Best_params for tinyimagenet

[bhat-prashant]

Hi, I would like to reproduce TinyImageNet results. But, I do not see best hyperparameters for TimyImageNet. Please update them. Thanks in advance,

[lijincm]

Thanks for your interest in our work! Firstly, following CoCa [a], to extract the complete features, we replace the rotate_img function in ./models/ocdnet.py with the following form:

• [rotate_img ]

            def rotate_img(img, s):
                  if s//4 == 0:
                      transform = transforms.RandomResizedCrop(size=(64, 64), scale=(0.66, 0.67), ratio = (0.99,1.00))
                  elif s // 4 == 1:
                      transform = transforms.RandomResizedCrop(size=(64, 64), scale=(0.99, 1.00), ratio=(0.66, 0.67))
                  elif s // 4 == 2:
                      transform = transforms.RandomResizedCrop(size=(64, 64), scale=(0.99, 1.00), ratio=(1.32, 1.33))
                  else:
                      transform = lambda x: x
                  img = transform(img)
                  return torch.rot90(img, s%4, [-1, -2])

  Secondly, Bernoulli_probability is an extremely sensitive hyperparameter for Tiny-ImageNet, the reference values are as follows:

• [hyperparameters ]

              'aod': {
                  200: {
                    'Bernoulli_probability': 0.04,
                   'ER_weight': 0.5,
                   'lr': 0.1,
                   'minibatch_size': 32,
                   'batch_size': 32,
                   'n_epochs': 50,
                  },
                  500: {
                    'Bernoulli_probability': 0.05,
                   'ER_weight': 0.5,
                   'lr': 0.1,
                   'minibatch_size': 32,
                   'batch_size': 32,
                   'n_epochs': 50,
                  },
                  5120: {
                    'Bernoulli_probability': 0.07,
                   'ER_weight': 0.5,
                   'lr': 0.1,
                   'minibatch_size': 32,
                   'batch_size': 32,
                   'n_epochs': 50,
                  }
              },

  [a] Complementary Calibration: Boosting General Continual Learning with Collaborative Distillation and Self-Supervision [J]. IEEE Transactions on Image Processing, 2022. (Online available about 2022.12.31)