Loading a dataset from a task does not (seem to) allow the GEOBENCH_DIR to be overwritten

[gabrieltseng]

Hi -

Thanks for this benchmark !

When I load tasks using geobench.task_iterator, I can pass a benchmark_dir which overrides the default benchmark directory when I load my tasks. However, when I then use a task to load the dataset this is ignored.

Specifically, if I have my geobench data in some_dir (where some_dir != $GEO_BENCH_DIR), then this doesn't work:

my_dataset= "m-eurosat"
for task in geobench.task_iterator(benchmark_name=my_dataset, benchmark_dir=some_dir):
    if task.dataset_name == my_dataset:
        break
dataset = task.get_dataset(split=self.split, partition_name=self.partition) # this fails with "dataset_dir does not exist"

Is there a way I can also load my dataset from some_dir?