Class-IL Permuted MNIST

[en-tropyc]

I'm trying to set up a class incremental learning scenario using the Permuted MNIST task protocol. I modified the PermutedMNIST() method from cmnist.py so that an NCScenario instance is returned as follows:

return nc_benchmark(
    list_train_dataset, 
    list_test_dataset, 
    n_experiences=len(list_train_dataset), 
    shuffle=False, 
    task_labels=False,
    class_ids_from_zero_from_first_exp=False,
    class_ids_from_zero_in_each_exp=False,
    one_dataset_per_exp=True
)

Below are some of the sampled image/label pairs from each of the resulting experiences. This looks correct to me since at test time the network will only see Task 0 for each different permutation it's confronted with, but it will need to correctly infer the digit and infer permutation it belongs to (ex. digit 7 from permutation 2 for output 27). Am I thinking about this correctly?

Thank you in advance.

[AndreaCossu]

Hi @christam96 ! Yes, by looking at the labels tensor you have the result seems to be correct. You are in a class-incremental setting, just set the output layer of your model to have the appropriate number of units and you are good to go.

[en-tropyc]

Will do, thanks for the help 😃