astromlp-models

Deep learning models for astrophysics applications.

Topic: Galaxies

Collection of Deep Learning models to characterize different proprieties of galaxies, based on data from the Sloan Digital Sky Survey (SDSS).

The SDSS Galaxy Subset dataset is used to train and explore available models (the default location of the dataset w.r.t. to this repository is ../sdss-gs).

Models

The following table quickly describes single input/output models available.

Model	Input	Output	Type	Description
i2r	img	redshift	regression	infer object redshift from RGB image
f2r	fits	redshift	regression	infer object redshift from FITS data
s2r	spectra	redshift	regression	infer object redshift from spectra data
ss2r	ssel	redshift	regression	infer object redshift from selected spectra data
b2r	bands	redshift	regression	infer object redshift from bands data
w2r	wise	redshift	regression	infer object redshift from WISE data
i2sm	img	smass	regression	infer object stellar mass from RGB image
f2sm	fits	smass	regression	infer object stellar mass from FITS data
s2sm	spectra	smass	regression	infer object stellar mass from spectra data
ss2sm	ssel	smass	regression	infer object stellar mass from selected spectra data
b2sm	bands	smass	regression	infer object stellar mass from bands data
w2sm	wise	smass	regression	infer object stellar mass from WISE data
i2s	img	subclass	classification	infer object sub-class from RGB image
f2s	fits	subclass	classification	infer object sub-class from FITS data
s2s	spectra	subclass	classification	infer object sub-class from spectra data
ss2s	ssel	subclass	classification	infer object sub-class from selected spectra data
b2s	bands	subclass	classification	infer object sub-class from bands data
w2s	wise	subclass	classification	infer object sub-class from WISE data
i2g	img	gz2c	classification	infer Galaxy Zoo 2 simplified class from RGB image
f2g	fits	gz2c	classification	infer Galaxy Zoo 2 simplified class from FITS data
s2g	spectra	gz2c	classification	infer Galaxy Zoo 2 simplified class from spectra data
ss2g	ssel	gz2c	classification	infer Galaxy Zoo 2 simplified class from selected spectra data
b2g	bands	gz2c	classification	infer Galaxy Zoo 2 simplified class from bands data
w2g	wise	gz2c	classification	infer Galaxy Zoo 2 simplified class from WISE data

The following table quickly summarizes the multi-input/output models available.

Model	Input(s)	Output(s)
fSbW2rSM	fits, spectra, bands, wise	redshift, smass
fSbW2sG	fits, spectra, bands, wise	subclass, gz2c
iFsSSbW2r	img, fits, spectra, ssel, bands, wise	redshift
iFsSSbW2sm	img, fits, spectra, ssel, bands, wise	smass
iFsSSbW2s	img, fits, spectra, ssel, bands, wise	subclass
iFsSSbW2g	img, fits, spectra, ssel, bands, wise	gz2c
iFsSSbW2rSMsG	img, fits, spectra, ssel, bands, wise	redshift, smass, subclass, gz2c

Inputs & Outputs

The following tables describe the data used as inputs for the models, and the outputs for regression and classification models.

Input	Shape	Description
img	150x150x3	RGB image from the object in JPEG format, 150x150 pixels, generated using the SkyServer DR16 API
fits	61x61x5	FITS data subset around the object across the u, g, r, i, z bands; cut is done using the ImageCutter library
spectra	3225x1	full best fit spectra data from SDSS between 4000 and 9000 wavelengths
ssel	1225x1	best fit spectra data from SDSS for specific selected intervals of wavelengths discussed by Sánchez Almeida 2010
bands	5x1	photometric values from SDSS data: [modelMag_u, modelMag_g, modelMag_r, modelMag_i, modelMag_z] (data table details)
wise	4x1	list of bands values from WISE data: [w1mag, w2mag, w3mag, w4mag] (data table details)

Output	Type	Description
redshift	regression	final redshift from SDSS data z (data table details)
subclass	classification	subset of sub-class from SDSS data for the galaxy objects subClass (data table details)
smass	regression	stellar mass extracted from the eBOSS Firefly catalog
gz2c	classification	simplified version of the Galaxy Zoo 2 classification, from Willett et al 2013 (see class sets section below for details)

Class Sets

The following tables describe the class set labels for the classification outputs.

