pierrot0
commented
1 year ago Hey, thanks for the request. At the moment it's unlikely that we will get any bandwidth to work on this though. However if you wish to send a PR, we will look at it.
From a quick look at it, one would need to download the identities file from around https://github.com/tensorflow/datasets/blob/master/tensorflow_datasets/datasets/celeb_a/celeb_a_dataset_builder.py#L93, then read that file and do the stitching around https://github.com/tensorflow/datasets/blob/master/tensorflow_datasets/datasets/celeb_a/celeb_a_dataset_builder.py#L180. Then the test and checksums file would need to be updated.
OmkarBorhade98
ccl-core
andergalvaan
Maybe I am missing something, so correct me if I am wrong, but I noticed that there is no identity attribute in the tfds 'celeb_a' dataset, despite it being mentioned in the dataset description. I quote from here: "CelebA has large diversities, large quantities, and rich annotations, including - 10,177 number of identities, - 202,599 number of face images, and - 5 landmark locations, 40 binary attributes annotations per image."
I do not know if this is intended or not, anyway I find all these tfds datasets very convenient, but I also need to use the identity label for my use case. Would it be possible to add an identity index to the current set of features? Another idea would be to directly add the original file name to the attributes (the identity can be inferred from that), as it was suggested, among other things, here .