conditionwang / fcil Goto Github PK

I was wondering that in the paper when you mention that you add the same number of classes in the clients to perform incremental in classes, the new classes added in the same client, are they of the same category? For e.g. suppose I have a model that predicts dog breeds. Suppose client 1 receives data on breed x and incremental learning is performed with this new data for breed and client 2 received data on breed y, and similarly incremental learning is performed for breed y, so are breed x= breed y or can they even be different breeds?

Hey, I am also researching in the federated class incremental learning domain and found your paper to be very interesting. I tried running the code on A-100 80 GB machine but it throws CUDA out of memory error even with a batch size of 32 instead of 128. Can you tell me about the gpu requirements for training on tiny-imagenet-200?

May I ask if your processing of the tinyimagenet dataset is to separate 50 samples from each class in its training,and then use them as test set, which is the testing effect of tinyimagenet in your paper

why you have initialized 125 GLFC models where they are getting used( confusion is that there are only 30 clients and we have 125 models), how those 125 models are used?

Are you using multiple GPUs? What is the video memory of the GPU used?

May I ask if the learning rate is 2.0 above T=5 and T=20.

In federated learning, the dataset is commonly partitioned among clients. However, I am unable to find the code related to dataset partitioning. It appears that all clients are sharing the same dataset. Could you tell me what I'm missing?