SDSS sub-class subset (subclass)

The sub-class parameter for each object is available from the SDSS spectroscopic catalogs.

Label	Description
AGN	has detectable emission lines that are consistent with being a Seyfert or LINER
BROADLINE	has lines detected at the 10-sigma level with sigmas > 200 km/sec at the 5-sigma level
STARBURST	galaxy is star-forming
STARFORMING	has detectable emission lines that are consistent with star-formation criteria

Galaxy Zoo 2 Simplified Classes (gz2c)

This is a simplified version of the Galaxy Zoo 2 classification tree, detailed information on the original data is available here.

Label	Description
A	artifact, star
Ec	smooth, cigar-shaped
Ei	smooth, in-between
Er	smooth, completely round
SBa	with features/disks, has bar, dominant bulge prominence
SBaR	with features/disks, has bar, dominant bulge prominence, has spiral structure
SBb	with features/disks, has bar, obvious bulge prominence
SBbR	with features/disks, has bar, obvious bulge prominence, has spiral structure
SBc	with features/disks, has bar, just noticeable bulge prominence
SBcR	with features/disks, has bar, just noticeable bulge prominence, has spiral structure
SBd	with features/disks, has bar, no bulge prominence
SBdR	with features/disks, has bar, no bulge prominence, has spiral structure
Sa	with features/disks, dominant bulge prominence
SaR	with features/disks, dominant bulge prominence, has spiral structure
Sb	with features/disks, obvious bulge prominence
SbR	with features/disks, obvious bulge prominence, has spiral structure
Sc	with features/disks, just noticeable bulge prominence
ScR	with features/disks, just noticeable bulge prominence, has spiral structure
Sd	with features/disks, no bulge prominence
SdR	with features/disks, no bulge prominence, has spiral structure
Seb	with features/disks, edge-on, boxy bulge
Sen	with features/disks, edge-on, no bulge
Ser	with features/disks, edge-on, round bulge

Fitting Models and Visualizing Metrics

The models available in this repository are implemented using Keras. To fit the models available in this repository the astromlp Python package that includes all the helper classes is also required.

You can fit a model using mlflow, for example to fit the i2r model using your current python (i.e. don't create a new environment using conda) you can run from the repository directory, and also include this run in the i2r experiment:

$ mlflow run i2r --experiment-name i2r --env-manager=local

You can also change the parameters to run the model, namely the number of epochs, the batch size, the loss function and optimizer to use, for example:

$ mlflow run i2r -P epochs=10 -P batch_size=32 -P loss=mse -P optimizer=adam --experiment-name i2r --no-conda

You can also change the location of the dataset by setting the ds parameter:

$ mlflow run i2r -P ds=/tmp/sdss-gs --experiment-name i2r --no-conda

To view the data concerning the fitting of the available models you can use mlflow user interface:

$ mlflow ui --backend-store-uri sqlite:///mlruns.db

To fit models and include data in this database you can set the MLFLOW_TRACKING_URI environment variable to this file, remember to use an absolute path, for example:

$ export MLFLOW_TRACKING_URI=sqlite:////home/nrc/astromlp-models.git/mlruns.db

And can also check the generated tensorboard logs, for example:

$ tensorboard --logdir i2r/logs/

Single Input/Output Models Hyper-parameters Exploration

The following table summarizes different combinations of batch size (32 or 64), optimizer (RMSProp or Adam) and loss functions (Mean Squared Error or Mean Absolute Error for regression, categorical crossentropy is always used for multi-label classification) hyper-parameters exploration, the score is the best validation value for the validation dataset. The highlighted row describes the final hyper-parameters used for bootstrapping each model and the corresponding evaluation score (accuracy for classification models) on the test set, never seen by any model during the exploration steps.

