high-resolution images & a 2d discrete variational factor #1068

jeff-regier · 2024-09-03T18:00:29Z

Assorted changes:

Allows the decoder to efficiently generate high-resolution images (by injecting small stamps into large images)
Generates data by launching 16 worker processes (from Python; no need to launch multiple processes using a shell script)
Improves the neural network architecture, in part by making it more shallow and adding a skip connection earlier. It also removes CatalogNet.
Adds a new variational factor for locations/position that is based on a discrete 2d grid; it works better than the TruncationBivariateGaussian in my experiments.

codecov · 2024-09-03T18:27:23Z

Codecov Report

Attention: Patch coverage is 54.83871% with 112 lines in your changes missing coverage. Please review.

Project coverage is 76.40%. Comparing base (bd96328) to head (2656a3d).
Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
bliss/encoder/encoder.py	68.13%	29 Missing ⚠️
bliss/main.py	25.71%	26 Missing ⚠️
bliss/catalog.py	19.23%	21 Missing ⚠️
bliss/simulator/prior.py	0.00%	18 Missing ⚠️
bliss/encoder/variational_dist.py	34.61%	17 Missing ⚠️
bliss/simulator/decoder.py	95.65%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1068      +/-   ##
==========================================
- Coverage   81.25%   76.40%   -4.86%     
==========================================
  Files          23       22       -1     
  Lines        3051     3085      +34     
==========================================
- Hits         2479     2357     -122     
- Misses        572      728     +156

Flag	Coverage Δ
unittests	`76.40% <54.83%> (-4.86%)`	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

jeff-regier · 2024-09-04T13:39:04Z

case_studies/weak_lensing/lensing_prior.py

-            "galaxy_params": galaxy_params,
-            "star_fluxes": star_fluxes,
-        }
+        catalog_params["shear"] = shear


@shreyasc30 @timwhite0 The function _sample_galaxy_prior() no longer returns fluxes. I updated this code to account for the change but couldn't test it, so there may be minor syntax errors here.

jeff-regier · 2024-09-04T13:40:28Z

case_studies/weak_lensing/lensing_convnet.py

-                ConvBlock(nch_hidden, 64, kernel_size=5, padding=2),
-                nn.Sequential(*[ConvBlock(64, 64, kernel_size=5, padding=2) for _ in range(1)]),
+                ConvBlock(nch_hidden, 64, kernel_size=5),
+                nn.Sequential(*[ConvBlock(64, 64, kernel_size=5) for _ in range(1)]),


@shreyasc30 @timwhite0 The appropriate padding gets determined automatically now by ConvBlock, so padding is no longer an available argument.

jeff-regier · 2024-09-04T13:47:24Z

case_studies/galaxy_clustering/encoder/encoder.py

            tile_slen=self.tile_slen,
            downsample_at_front=self.downsample_at_front,
        )
-        n_params_per_source = self.var_dist.n_params_per_source
-        self.catalog_net = CatalogNet(num_features, n_params_per_source)


@kapnadak @gapatron CatalogNet is no longer around. To account for this, I adjusted your code by removing its references to CatalogNet and by making the output of your GalaxyClusterFeaturesNet have the right output dimension to be used directly to parameterize the variational distribution. I haven't tested these changes though: some additional modifications may be needed.

jeff-regier · 2024-09-04T14:32:18Z

bliss/main.py

+
+
+def generate_one_file(gen_cfg: DictConfig, file_idx: int):
+    prior = instantiate(gen_cfg.prior)


@kapnadak @gapatron The BLISS decoder can efficiently generate high-resolution images now. If you want to plug in your galaxy cluster prior, it just requires adjusting the config files so that cfg.generate.prior points to your prior rather than the default one.

jeff-regier · 2024-09-04T18:54:09Z

bliss/encoder/convnets.py

+            nn.Upsample(scale_factor=2, mode="nearest"),  # 8
+            C3(512, 256, n=3, shortcut=False),
+            nn.Upsample(scale_factor=2, mode="nearest"),  # 10
+            C3(384, num_features, n=3, shortcut=False),


@YicunDuanUMich @georgeyfly This new network architecture works quite a bit better for me (on SDSS, I haven't tried in on DC2). Will you try it out too?

