This is the repository for a Continual Learning(CL) method
Use cosine as the measure in the margin loss, and it aware the sample nums for each class. See CosineLabelAwareMarginLoss in
Loss.py. The method contains cosface and LabelDistributionAwareMarginLoss
Use a Transformer Decoder to replace the pooling and fully connected layers in the network(in this case the ResNet18).
See TransformerDecoderFundus in Transformer.py.
Use a Contrastive Loss to accelerate the process of adapting to the new data. The replay data doing this can be seen as knowledge distillation. And the pseudo feature sampled from each class's distribution can do the similar thing to the Prototype Augmentation
This part is implemented in GeneralizationAwareContinualLearning.py, inside the _compute_loss() function.
Run main.py, there are several settings can be changed directly in main.py.
- dataset_order = [0, 1, 2], 0-ODIR, 1-RR, 2-ISEE
- debug = False, IF TRUE, then only run 1 epoch with several iters, the evaluation may raise error since the prediction may not include all labels
- tasks = 3, maximum tasks, if x only the first x in dataset_order will be used
- load_sub_exp_ckpt = False, if there exists a subset experiment for current ones, load the checkpoint to avoid rerun(e.g. 0->1 is the subset of 0->1->2)
- sample_ratio = 0.1, replay sample ratio for each class
- epochs = 120
- device = 1, no DDP implementation in these codes, if you want, you can modify the code yourself