Model	Input	Output	Epochs	Batch Size	Optimizer	Loss	Score
i2r	img	redshift	10	32	RMSProp	MSE	0.003365
i2r	img	redshift	10	32	RMSProp	MAE	0.030170
i2r	img	redshift	10	32	Adam	MSE	0.016661
i2r	img	redshift	10	32	Adam	MAE	0.071612
i2r	img	redshift	10	64	RMSProp	MSE	0.003508
i2r	img	redshift	10	64	RMSProp	MAE	0.030480
i2r	img	redshift	10	64	Adam	MSE	0.003498
i2r	img	redshift	10	64	Adam	MAE	0.071558
i2r	img	redshift	20	32	RMSProp	MSE	0.002377
f2r	fits	redshift	10	32	RMSProp	MSE	0.001930
f2r	fits	redshift	10	32	RMSProp	MAE	0.027434
f2r	fits	redshift	10	32	Adam	MSE	0.001917
f2r	fits	redshift	10	32	Adam	MAE	0.024896
f2r	fits	redshift	10	64	RMSProp	MSE	0.002075
f2r	fits	redshift	10	64	RMSProp	MAE	0.028206
f2r	fits	redshift	10	64	Adam	MSE	0.001870
f2r	fits	redshift	10	64	Adam	MAE	0.024997
f2r	fits	redshift	20	64	Adam	MSE	0.002020
s2r	spectra	redshift	10	32	RMSProp	MSE	0.004649
s2r	spectra	redshift	10	32	RMSProp	MAE	0.033334
s2r	spectra	redshift	10	32	Adam	MSE	0.006617
s2r	spectra	redshift	10	32	Adam	MAE	0.040978
s2r	spectra	redshift	10	64	RMSProp	MSE	0.005554
s2r	spectra	redshift	10	64	RMSProp	MAE	0.038496
s2r	spectra	redshift	10	64	Adam	MSE	0.014275
s2r	spectra	redshift	10	64	Adam	MAE	0.044180
s2r	spectra	redshift	20	32	RMSProp	MSE	0.004401
ss2r	ssel	redshift	10	32	RMSProp	MSE	0.004343
ss2r	ssel	redshift	10	32	RMSProp	MAE	0.030403
ss2r	ssel	redshift	10	32	Adam	MSE	0.008098
ss2r	ssel	redshift	10	32	Adam	MAE	0.045030
ss2r	ssel	redshift	10	64	RMSProp	MSE	0.005174
ss2r	ssel	redshift	10	64	RMSProp	MAE	0.031783
ss2r	ssel	redshift	10	64	Adam	MSE	0.006342
ss2r	ssel	redshift	10	64	Adam	MAE	0.053812
ss2r	ssel	redshift	20	32	RMSProp	MSE	0.003700
b2r	bands	redshift	10	32	RMSProp	MSE	0.003765
b2r	bands	redshift	10	32	RMSProp	MAE	0.027912
b2r	bands	redshift	10	32	Adam	MSE	0.004437
b2r	bands	redshift	10	32	Adam	MAE	0.027528
b2r	bands	redshift	10	64	RMSProp	MSE	0.004726
b2r	bands	redshift	10	64	RMSProp	MAE	0.029070
b2r	bands	redshift	10	64	Adam	MSE	0.004593
b2r	bands	redshift	10	64	Adam	MAE	0.027921
b2r	bands	redshift	20	32	RMSProp	MSE	0.002806
w2r	wise	redshift	10	32	RMSProp	MSE	0.011475
w2r	wise	redshift	10	32	RMSProp	MAE	0.055713
w2r	wise	redshift	10	32	Adam	MSE	0.011523
w2r	wise	redshift	10	32	Adam	MAE	0.055016
w2r	wise	redshift	10	64	RMSProp	MSE	0.011657
w2r	wise	redshift	10	64	RMSProp	MAE	0.056263
w2r	wise	redshift	10	64	Adam	MSE	0.011557
w2r	wise	redshift	10	64	Adam	MAE	0.055382
w2r	wise	redshift	20	32	RMSProp	MSE	0.012004
i2sm	img	smass	10	32	RMSProp	MSE	40929.859
i2sm	img	smass	10	32	RMSProp	MAE	26.048
i2sm	img	smass	10	32	Adam	MSE	40813.363
i2sm	img	smass	10	32	Adam	MAE	26.695
i2sm	img	smass	10	64	RMSProp	MSE	40982.949
i2sm	img	smass	10	64	RMSProp	MAE	26.736
i2sm	img	smass	10	64	Adam	MSE	42835.566
i2sm	img	smass	10	64	Adam	MAE	26.238
i2sm	img	smass	20	64	Adam	MAE	22.537
f2sm	fits	smass	10	32	RMSProp	MSE	40600.176
