train some models on CIFAR, but if you are interested, maybe the following config can be helpful as a baseline setting (ResNet110).

DATASET:
  NAME: CIFAR10
  NUM_CLASSES: 10
SOLVER:
  BASE_LR: 0.005
  LR_MILESTONES: [120, 240]
  MAX_EPOCHS: 300
  TRAIN_BATCH_SIZE: 128
  WARMUP: True
  WARMUP_EPOCH: 10
  WARMUP_FACTOR: 0.1
ISON:
  DEPTH: 110
  ORTHO_COEFF: 1e-4
  HAS_BN: False
  HAS_ST: False
  SReLU: True

from isonet.

Thanks for the quick response. I'll give it a try.

from isonet.