Trying to reproduce continual learning for Human State Monitoring with Avalanche 0.4 (NIScenario - which benchmark generator to use?)

[KristienR]

Hello everyone, I tried to make the code "Continual learning for Human State Monitoring" working with Avalanche 0.4 for the Wesad dataset. noticed that the original code was probably implemented with an avalanche version prior to 0.3. The benchmark dataset_benchmark, apparantly used to have AvalancheDatasets as input parameters, but that does not work in Avalanche 0.4.0 I tried to replace the benchmark generator in the original wesad_train_test.py by as wel the ni_benchmark as the new dataset_benchmark, but neither worked . (see code sample below). The niscenario seems to concatenate the different experiences in one dataset, but there seems to go something wrong, resulting in out of bound errors. (this works with only one experience, but that is obviously not the point of continual learning). (I extended WesadTsSet to be able to have Testsets of the selected pairs only for this). The clscenario has following runtime error "File "C:\Users\romboukr.conda\envs\Python37\lib\site-packages\torch\nn\functional.py", line 3026, in cross_entropy
return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index, label_smoothing)
RuntimeError: expected scalar type Long but found Int"
ds0 = WESADTrSet(pair=0)
ds1 = WESADTrSet(pair=1)
dst= WESADTsSet()
dst0 = WESADTsSet(pair=0)
dst1: WESADTsSet = WESADTsSet(pair=1)

traininglist= [ds0,ds1]
testlist = [dst0,dst1]
niscenario = ni_benchmark(traininglist, testlist,n_experiences=len(traininglist), task_labels=True)
clscenario = dataset_benchmark (traininglist, [dst], complete_test_set_only=True)

I would like to use Avalanche for my thesis regarding continual learning for freeze of gait, my data also consists of timeseries off accelerometerdata. any help would be much appreciated!