f2sm	fits	smass	10	32	RMSProp	MAE	27.171
f2sm	fits	smass	10	32	Adam	MSE	40491.633
f2sm	fits	smass	10	32	Adam	MAE	25.758
f2sm	fits	smass	10	64	RMSProp	MSE	40679.352
f2sm	fits	smass	10	64	RMSProp	MAE	26.600
f2sm	fits	smass	10	64	Adam	MSE	40903.289
f2sm	fits	smass	10	64	Adam	MAE	25.931
f2sm	fits	smass	20	32	Adam	MAE	21.753
s2sm	spectra	smass	10	32	RMSProp	MSE	2963.335
s2sm	spectra	smass	10	32	RMSProp	MAE	20.871
s2sm	spectra	smass	10	32	Adam	MSE	3432.733
s2sm	spectra	smass	10	32	Adam	MAE	20.129
s2sm	spectra	smass	10	64	RMSProp	MSE	3159.474
s2sm	spectra	smass	10	64	RMSProp	MAE	21.317
s2sm	spectra	smass	10	64	Adam	MSE	3518.646
s2sm	spectra	smass	10	64	Adam	MAE	20.420
s2sm	spectra	smass	20	32	Adam	MAE	19.577
ss2sm	ssel	smass	10	32	RMSProp	MSE	40206.445
ss2sm	ssel	smass	10	32	RMSProp	MAE	22.797
ss2sm	ssel	smass	10	32	Adam	MSE	3018.159
ss2sm	ssel	smass	10	32	Adam	MAE	22.728
ss2sm	ssel	smass	10	64	RMSProp	MSE	40271.656
ss2sm	ssel	smass	10	64	RMSProp	MAE	24.099
ss2sm	ssel	smass	10	64	Adam	MSE	40798.273
ss2sm	ssel	smass	10	64	Adam	MAE	22.380
ss2sm	ssel	smass	20	64	Adam	MAE	18.793
b2sm	bands	smass	10	32	RMSProp	MSE	3387.860
b2sm	bands	smass	10	32	RMSProp	MAE	26.523
b2sm	bands	smass	10	32	Adam	MSE	5747.023
b2sm	bands	smass	10	32	Adam	MAE	25.930
b2sm	bands	smass	10	64	RMSProp	MSE	5520.346
b2sm	bands	smass	10	64	RMSProp	MAE	26.694
b2sm	bands	smass	10	64	Adam	MSE	5725.097
b2sm	bands	smass	10	64	Adam	MAE	25.151
b2sm	bands	smass	20	64	Adam	MAE	25.344
w2sm	wise	smass	10	32	RMSProp	MSE	4323.260
w2sm	wise	smass	10	32	RMSProp	MAE	30.034
w2sm	wise	smass	10	32	Adam	MSE	4391.042
w2sm	wise	smass	10	32	Adam	MAE	29.858
w2sm	wise	smass	10	64	RMSProp	MSE	4423.225
w2sm	wise	smass	10	64	RMSProp	MAE	30.605
w2sm	wise	smass	10	64	Adam	MSE	4433.951
w2sm	wise	smass	10	64	Adam	MAE	29.800
w2sm	wise	smass	20	64	Adam	MAE	27.719
i2s	img	subclass	10	32	RMSProp	categorical crossentropy	0.771606
i2s	img	subclass	10	32	Adam	categorical crossentropy	0.778097
i2s	img	subclass	10	64	RMSProp	categorical crossentropy	0.767020
i2s	img	subclass	10	64	Adam	categorical crossentropy	0.787760
i2s	img	subclass	20	64	Adam	categorical crossentropy	0.765625
f2s	fits	subclass	10	32	RMSProp	categorical crossentropy	0.771977
f2s	fits	subclass	10	32	Adam	categorical crossentropy	0.782177
f2s	fits	subclass	10	64	RMSProp	categorical crossentropy	0.766183
f2s	fits	subclass	10	64	Adam	categorical crossentropy	0.778739
f2s	fits	subclass	20	32	Adam	categorical crossentropy	0.785590
s2s	spectra	subclass	10	32	RMSProp	categorical crossentropy	0.765579
s2s	spectra	subclass	10	32	Adam	categorical crossentropy	0.767897
s2s	spectra	subclass	10	64	RMSProp	categorical crossentropy	0.766648
s2s	spectra	subclass	10	64	Adam	categorical crossentropy	0.713914
s2s	spectra	subclass	20	32	Adam	categorical crossentropy	0.763455
ss2s	ssel	subclass	10	32	RMSProp	categorical crossentropy	0.765764
