MetaModelInc module offline training should call eval for validation predict.

🐛 Bug Description

To Reproduce

Steps to reproduce the behavior:

1. create any module with bn or dropout layer
2. run offline training
3. inloop training missed model.eval(), caused batch normalization or dropout layer cannot work as predict model.

Expected Behavior

prediction should not dropout any value or using dynamic batch mean and var for batch normalization. The run_task function should be modified from :

def run_task(self, meta_input, phase):
""" A single naive incremental learning task """
.........
with torch.no_grad():
pred = self.framework(meta_input["X_test"].to(self.framework.device), None)
return pred.detach().cpu().numpy()
to
def run_task(self, meta_input, phase):
""" A single naive incremental learning task """
.........
with torch.no_grad():
self.framework.eval()
pred = self.framework(meta_input["X_test"].to(self.framework.device), None)
return pred.detach().cpu().numpy()

Start by locating MetaModelInc's run_task method and reproduce offline training with a module containing batch normalization or dropout. Verify the validation prediction path switches the framework to evaluation mode and confirm predictions no longer apply dropout or training-time batch statistics.