Problem with Benchmark Creation after Upgrade to 0.3.1

[jannishag]

Hi all,
I am experiencing the following error when creating an nc_benchmark using avalanche-lib 0.3.1:
AttributeError: ‘AvalancheDataset’ object has no attribute ‘targets’

I did not experience this using avalanche-lib 0.2.1, so maybe there is an issue with the update?
I would appreciate any help on that!

Here is the traceback:

---> 68 scenario = nc_benchmark(
69 Train_List, Test_List, n_experiences=1, one_dataset_per_exp=True, shuffle=False, seed=1234, task_labels=True
70 )

/usr/local/lib/python3.8/dist-packages/avalanche/benchmarks/generators/benchmark_generators.py in nc_benchmark(train_dataset, test_dataset, n_experiences, task_labels, shuffle, seed, fixed_class_order, per_exp_classes, class_ids_from_zero_from_first_exp, class_ids_from_zero_in_each_exp, one_dataset_per_exp, train_transform, eval_transform, reproducibility_data)
194 seq_test_dataset,
195 mapping,
--> 196 ) = concat_datasets_sequentially(train_dataset, test_dataset)
197
198 if one_dataset_per_exp:

/usr/local/lib/python3.8/dist-packages/avalanche/benchmarks/utils/utils.py in concat_datasets_sequentially(train_dataset_list, test_dataset_list)
202
203 # Obtain the number of classes of each dataset
--> 204 classes_per_dataset = [
205 _count_unique(
206 train_dataset_list[dataset_idx].targets,

/usr/local/lib/python3.8/dist-packages/avalanche/benchmarks/utils/utils.py in (.0)
204 classes_per_dataset = [
205 _count_unique(
--> 206 train_dataset_list[dataset_idx].targets,
207 test_dataset_list[dataset_idx].targets,
208 )

AttributeError: 'AvalancheDataset' object has no attribute 'targets'

EDIT:
From the API Docs I read that the attribute 'targets' was removed from AvalancheDatasets in the new release. Thus, I guess the utils.py has to be updated, right?

[AntonioCarta]

Hi @jannishag, can you give more details on your data? Without knowing your dataset we can't say anything about the error.

From the API Docs I read that the attribute 'targets' was removed from AvalancheDatasets in the new release. Thus, I guess the utils.py has to be updated, right?

We didn't remove the attribute, we made it optional. You can use avalanche.benchmarks.utils.make_classification_dataset to easily instantiate datasets with class and task labels. This is equivalent to the old AvalancheDataset, while the new AvalancheDataset is a generic object which doesn't necessarily have targets.

[jannishag]

Ah, alright! Thanks Antonio for the hint, I got it now.

If anyone experiences the same issue:
I did create my AvalancheDataset directly from a TensorDataset (code like, it was AvalancheDataset(TensorDataset)). Using avalanche.benchmarks.utils.make_classification_dataset instead of casting solved the problem.