task_incremental can't use "JointTraining" strategy as multitask

[zhaoxiongjun]

🐛 Describe the bug I use the avalanche/examples/task_incremental.py code, and change the strategy to "JointTraining", the follow error is occurs: AttributeError: 'AvalancheSubset' object has no attribute 'dataset'

🐝 Expected behavior When I commented the code model = as_multitask(model, 'classifier'), the problem disappeared. But I don't think this is the result I expected. How should the upper bound of the task increment be defined base Avalanche?

🐞 Screenshots

🦋 Additional context this is my code. ` from future import absolute_import from future import division from future import print_function

import argparse import torch from torch.nn import CrossEntropyLoss from torch.optim import Adam

from avalanche.benchmarks.classic import SplitCIFAR10 from avalanche.models import SimpleMLP, as_multitask from avalanche.training.supervised import JointTraining, Cumulative

device = "cpu"

model = SimpleMLP(input_size=32 32 3, num_classes=10) model = as_multitask(model, 'classifier')

scenario = SplitCIFAR10(n_experiences=5, return_task_id=True) train_stream = scenario.train_stream test_stream = scenario.test_stream

optimizer = Adam(model.parameters(), lr=0.01) criterion = CrossEntropyLoss()

strategy = JointTraining( model=model, optimizer=optimizer, criterion=criterion, train_mb_size=128, train_epochs=3, eval_mb_size=128, device=device, )

strategy.train(train_stream) strategy.eval(test_stream) `

[HamedHemati]

Hi @zhaoxiongjun !

I tested your code with the master branch, and it works fine. It seems like this was an issue in version 0.2.0.

Which version are you using?

[zhaoxiongjun]

Oh yes, it is indeed a version issue. I am playing with it on Google Colaboratory from official documentation. The version installed inside is 0.2.0.

[zhaoxiongjun]

Hi, @HamedHemati , Sorry, I have another question to ask you, that is, for the task incremental scenario (each task has a different label category), is it possible to use the "JointTraining" strategy and with the "as_multitask" function as the upper bound of performance.

[HamedHemati]

@zhaoxiongjun that's a common practice to use JointStrategy as an upper-bound in task-incremental setups. However, some researchers may consider it as a "soft" upper bound.