This repository was made to reproduce results of our paper Multiband galaxy morphologies for CLASH: a convolutional neural network transferred from CANDELS (currently accepted to Publications of the Astronomical Society of the Pacific)
To reproduce our results you have to install following python3 packages:
- numpy==1.15.2
- pandas==0.23.4
- matplotlib==2.2.3
- Keras==2.2.2
- scipy==1.1.0
- astropy==2.0.8
- Theano==1.0.4
- Important: This section provides all you need to reproduce our results from scratch. If you want to use our model to predict morphologies in your own data, please go to the next section.
We trained a baseline model on CANDELS images and transferred it to CLASH images. In order to reproduce our results, the first step is to download CANDELS/CLASH data. You must have images and labels for each survey.
Our baseline model was trained using HST images from a CANDELS GOODS-S field (Giavalisco et al. 2004), taken with WFC3 in the F160W band from Hubble Legacy Fields (HLF) Data Release 1.5 for the GOODS-S region (HLF-GOODS-S). We used galaxies from Kartaltepe et al. 2015 catalog, selecting galaxies with F160W magnitudes Hmag < 24.5, which corresponds to the flux limit for reliable visual morphological classifications. We created postage-stamp images from the GOODS-S mosaic setting the size to four times the Petrosian radius as reported in the catalog of Guo et al. 2013.
For the transfer learning, we used images from the CLASH Multi-Cycle Treasury program Postman et al. 2012. CLASH observed 25 clusters of galaxies in up to 16 filters, namely F225W, F275W, F336W, F390W, F435W, F475W, F606W, F625W, F775W, F814W, F850W, F105W, F110W, F125W F140W and F160W, covering the ultraviolet (UV), optical (OPT) and near-infrared (NIR) regions of the spectrum. Molino et al. 2017 published accurate multiwavelength photometric catalogs for these clusters which also provide the Petrosian radius. With this data, we created postage-stamp using mosaics from MAST for each filter separately following the same criterion for the magnitude cut and the size that we adopted for CANDELS ending up with a sample of 68, 531 galaxies.
Both CANDELS and CLASH data used in this work can be downloaded here
Once you have CANDELS/CLASH images, you must put stamps in following path './CANDELS/stamps/' or './CLASH/stamps/' and excecute train.py -dataset-, where -dataset- represent either CANDELS or CLASH. If you want to train only using CLASH data, you have to download CANDELS weights in order to initialize the model.
If you want to test the model using your own CLASH images without training a new model, you must initialize the algorithm using parameters presented in our paper, downloading them here. Once parameters are downloaded, you have to place your CLASH images in './galaxies_to_predict/' and follow example provided in Predict_example.ipynb
You can download the catalog presented in our paper here.
If you have a question do not hesitate to contact us at [email protected].
- Manuel Pérez Carrasco - Msc. student, department of Computer Science, University of Concepción - Github
- Guillermo Cabrera Vives - Assistant professor, department of Computer Science, University of Concepción
- Monserrat Martinez Marín - Msc. student, department of Astronomy, University of Concepción
- Pierluigi Cerulo - Postdoctoral fellow, department of Astronomy, University of Concepción
- Ricardo Demarco - Assistant professor, department of Astronomy, University of Concepción
- Pavlos Protopapas - Scientific director, Institute for applied computational sciences, Harvard University
- Julio Godoy - Assistant professor, department of Computer Science, University of Concepción
- Marc Huertas-Company - Assistant professor, department of Astronomy, University Paris Diderot