BLISS User API

Bayesian Light Source Separator (BLISS) is a Bayesian method for deblending and cataloging light sources.

Installation

[ ]:

!pip install -e /home/zhteoh/770-bulk-predict

[ ]:

!pip install bliss-deblender

Tutorial

[6]:

from bliss.api import BlissClient

# bliss_client = BlissClient(cwd="/data/scratch/zhteoh/tutorial")
bliss_client = BlissClient(cwd="/tmp/pytest-of-zhteoh/pytest-417")

Train the model

Generate synthetic image data

[5]:

bliss_client.generate(
    n_batches=3,
    batch_size=64,
    max_images_per_file=128
)

Data will be saved to /data/scratch/zhteoh/tutorial/data/cached_dataset

Simulating images in batches for file: 100%|██████████| 2/2 [04:34<00:00, 137.09s/it]
Simulating images in batches for file: 100%|██████████| 2/2 [04:41<00:00, 140.52s/it]3s/it]
Generating and writing cached dataset files: 100%|██████████| 2/2 [09:15<00:00, 277.74s/it]

Pass additional custom configuration parameters

[7]:

# Alter default cached_data_path
bliss_client.cached_data_path = "/data/scratch/zhteoh/tutorial/data/cached_dataset_ms0.02"

bliss_client.generate(
    n_batches=3,  # required
    batch_size=64,  # required
    max_images_per_file=128,  # required
    simulator={"survey": {"prior_config": {"mean_sources": 0.02}}},  # optional
    generate={"file_prefix": "dataset"},  # optional
)

Data will be saved to /data/scratch/zhteoh/tutorial/data/cached_dataset_ms0.02

Simulating images in batches for file: 100%|██████████| 2/2 [01:01<00:00, 30.78s/it]
Simulating images in batches for file: 100%|██████████| 2/2 [01:06<00:00, 33.06s/it]7s/it]
Generating and writing cached dataset files: 100%|██████████| 2/2 [02:07<00:00, 63.95s/it]

[ ]:

bliss_client.cached_data_path = "/data/scratch/zhteoh/tutorial/data/cached_dataset_ms0.02"

[ ]:

# Check that the dataset is generated
!ls /data/scratch/zhteoh/tutorial/data/cached_dataset_ms0.02
!du -sh /data/scratch/zhteoh/tutorial/data/cached_dataset_ms0.02
# !cat /data/scratch/zhteoh/tutorial/dataset/hparams.yaml

print("Dataset:", bliss_client.cached_data_path)
dataset_0 = bliss_client.get_dataset_file(filename="dataset_0.pt")
print(" Size:", len(dataset_0))
print(" Shape:", dataset_0[0]["images"].shape)

Train the model

Without pretrained weights

[ ]:

bliss_client.train(weight_save_path="tutorial_encoder/0.pt")

With pretrained weights

Download our relevant pretrained weights for your sky survey.

[ ]:

import os
assert os.path.exists("/data/scratch/zhteoh/tutorial/data/pretrained_models")

bliss_client.load_pretrained_weights_for_survey(survey="sdss", filename="sdss_pretrained.pt")

!ls /data/scratch/zhteoh/tutorial/data/pretrained_models

Train on cached generated disk dataset

[ ]:

bliss_client.train_on_cached_data(
    weight_save_path="tutorial_encoder/0.pt",
    train_n_batches=2,
    batch_size=64,
    val_split_file_idxs=[1],
    pretrained_weights_filename=None,
)

