ICaRL Scenario with Custom Dataset

[Sean3216]

Hi, I'm trying to implement ICaRL on a custom dataset. So far I've successfully used the package for implementing LwF and EWC.

so far, this is how I prepare the dataset for training:

---

xtrainex = torch.tensor(xtrain1.values,dtype = torch.float32)
ytrainex = torch.tensor(ytrain1, dtype = torch.long)

xtestex = torch.tensor(xtest1.values,dtype = torch.float32)
ytestex = torch.tensor(ytest1, dtype = torch.long)

trainsetex = TensorDataset(xtrainex,ytrainex)
testsetex = TensorDataset(xtestex, ytestex)

trainex = make_classification_dataset(trainsetex,task_labels = 0)
testex = make_classification_dataset(testsetex,task_labels = 0)

=====================================================================
xtrainreal = torch.tensor(xtrain2.values,dtype = torch.float32)
ytrainreal = torch.tensor(ytrain2, dtype = torch.long)

xtestreal = torch.tensor(xtest2.values,dtype = torch.float32)
ytestreal = torch.tensor(ytest2, dtype = torch.long)

trainsetreal = TensorDataset(xtrainreal,ytrainreal)
testsetreal = TensorDataset(xtestreal, ytestreal)

trainreal = make_classification_dataset(trainsetreal, task_labels = 1)
testreal = make_classification_dataset(testsetreal, task_labels = 1)

=====================================================================
scenario_custom_task_labels = Generator.dataset_benchmark([trainex,trainreal],
[testex,testreal])

=====================================================================
eval_plugin = EvaluationPlugin(accuracy_metrics(minibatch=True, epoch=True, experience=True, stream=True),
loss_metrics(minibatch=True, epoch=True, experience=True, stream=True),
timing_metrics(epoch=True, epoch_running=True),
forgetting_metrics(experience=True, stream=True),
loggers=[interactive_logger, tb_logger]
)

cl_strategy = ICaRL(model.feature_extractor,
model.classifier,
optimizer=torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9),
device=device,
train_mb_size=64,
train_epochs=20,
eval_mb_size=64,
evaluator=eval_plugin,
memory_size = 32,
fixed_memory = False,
buffer_transform = None)

---

I have several questions as I'm quite new to continual learning.

1. "from avalanche.training import ICaRL" ---> when I try to implement this, I don't really understand how to fill out the buffer_transform parameter. Reading from the documentation https://avalanche-api.continualai.org/en/v0.3.1/generated/avalanche.training.ICaRL.html#avalanche.training.ICaRL, if I understood correctly, it says that it is the transformation that is applied to the buffer elements or the samples for Replay.
   I tried to take a look here https://github.com/ContinualAI/avalanche/blob/master/examples/icarl.py and the parameter is filled with a function to transform the data. I don't use any kind of transformation for my data. Do you have any suggestions regarding what to fill if this is the case?

2. I tried to run ICaRL with the code above (buffer_transform set to just None) but encountered an error because of my benchmark type (I think). The following is the error:
   

it was an error on the data type that pops when I start the training. End of the error says:
AttributeError: 'GenericCLScenario' object has no attribute 'n_classes_per_exp'

Is there a way to convert GenericCLScenario to NCScenario?

Thank you so much! I'm sorry for the long question. I used bold on the important words to help faster reading in case it's too long.
Cheers! Sorry for the trouble.