Datasets:

imagenet-1k

like 309

Hi all,

I recently started using HuggingFace and was excited to try out the HF datasets. However, when I used datasets.load_dataset to load the “ImageNet-1K” dataset with the default configuration, one of the archive files could not be downloaded due to a connection error. When I tried to load the dataset again, the corrupted files were skipped and the next step (generating the train split) was processed, resulting in errors.

As a workaround, I cloned the dataset from HuggingFace using Git LFS and used the local path to load it. However, I encountered another issue where the loading script “imagenet-1k.py” always checked for online data files instead of looking for the downloaded files in ./data. To resolve this, I modified the _DATA_URL to force the loader to look in ./data. This changed the checksum of the dataset and caused it to be placed in a different folder from the default path. As a result, I had to specify the path to “imagenet-1k.py” every time I used the dataset.

Is there a solution for using cloned datasets in the same way as datasets loaded by load_dataset with an internet connection? Any help would be appreciated. Thanks!