Run the model

Using sample SDSS dataset

Get predictions for the sample dataset

[3]:

est_cat, est_cat_table, pred_tables = bliss_client.predict_sdss(
    weight_save_path="tutorial_encoder/zscore_five_band.pt",
    # predict={"dataset": {"sdss_fields": [{"run": 94, "camcol": 1, "fields": [12]}, {"run": 3900, "camcol": 6, "fields": [296]}]}},
)


                 from  n    params  module                                  arguments
  0                -1  1     16128  yolov5.models.common.Conv               [10, 64, 5, 1]
  1                -1  3     12672  yolov5.models.common.Conv               [64, 64, 1, 1]
  2                -1  1     73984  yolov5.models.common.Conv               [64, 128, 3, 2]
  3                -1  1    147712  yolov5.models.common.Conv               [128, 128, 3, 1]
  4                -1  1    295424  yolov5.models.common.Conv               [128, 256, 3, 2]
  5                -1  6   1118208  yolov5.models.common.C3                 [256, 256, 6]
  6                -1  1   1180672  yolov5.models.common.Conv               [256, 512, 3, 2]
  7                -1  9   6433792  yolov5.models.common.C3                 [512, 512, 9]
  8                -1  1   4720640  yolov5.models.common.Conv               [512, 1024, 3, 2]
  9                -1  3   9971712  yolov5.models.common.C3                 [1024, 1024, 3]
 10                -1  1   2624512  yolov5.models.common.SPPF               [1024, 1024, 5]
 11                -1  1    525312  yolov5.models.common.Conv               [1024, 512, 1, 1]
 12                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']
 13           [-1, 6]  1         0  yolov5.models.common.Concat             [1]
 14                -1  3   2757632  yolov5.models.common.C3                 [1024, 512, 3, False]
 15                -1  1    131584  yolov5.models.common.Conv               [512, 256, 1, 1]
 16                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']
 17        [-1, 4, 5]  1         0  yolov5.models.common.Concat             [1]
 18                -1  3    756224  yolov5.models.common.C3                 [768, 256, 3, False]
 19              [17]  1     29222  yolov5.models.yolo.Detect               [33, [[4, 4]], [768]]
Model summary: 275 layers, 30795430 parameters, 30795430 gradients, 374.3 GFLOPs

GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1,2,3,4,5,6,7]


                 from  n    params  module                                  arguments
  0                -1  1     16128  yolov5.models.common.Conv               [10, 64, 5, 1]
  1                -1  3     12672  yolov5.models.common.Conv               [64, 64, 1, 1]
  2                -1  1     73984  yolov5.models.common.Conv               [64, 128, 3, 2]
  3                -1  1    147712  yolov5.models.common.Conv               [128, 128, 3, 1]
  4                -1  1    295424  yolov5.models.common.Conv               [128, 256, 3, 2]
  5                -1  6   1118208  yolov5.models.common.C3                 [256, 256, 6]
  6                -1  1   1180672  yolov5.models.common.Conv               [256, 512, 3, 2]
  7                -1  9   6433792  yolov5.models.common.C3                 [512, 512, 9]
  8                -1  1   4720640  yolov5.models.common.Conv               [512, 1024, 3, 2]
  9                -1  3   9971712  yolov5.models.common.C3                 [1024, 1024, 3]
 10                -1  1   2624512  yolov5.models.common.SPPF               [1024, 1024, 5]
 11                -1  1    525312  yolov5.models.common.Conv               [1024, 512, 1, 1]
 12                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']
 13           [-1, 6]  1         0  yolov5.models.common.Concat             [1]
 14                -1  3   2757632  yolov5.models.common.C3                 [1024, 512, 3, False]
 15                -1  1    131584  yolov5.models.common.Conv               [512, 256, 1, 1]
 16                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']
 17        [-1, 4, 5]  1         0  yolov5.models.common.Concat             [1]
 18                -1  3    756224  yolov5.models.common.C3                 [768, 256, 3, False]
 19              [17]  1     29222  yolov5.models.yolo.Detect               [33, [[4, 4]], [768]]
Model summary: 275 layers, 30795430 parameters, 30795430 gradients, 374.3 GFLOPs

GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1,2,3,4,5,6,7]

[5]:

bliss_client.plot_predictions_in_notebook()

