Datasets:

---

torchgeo

/

eurosat

like 2

Follow

TorchGeo 133

The dataset has multiple splits, including the original split and a geographic split. There should only be 27k images and 27k labels for each split. Which file are you trying to use?

@ajstewart that is strange, for some reason HF is finding 2 rows for each 1 image.

The HuggingFace website shows the train split has 32.5k rows.

Loading the dataset using the datasets package also shows duplicates for each image:

from datasets import load_dataset
ds = load_dataset("torchgeo/eurosat")
print (ds)

# Output
DatasetDict({
    train: Dataset({
        features: ['text'],
        num_rows: 32460
    })
    validation: Dataset({
        features: ['text'],
        num_rows: 10820
    })
    test: Dataset({
        features: ['text'],
        num_rows: 10820
    })
})

Is it possible that HF is using both eurosat-spatial-train.txt and eurosat-train.txt for the training split?

I've never used the datasets package before so can't comment on that. This dataset is designed for use with torchgeo.datasets, which is well-tested and stable. See https://docs.torchgeo.org/en/stable/api/datasets/eurosat.html for docs.