Overview

TransferZ-Images is a machine learning ready dataset of 100,442 galaxy images and photometry from the Hyper Suprime-Cam (HSC) second public data release in the 2 sq. deg. COSMOS field. In addition, it contains redshift measurements dervied in COSMOS2020 using up to 30 band photometry (Weaver et. al. 2022). This dataset extends from TransferZ (created by Soriano et. al. 2024). This dataset is submitted along with the NeurIPS ML4PS 2025 submission titled "Combining datasets with different ground truths using Low-Rank Adaptation to generalize image-based CNN for photometric redshift prediction." This dataset is used to test the first implementation of Low Rank Adapation (LoRA) for fine-tuning a CNN on different sources of ground truth. TransferZ serves as the dataset to train the baseline dataset. We provide the same train test splits used in the paper.

Features

•  Photometry measurement and images for 100,442 galaxies in five photometric bands (g: 4754 \AA, r: 6175 \AA, i: 7711 \AA, z: 8898 \AA, y: 9762 \AA)
• Photometric redshifts for each galaxy derived from 35-band photometry (see Weaver+22). Redshifts range from 0.01 to 4
• Images are provided in 64x64 pixel format

Citation

If you make use of these products, please cite this repository. In addition, for COSMOS2020 data products, please cite Weaver et. al. (2022; for HSC-PDR2, please cite Aihara et. al. (2019).

References:

1. Aihara et al. (2019). PASJ, 71, 6. [arXiv:1905.12221][doi]
2. Weaver et al. (2022).ApJS, 258, 11. [arXiv:2110.13923][doi]