Applicability of Avalanche library to regression problem?

[highclef]

Hi all,
I read a post #407 and it seems like Avalanche is suitable for classification problems, not for regression problems. Also the 91th line of base_strategy.py uses CrossEntropyLoss(), which is actually for classification : https://github.com/ContinualAI/avalanche/blob/master/avalanche/training/strategies/base_strategy.py

Here are my questions:

1. Can I still use Avalanche for regression problem?
2. I used MSELoss() for my regression problem and here is the code:

# define optimizer and loss function for training
optimizer = SGD(model.parameters(), lr=0.001, momentum=0.9)
criterion = MSELoss()

# Instantiate Continual learning strategy
strategy = Naive(
     model, optimizer, criterion, train_mb_size=32, train_epochs=2,
     eval_mb_size=32)

Does base_strategy overwrite my criterion and applies CrossEntropyLoss() automatically, no matter which criterion I defined, or am I still having MSELoss() on my result?
3. According to https://pytorch.org/docs/stable/generated/torch.nn.MSELoss.html, MSELoss() measures the mean squared error. But what does "Accuracy" mean in this case?

{'Loss_Epoch/train_phase/train_stream/Task000': 0.19414528276603987,
  'Loss_Exp/eval_phase/test_stream/Task000/Exp000': 0.24790968985511008,
  'Loss_Exp/eval_phase/test_stream/Task000/Exp001': 0.2299130112751,
  'Loss_Exp/eval_phase/test_stream/Task000/Exp002': 0.2401025802251622,
  'Loss_Stream/eval_phase/test_stream/Task000': 0.23930842711845743,
  'Top1_Acc_Epoch/train_phase/train_stream/Task000': 0.35983479607640684,
  'Top1_Acc_Exp/eval_phase/test_stream/Task000/Exp000': 0.2794386821232459,
  'Top1_Acc_Exp/eval_phase/test_stream/Task000/Exp001': 0.16015863331299574,
  'Top1_Acc_Exp/eval_phase/test_stream/Task000/Exp002': 0.2684563758389262,
  'Top1_Acc_Stream/eval_phase/test_stream/Task000': 0.2360178970917226}

It doesn't look like Acc = 1 - Loss. What is calculcated for Acc in this case and how should I interpret this result? Should I just accept the losses and ignore accs?

[AndreaCossu]

Hi @highclef ! Yes, you can use Avalanche for regression. The MSELoss you are passing is used as loss during the strategy execution, so it should work fine: the BaseStrategy does not override your loss with CrossEntropyLoss.

Currently, the default logger used by the strategy returns the loss and the accuracy. In your case, you can either discard the accuracy values (since they are not meaningful for your problem) or, better, feed to the strategy your custom EvaluationPlugin instance with the metrics you want (take a look at this tutorial if you are not sure how to do this).

Hope this helps!