A brain-inspired version of generative replay for continual learning with deep neural networks (e.g., class-incremental learning on CIFAR-100; PyTorch code).
Hi is it possible to allow an option for users to input their own dataset into the model rather than the default CIFAR10 and MNIST? If not, is there a minimal way that we can modify the scripts to accommodate our own datasets, especially inputs that are not image based , eg 1D arrays? Thanks!!!
Hi, I want to ask about the pre-trained models. The paper says you pre-train on cifar10, but the default pre-trained model, in the store/models/ file, C3-5×16-bn, seems a model of pre-training on cifar100. I have run main_pretrain.py on cifar10 to change the pre-trained model and it worked worse than the default. Did you use pre-trained model of cifar100 rather than cifar10 on paper's experiments?
I am trying to reproduce results on Class-IL for EWC and LWF for 10 tasks. I saw in your article that for 10 classes LWF and EWC get higher accuracy than I am able to achieve. Am I doing something wrong?
Hi, I refer to Fig 4 of the main paper, where it states "As a control, also shown is a variant of generative replay whereby the networks are reinitialized before each new task/episode". By reinitialization, does this mean that if we have that model after trained on task1, if we have to train on the next task, then we create a "brand new" model with the exact same architecture and retrain? If it is, then how different will it be from retraining from the current model? Is it likely that there will be no difference in performance with and without reinitialization for other cases because I tried on other types of dataset and the 2 curves kind of overlap rather than show one is better than the other?
Hi,
When I ran compare_permMNIST100_bir.py, and set --seed=12 --n-seeds=5.
Then an error occureed,
File "/media/wuya/DATA/Code/continual_learning/brain-inspired-replay/options.py", line 350, in set_defaults
args.xdg_prop = 0. if args.scenario=="task" and args.xdg_prop is None else args.xdg_prop
AttributeError: 'Namespace' object has no attribute 'xdg_prop'
I don't know where to set xdg_prop, can you tell me how to do it?
Thank you very much!
Hi, I am learning on your codes for a week and I have run several experiments so far. I have some questions.
The setting of 'scenario='task'' works well, but the other one of 'scenario=class' is what I am confused with :
For example, if I set a setting: 'expri..=splitMNIST, tasks=5, scenario=class ' , the result is "-task1: xx. xx -task2: xx. xx task3: xx. xx -task4: xx. xx -task5: xx. xx". I don't understand what it means as the result format is the same with 'scenario=task'. Does the classifier uses single-head or multi-head for class-IL ?
In my experiments, EWC, SI, LwF had no effect on class-IL , all forgot. Is it normal?
Hi, I am really amazed by the number of options that are being provided in the code to run the experiments. I was wondering how to run the baselines, that is, Joint Training and None [simple sequential learning], as mentioned in the article. What is the code and the parameters to run these baseline algorithms seperately, like we can run algorithms like GR, LWF, etc?