Bokeh Plot

[4]:

print("Number of entries:", len(est_cat_table))
est_cat_table[:5].show_in_notebook(display_length=5)

Number of entries: 254

[4]:

Table length=5

idx	plocs	source_type	star_flux_u	galaxy_flux_u	star_flux_g	galaxy_flux_g	star_flux_r	galaxy_flux_r	star_flux_i	galaxy_flux_i	star_flux_z	galaxy_flux_z	galaxy_disk_frac	galaxy_beta_radians	galaxy_disk_q	galaxy_a_d	galaxy_bulge_q	galaxy_a_b
			nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy
0	tensor([266.37012, 101.83371])	tensor([1])	1.9210918e-07	1.8825295	0.65101457	4.082949	0.8307042	5.4938326	3.4831977	23.921375	3.7906942	42.90599	0.004128876	0.003471673	0.00076144224	0.0021450003	0.0032806138	1.7661792
1	tensor([549.19055, 274.24149])	tensor([1])	1.3584375e-05	0.26326695	0.121947005	0.7043628	0.464434	1.7592952	1.9545175	8.922166	2.7667255	16.547413	0.0028373753	0.0025506346	0.0007314465	0.005454664	0.0045369393	1.13643
2	tensor([236.75182, 576.83881])	tensor([1])	0.00011949862	0.08514384	0.36902255	0.8045571	0.59973806	1.0820206	1.4834237	3.1211157	0.009948616	2.3634646	0.0020792615	0.0025121698	0.0005975204	0.0025347031	0.0041335626	0.96826047
3	tensor([248.72098, 260.99545])	tensor([1])	4.0953696e-06	0.3688059	0.42916834	0.97960687	0.5803684	1.0114436	1.1978112	2.8975894	0.0026258775	3.1288908	0.00235916	0.002701597	0.00058388285	0.0032304921	0.0037438122	1.0822673
4	tensor([ 7.23273, 211.16307])	tensor([1])	0.17072378	0.88529915	0.2693183	0.6426919	0.5149893	0.95341843	0.8936408	2.1601653	0.015211616	3.36091	0.0024914178	0.0045036403	0.00030559252	0.0028280267	0.0030566908	1.2435206

Inspect probabilistic predictions

BLISS produces probability distributions on the predicted latent variables.

[5]:

print("Number of entries (RCF (94, 1, 12)):", len(pred_tables[(94, 1, 12)]))
pred_tables[(94, 1, 12)][:5].show_in_notebook(display_length=5)

Number of entries (RCF (94, 1, 12)): 24964

[5]:

