Hi, thank you for releasing DeepAstroUDA.
I am trying to reproduce the SDSS → DECaLS experiment from the paper, but I noticed several issues regarding the released SDSS dataset on Zenodo.
Specifically:
1sdss_1.h5 and sdss_2.h5 appear to be completely identical (same MD5 hash).
2 The class distributions seem inconsistent with the paper description. For example, one class contains only a few dozen samples, while the paper mentions that most classes should contain 1k–2.6k images, with the smallest class containing 334 images.
3 The paper describes the gravitational lens class as a target-only unknown class, but the released source-domain data appears to contain all 10 classes.
Because of this, I am unsure whether:
- the uploaded SDSS files are duplicated accidentally,
- the released dataset is only a reduced/demo subset,
- or some preprocessing/splitting step is missing from the repository.
Could you clarify which dataset should be used for reproducing the SDSS → DECaLS experiment, or whether the full source-domain dataset is available somewhere else?
Thank you very much.
Hi, thank you for releasing DeepAstroUDA.
I am trying to reproduce the SDSS → DECaLS experiment from the paper, but I noticed several issues regarding the released SDSS dataset on Zenodo.
Specifically:
1
sdss_1.h5andsdss_2.h5appear to be completely identical (same MD5 hash).2 The class distributions seem inconsistent with the paper description. For example, one class contains only a few dozen samples, while the paper mentions that most classes should contain 1k–2.6k images, with the smallest class containing 334 images.
3 The paper describes the gravitational lens class as a target-only unknown class, but the released source-domain data appears to contain all 10 classes.
Because of this, I am unsure whether:
Could you clarify which dataset should be used for reproducing the SDSS → DECaLS experiment, or whether the full source-domain dataset is available somewhere else?
Thank you very much.