That's nice to hear. I am glad to have a try on DC2 to see if it works better.

Thanks! I'll try it.

jeff-regier added 30 commits August 5, 2024 10:59

null sample renders bug

5b4ca23

fix annoying data_source warning

82ded83

aug5_discretizedbox_quarterpixels

79f7d20

not much, some m2 notebook stuff

aa4cadd

sdss case study

16d2083

revert convnets to master

bf40a00

merge

c8635eb

fixed some merge discrepencies

8584488

bring back former to_tile_catalog

8b29fd0

self.double_sample_prob and nll hack

0d62d71

sdss demo needs retrain with one band fluxes

49cd763

tiny changes to m2 case study

095fcbd

revert problematic changes

546f48a

fix bin_cutoffs

cd91253

logsumexp for double detections

2a6ee90

back to 32-true

77b26d8

lots of 1x1 kernels

df502bc

updated sample for logsumexp double detect

dab7ff7

wider context net

4032fa2

removed incorrect masking

d1848d4

gate new_est_cat2 during sampling

6cd81f1

simplify color context

2540ad4

count_net

168b77b

countnet refactor

5027822

richer color history

d341192

no groupnorm in heads

d5dcfe8

minimalist_history

e3245c6

deeper

c13f441

1x1 for real

6c594a6

centered locs for colors

16dac53

jeff-regier added 22 commits August 23, 2024 14:38

earlier first skip connection

d44461d

null normalizer

1db8498

no groupnorm for localnet or detectionnet

e6b3e70

restore normalizers

80fb47d

flux in color context

2be7fa1

some merging; give decoder bounds

7c94649

recovered from merge?

317bba8

use multiprocessing for generate

3db820e

add mask patterns

62f90fb

pylint

e22be1d

flake8

fe0cc24

remove simulator

b895d33

forgot to remove line

b606b0a

wrong path

201c3d1

simplify prior

a8c1ad5

restored double downsample

e82dfd9

remove fixtures; fix test paths

d06402b

revert log10 division change

d1f5496

tiles_to_crop

a09de3e

smaller images

efed656

new base_config encoder

4b9dfcf

notebooks running again

2656a3d

jeff-regier merged commit 84ca732 into master Sep 3, 2024
2 of 3 checks passed

jeff-regier deleted the jr/discrete_locs branch September 3, 2024 18:31

jeff-regier commented Sep 4, 2024

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

high-resolution images & a 2d discrete variational factor #1068

high-resolution images & a 2d discrete variational factor #1068

jeff-regier commented Sep 3, 2024 •

edited

Loading

codecov bot commented Sep 3, 2024 •

edited

Loading

jeff-regier Sep 4, 2024

jeff-regier Sep 4, 2024

jeff-regier Sep 4, 2024

jeff-regier Sep 4, 2024 •

edited

Loading

jeff-regier Sep 4, 2024

georgeyfly Sep 4, 2024

YicunDuanUMich Sep 5, 2024



		def generate_one_file(gen_cfg: DictConfig, file_idx: int):
		prior = instantiate(gen_cfg.prior)

high-resolution images & a 2d discrete variational factor #1068

high-resolution images & a 2d discrete variational factor #1068

Conversation

jeff-regier commented Sep 3, 2024 • edited Loading

codecov bot commented Sep 3, 2024 • edited Loading

Codecov Report

jeff-regier Sep 4, 2024

Choose a reason for hiding this comment

jeff-regier Sep 4, 2024

Choose a reason for hiding this comment

jeff-regier Sep 4, 2024

Choose a reason for hiding this comment

jeff-regier Sep 4, 2024 • edited Loading

Choose a reason for hiding this comment

jeff-regier Sep 4, 2024

Choose a reason for hiding this comment

georgeyfly Sep 4, 2024

Choose a reason for hiding this comment

YicunDuanUMich Sep 5, 2024

Choose a reason for hiding this comment

jeff-regier commented Sep 3, 2024 •

edited

Loading

codecov bot commented Sep 3, 2024 •

edited

Loading

jeff-regier Sep 4, 2024 •

edited

Loading