Table length=5

idx	on_prob_false	on_prob_true	galaxy_prob_false	galaxy_prob_true	galsim_disk_frac_mean	galsim_disk_frac_std	galsim_beta_radians_mean	galsim_beta_radians_std	galsim_disk_q_mean	galsim_disk_q_std	galsim_a_d_mean	galsim_a_d_std	galsim_bulge_q_mean	galsim_bulge_q_std	galsim_a_b_mean	galsim_a_b_std	star_flux_u_mean	star_flux_u_std	star_flux_g_mean	star_flux_g_std	star_flux_r_mean	star_flux_r_std	star_flux_i_mean	star_flux_i_std	star_flux_z_mean	star_flux_z_std	galaxy_flux_u_mean	galaxy_flux_u_std	galaxy_flux_g_mean	galaxy_flux_g_std	galaxy_flux_r_mean	galaxy_flux_r_std	galaxy_flux_i_mean	galaxy_flux_i_std	galaxy_flux_z_mean	galaxy_flux_z_std
							rad	rad			arcsec	arcsec			arcsec	arcsec	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy
0	0.9967802	0.0032197654	0.00043278933	0.9995672	0.5724318	1.8524451	0.3748145	1.6649439	0.47005343	2.0741816	1.2313402	0.8103968	0.06683135	2.0949547	0.45463657	1.2343122	3.3472466	3.8983898	4.3143854	1.1822239	5.656625	0.49325925	5.869922	0.75289595	4.5321865	6.960256	5.0434275	1.5203606	5.571418	0.85820246	6.7504387	0.37685403	7.301384	0.46073246	6.1269445	2.2173305
1	0.9998033	0.00019671048	0.000100016594	0.9999	1.5998857	2.19047	0.5015998	1.7545787	0.6687057	1.8649726	1.6300843	0.69789016	0.43426037	1.9982051	0.64819956	1.3581581	1.5656762	4.7870736	3.5255218	1.0330999	4.566142	0.93675077	3.7978745	2.573515	3.59149	3.6905699	4.3889713	2.0829914	5.9058695	0.6866056	6.6050215	0.33476228	6.9426365	0.46526137	6.5928583	1.2505308
2	0.9999	1e-04	0.00017261505	0.9998274	1.8666542	2.0444577	0.43779612	1.7213758	0.6555178	1.7415044	1.8474009	0.64444655	0.18162489	1.8313138	0.68119955	1.2792516	1.2153401	6.0918665	3.1750164	1.3480045	4.1489496	0.9990819	3.697618	2.7583508	4.2246027	2.57833	3.9568303	2.85925	5.9339905	0.81570804	6.507223	0.43099797	6.7723436	0.70373243	7.2269716	1.1791977
3	0.99984646	0.0001535381	0.0001039505	0.99989605	1.6280034	1.8893498	-0.23069859	2.1657896	0.47975278	1.8008102	2.1519773	0.78149307	0.6024122	1.8148123	0.9013057	1.4662154	1.8403759	4.1638284	3.7711148	0.9987253	4.8116055	0.43048808	4.6379046	1.1526022	4.8451996	1.1440631	4.042429	3.6539454	5.9207506	0.9708037	6.7813787	0.5063409	7.1495094	0.9057231	7.574296	1.3889303
4	0.9997331	0.0002669305	0.000100016594	0.9999	1.4248388	1.7584034	-0.0010755062	1.9595318	0.63130164	1.4945648	2.0788147	0.80404717	0.58558655	1.8160628	0.75578976	1.3361301	2.6542282	5.4298334	3.8130317	0.9706262	5.1022377	0.41248718	4.725227	1.6942276	5.4294443	0.87989354	5.01033	2.4770377	5.858899	0.7944418	6.7706757	0.43073267	7.1094685	0.7354714	7.7826433	0.8656995

[9]:

print("Number of entries (RCF (3900, 6, 269)):", len(pred_tables[(3900, 6, 269)]))
pred_tables[(3900, 6, 269)][:5].show_in_notebook(display_length=5)

Number of entries (RCF (3900, 6, 269)): 24964

[9]:

