Questions about reproducing about cls-er HOT 1 CLOSED

Hi, Thank you for your interest in our work. Are you trying to reproduce the results for CLS-ER? I see from the command that you are running the code for LwF.

On the same dataset, you can run for our method using the following command:
python main.py --dataset seq-mnist --model clser --buffer_size 500 --load_best_args

For the mode general queries, yes task5 results show the final accuracy of the method on all the tasks trained for seq MNIST, for Rot-MNIST for instance it would be task20.

Our method does not use a pre-trained model, in the mammoth repo, the accuracies before training are utlized to calculate the fwd transfer of the model.

Also, we do not use specific seeds in order to avoid introducing bias in evaluation, similar to DER, we run the method 10 different random seeds (this would be the case if you do not set a particular seed) and provide the average performance and standard deviation.

FTI, when you run for our method, you will get the evaluation results for all three models, working model, stable model and the plastic model. For inference we use the stable model, Table S1 in the appendix provides the results for all the three components for the considerd CL settings.

from cls-er.