ss2s	ssel	subclass	10	32	Adam	categorical crossentropy	0.768824
ss2s	ssel	subclass	10	64	RMSProp	categorical crossentropy	0.761719
ss2s	ssel	subclass	10	64	Adam	categorical crossentropy	0.768415
ss2s	ssel	subclass	20	32	Adam	categorical crossentropy	0.756076
b2s	bands	subclass	10	32	RMSProp	categorical crossentropy	0.758995
b2s	bands	subclass	10	32	Adam	categorical crossentropy	0.762148
b2s	bands	subclass	10	64	RMSProp	categorical crossentropy	0.757720
b2s	bands	subclass	10	64	Adam	categorical crossentropy	0.762742
b2s	bands	subclass	20	32	Adam	categorical crossentropy	0.753472
w2s	wise	subclass	10	32	RMSProp	categorical crossentropy	0.772348
w2s	wise	subclass	10	32	Adam	categorical crossentropy	0.779303
w2s	wise	subclass	10	64	RMSProp	categorical crossentropy	0.753348
w2s	wise	subclass	10	64	Adam	categorical crossentropy	0.773345
w2s	wise	subclass	20	32	Adam	categorical crossentropy	0.784288
i2g	img	gz2c	10	32	RMSProp	categorical crossentropy	0.177365
i2g	img	gz2c	10	32	Adam	categorical crossentropy	0.179617
i2g	img	gz2c	10	64	RMSProp	categorical crossentropy	0.178693
i2g	img	gz2c	10	64	Adam	categorical crossentropy	0.179261
i2g	img	gz2c	20	32	Adam	categorical crossentropy	0.206522
f2g	fits	gz2c	10	32	RMSProp	categorical crossentropy	0.278153
f2g	fits	gz2c	10	32	Adam	categorical crossentropy	0.355011
f2g	fits	gz2c	10	64	RMSProp	categorical crossentropy	0.273864
f2g	fits	gz2c	10	64	Adam	categorical crossentropy	0.322159
f2g	fits	gz2c	20	32	Adam	categorical crossentropy	0.372283
s2g	spectra	gz2c	10	32	RMSProp	categorical crossentropy	0.268863
s2g	spectra	gz2c	10	32	Adam	categorical crossentropy	0.184685
s2g	spectra	gz2c	10	64	RMSProp	categorical crossentropy	0.267045
s2g	spectra	gz2c	10	64	Adam	categorical crossentropy	0.192330
s2g	spectra	gz2c	20	32	RMSProp	categorical crossentropy	0.225543
ss2g	ssel	gz2c	10	32	RMSProp	categorical crossentropy	0.261824
ss2g	ssel	gz2c	10	32	Adam	categorical crossentropy	0.182151
ss2g	ssel	gz2c	10	64	RMSProp	categorical crossentropy	0.261932
ss2g	ssel	gz2c	10	64	Adam	categorical crossentropy	0.256818
ss2g	ssel	gz2c	20	64	RMSProp	categorical crossentropy	0.259511
b2g	bands	gz2c	10	32	RMSProp	categorical crossentropy	0.246340
b2g	bands	gz2c	10	32	Adam	categorical crossentropy	0.245495
b2g	bands	gz2c	10	64	RMSProp	categorical crossentropy	0.245455
b2g	bands	gz2c	10	64	Adam	categorical crossentropy	0.243466
b2g	bands	gz2c	20	32	RMSProp	categorical crossentropy	0.275815
w2g	wise	gz2c	10	32	RMSProp	categorical crossentropy	0.257320
w2g	wise	gz2c	10	32	Adam	categorical crossentropy	0.262387
w2g	wise	gz2c	10	64	RMSProp	categorical crossentropy	0.258239
w2g	wise	gz2c	10	64	Adam	categorical crossentropy	0.259943
w2g	wise	gz2c	20	32	Adam	categorical crossentropy	0.254076

Acknowledgments

Thank you to Dr. Andrew Humphrey for helping spawning this project and his contributions that helped improve this work.