Table length=5

idx	on_prob_false	on_prob_true	galaxy_prob_false	galaxy_prob_true	galsim_disk_frac_mean	galsim_disk_frac_std	galsim_beta_radians_mean	galsim_beta_radians_std	galsim_disk_q_mean	galsim_disk_q_std	galsim_a_d_mean	galsim_a_d_std	galsim_bulge_q_mean	galsim_bulge_q_std	galsim_a_b_mean	galsim_a_b_std	star_flux_u_mean	star_flux_u_std	star_flux_g_mean	star_flux_g_std	star_flux_r_mean	star_flux_r_std	star_flux_i_mean	star_flux_i_std	star_flux_z_mean	star_flux_z_std	galaxy_flux_u_mean	galaxy_flux_u_std	galaxy_flux_g_mean	galaxy_flux_g_std	galaxy_flux_r_mean	galaxy_flux_r_std	galaxy_flux_i_mean	galaxy_flux_i_std	galaxy_flux_z_mean	galaxy_flux_z_std
							rad	rad			arcsec	arcsec			arcsec	arcsec	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy	nmgy
0	0.9998907	0.00010931422	0.000100016594	0.9999	1.8427546	1.8966767	0.1759355	1.5340765	1.0126817	2.1146162	1.5624831	0.7504867	0.013449192	1.813689	0.7457385	1.4330941	2.078672	4.371925	3.1335154	1.3468823	4.6861925	1.025893	3.8002343	3.3332293	4.4526873	3.0467448	4.9578514	2.1461434	5.680764	1.0542581	6.6223145	0.63501734	6.69081	0.9015477	7.4244366	1.308013
1	0.9996544	0.0003456013	0.0010736585	0.99892634	1.1580436	1.8981464	0.10356331	1.51382	-0.0015649796	1.7504221	1.3518832	0.7361354	0.023548365	2.1010633	0.38921642	1.1968479	3.0933747	3.3477535	4.4125013	0.7815108	5.5652084	0.42321062	5.5774903	0.88116187	5.87446	1.5581821	4.5780935	2.0108006	5.5708265	0.6634694	6.4561443	0.37752733	6.8191853	0.5379825	7.306822	1.0409187
2	0.99985003	0.00014994045	0.00061297417	0.999387	1.3557417	1.8018293	-0.26367164	1.7946986	0.30889058	2.0305235	1.4548128	0.69167155	-0.0173347	2.2034736	0.3886218	1.1588684	3.2807875	2.6192415	4.0223684	0.91126484	5.1866264	0.53087246	5.0306916	1.3611196	5.099345	2.653987	5.0178995	1.653454	5.6688538	0.6289276	6.583728	0.34923366	6.816272	0.5685118	7.0522823	1.2957809
3	0.9996386	0.0003613974	0.00012946129	0.99987054	2.0324132	1.7666548	-0.17425346	1.8910533	1.1925094	1.5656536	1.6915038	0.6388632	0.047097445	2.0907307	0.45591784	1.4602916	1.6957612	2.8256733	2.9859772	0.9797288	3.7973025	0.75521517	3.6166434	1.5853746	2.9077227	5.6961102	5.2335787	2.4630878	6.1985826	0.875432	6.861992	0.54272926	7.1172533	0.74267733	6.7640057	2.2090604
4	0.9988968	0.0011031964	0.000100016594	0.9999	1.7430322	1.6583972	-0.015733004	1.4824201	1.1148913	1.6724063	1.8329346	0.6921283	0.23822212	2.199088	0.6723771	1.4631884	2.6342945	3.1135128	3.8255234	0.80606526	5.134838	0.335847	5.2329555	0.8405285	5.072695	1.7161341	5.033725	3.064526	5.8001394	0.9634474	6.8403854	0.54909164	7.348344	0.72839123	7.5112333	1.2816015

Save predicted catalog to FITS file

[ ]:

est_cat_table.write("est_cat.fits", format="fits", overwrite=True)

[ ]:

# Check that catalog is saved as intended
from astropy.table import Table

est_cat_table = Table.read("est_cat.fits", format="fits")
print("Number of entries:", len(est_cat_table))
est_cat_table.show_in_notebook(display_length=5)

Evaluate prediction

[ ]:

import torch

from bliss.metrics import BlissMetrics
from bliss.surveys.sdss import PhotoFullCatalog

sdss_data_path = "/data/scratch/zhteoh/tutorial/data/sdss"
sdss = SloanDigitalSkySurvey()
photo_cat = PhotoFullCatalog.from_file()

est_cat_cuda = est_cat.to(torch.device("cpu"))
photo_cat_cuda = photo_cat.to(torch.device("cpu"))

metrics = BlissMetrics()
results = metrics(est_cat_cuda, photo_cat_cuda)

print(results)

Using user-specified SDSS dataset

Download online dataset

[ ]:

from astropy.coordinates import SkyCoord
from astroquery.sdss import SDSS
from pathlib import Path

pos = SkyCoord('0h8m05.63s +14d50m23.3s', frame='icrs') # 1011/3/44
# pos = SkyCoord("1h8m05.73s +13d10m20.3s", frame="icrs") # 4829/5/27
# pos = SkyCoord("1h2m05.83s -2d11m20.3s", frame="icrs") # 2699/4/71
region = SDSS.query_region(pos, radius="5 arcsec")
run, camcol, field = region["run"][0], region["camcol"][0], region["field"][0]
print("run:", run, "camcol:", camcol, "field:", field)
bliss_client.load_survey("sdss", run, camcol, field, download_dir=Path("data/sdss"))

Get predictions for the downloaded dataset

[ ]:

est_cat_dl, est_cat_table_dl, pred_tables_dl = bliss_client.predict_sdss(
    data_path="data/sdss",
    weight_save_path="tutorial_encoder/0.pt",
    predict={"dataset": {"run": 1011, "camcol": 3, "fields": [44]}}
)

[ ]:

bliss_client.plot_predictions_in_notebook()

Inspect probabilistic predictions

[ ]:

print("Number of entries:", len(pred_tables_dl[(1011, 3, 44)]))
pred_tables_dl[(1011, 3, 44)][:5].show_in_notebook(display_length=5)

Using sample DECaLS dataset

[4]:

est_cat, est_cat_table, pred_tables = bliss_client.predict_decals(
    weight_save_path="tutorial_encoder/single_band_base.pt",
    predict={
        "dataset": {
            "sky_coords": [
                # brick '3366m010' corresponds to SDSS RCF 94-1-12
                {"ra": 336.6643042496718, "dec": -0.9316385797930247},
                # brick '1358p297' corresponds to SDSS RCF 3635-1-169
                {"ra": 135.95496736941683, "dec": 29.646883837721347},
            ]
        }
    },
)


                 from  n    params  module                                  arguments
  0                -1  1      3328  yolov5.models.common.Conv               [2, 64, 5, 1]
  1                -1  3     12672  yolov5.models.common.Conv               [64, 64, 1, 1]
  2                -1  1     73984  yolov5.models.common.Conv               [64, 128, 3, 2]
  3                -1  1    147712  yolov5.models.common.Conv               [128, 128, 3, 1]
  4                -1  1    295424  yolov5.models.common.Conv               [128, 256, 3, 2]
  5                -1  6   1118208  yolov5.models.common.C3                 [256, 256, 6]
  6                -1  1   1180672  yolov5.models.common.Conv               [256, 512, 3, 2]
  7                -1  9   6433792  yolov5.models.common.C3                 [512, 512, 9]
  8                -1  1   4720640  yolov5.models.common.Conv               [512, 1024, 3, 2]
  9                -1  3   9971712  yolov5.models.common.C3                 [1024, 1024, 3]
 10                -1  1   2624512  yolov5.models.common.SPPF               [1024, 1024, 5]
 11                -1  1    525312  yolov5.models.common.Conv               [1024, 512, 1, 1]
 12                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']
 13           [-1, 6]  1         0  yolov5.models.common.Concat             [1]
 14                -1  3   2757632  yolov5.models.common.C3                 [1024, 512, 3, False]
 15                -1  1    131584  yolov5.models.common.Conv               [512, 256, 1, 1]
 16                -1  1         0  torch.nn.modules.upsampling.Upsample    [None, 2, 'nearest']
 17        [-1, 4, 5]  1         0  yolov5.models.common.Concat             [1]
 18                -1  3    756224  yolov5.models.common.C3                 [768, 256, 3, False]
 19              [17]  1     29222  yolov5.models.yolo.Detect               [33, [[4, 4]], [768]]
Model summary: 275 layers, 30782630 parameters, 30782630 gradients, 363.8 GFLOPs

GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0,1,2,3,4,5,6,7]

[ ]:

bliss_client.plot_predictions_in_notebook()