Update htcondor and some model code by transientlunatic · Pull Request #60 · transientlunatic/heron

transientlunatic · 2026-02-03T21:38:11Z

No description provided.

Major improvements: - Replace RBF kernel with Matérn(ν=2.5) kernel in GPyTorch models - Fixes 20.56% persistent mismatch bug (now 10.16%) - Fixes 0.08s timing error (now 0.0018s, 98% improvement) - Add HeronNonSpinningApproximantMatern class Root cause: RBF kernel (infinitely differentiable) cannot handle sharp merger peak + exponential ringdown structure in GW waveforms. Matérn(ν=2.5) allows twice-differentiable functions, more flexible. Testing infrastructure: - Add test_matern_single_iteration.py for training iteration tests - Add compare_kernels.py for kernel performance comparison - Add HTCondor submit file for cluster-based training - Add development diary documenting full investigation Updated .gitignore to exclude temporary files and HTCondor logs. Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

- Add example scripts for GPR training and IMRPhenomD mean function - Add test_mean_functions.py for mean function validation - Add documentation on deep kernel learning, mean functions, and kernel analysis Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Copilot

Pull request overview

This pull request adds support for Matérn kernels to the Heron GPR waveform model, includes HTCondor cluster submission scripts, adds comprehensive documentation about mean functions and kernel choices, and provides example configuration files and scripts for training and validation.

Changes:

Added ExactGPModelMatern class and HeronNonSpinningApproximantMatern class to support Matérn kernels (ν=1.5, ν=2.5) as alternatives to RBF kernels
Added extensive test suite for mean functions (though the implementation is missing)
Added HTCondor cluster scripts and Python testing scripts for kernel comparison and training
Added comprehensive documentation about mean functions, kernel analysis, deep kernel learning, and development notes
Updated example configuration files and added new example scripts
Updated .gitignore to exclude training data, logs, and temporary files

Reviewed changes

Copilot reviewed 16 out of 17 changed files in this pull request and generated 23 comments.

Show a summary per file

File	Description
tests/test_mean_functions.py	Test suite for IMRPhenomD mean functions (implementation missing)
scripts/matern_training_cluster.sub	HTCondor submit file for distributed Matérn kernel training
scripts/test_matern_single_iteration.py	Single training iteration test script
scripts/compare_kernels.py	Comparison script for RBF vs Matérn kernels
heron/models/gpytorch.py	Added ExactGPModelMatern and HeronNonSpinningApproximantMatern classes
examples/*.yml	Configuration files for training and plotting with mean function support
examples/*.py	Example script for GPR with IMRPhenomD mean (missing implementation)
docs/*.md	Extensive documentation on mean functions, kernels, and development notes
.gitignore	Updated to exclude training data, HTCondor logs, and temporary files

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-02-03T21:45:23Z

scripts/compare_kernels.py

+        # Warp the time axis
+        points[points[:, 1] < 0, 1] = points[points[:, 1] < 0, 1] / self.warp_scale
+
+        parameters.pop("time")


The parameters.pop("time") on line 86 mutates the input dictionary. This can cause unexpected behavior if the caller reuses the dictionary. Consider creating a copy of parameters at the start of the method or document this side effect clearly.

Copilot · 2026-02-03T21:45:23Z

docs/imrphenomd_mean_function.md

+
+### 1. Mean Function: `IMRPhenomDMeanFunction`
+
+Located in: [`heron/models/mean_functions.py`](../heron/models/mean_functions.py)


Documentation references a file heron/models/mean_functions.py that doesn't exist in the codebase. The mean function implementation is missing but is extensively documented and referenced throughout the documentation files. Either the implementation needs to be added or the documentation should be removed/updated.

Suggested change

Located in: [`heron/models/mean_functions.py`](../heron/models/mean_functions.py)

Implemented by the `IMRPhenomDMeanFunction` class in the mean-function module of the `heron.models` package.

Copilot · 2026-02-03T21:45:24Z

heron/models/gpytorch.py

+        ).cuda()
+        test_times = torch.linspace(
+            time_min, time_max, time_n, dtype=torch.float32
+        ).cuda()


The _make_evaluation_manifold method hardcodes .cuda() calls on lines 356 and 359, which will fail on CPU-only systems. Should use .to(self.device) instead for consistency. The same issue appears in the parent class HeronNonSpinningApproximant (lines 189, 192).

Suggested change

).cuda()

test_times = torch.linspace(

time_min, time_max, time_n, dtype=torch.float32

).cuda()

).to(self.device)

test_times = torch.linspace(

time_min, time_max, time_n, dtype=torch.float32

).to(self.device)

Copilot · 2026-02-03T21:45:24Z

scripts/matern_training_cluster.sub

+executable              = /data/wiay/conda_envs/heron2026/bin/python
+arguments               = scripts/test_matern_single_iteration.py $(iterations)
+
+# Use shared filesystem instead of file transfer
+should_transfer_files   = NO
+
+# Set initial directory (where the job will run and logs will be written)
+initialdir              = /home/daniel/repositories/ligo/heron


Hardcoded absolute paths to specific machines will fail when run on different systems. The paths /data/wiay/conda_envs/heron2026/bin/python and /home/daniel/repositories/ligo/heron are specific to one user's setup and should be parameterized or documented as needing customization before use.

Copilot · 2026-02-03T21:45:24Z

scripts/compare_kernels.py

+        self.train_y_plus = train_y_plus.cuda() * self.output_scale
+        self.train_y_cross = train_y_cross.cuda() * self.output_scale
+
+        # Use Matérn(ν=1.5) kernels
+        self.models = {}
+        self.models["plus"] = ExactGPModelMatern(
+            self.train_x_plus, self.train_y_plus, nu=1.5
+        ).to(self.device)
+        self.models["cross"] = ExactGPModelMatern(
+            self.train_x_cross, self.train_y_cross, nu=1.5
+        ).to(self.device)
+
+        for polarisation in ("plus", "cross"):
+            self.models[polarisation].likelihood.cuda()


The HeronMatern15 class hardcodes .cuda() calls on lines 45-46 and 58 without checking CUDA availability. While line 22 defines a device variable that respects torch.cuda.is_available(), the .cuda() calls bypass this check. Use .to(self.device) instead for consistency and to avoid runtime errors on CPU-only systems.

Copilot · 2026-02-03T21:45:28Z

heron/models/gpytorch.py

+            times_i = times
+            # parameters.pop("time")
+        else:
+            times_i = (
+                torch.tensor(times.value, dtype=torch.float32) / mass_factor
+            ) - epoch


Variable times_i is not used.

Suggested change

times_i = times

# parameters.pop("time")

else:

times_i = (

torch.tensor(times.value, dtype=torch.float32) / mass_factor

) - epoch

# parameters.pop("time")

Copilot · 2026-02-03T21:45:28Z

tests/test_mean_functions.py

+
+        # First evaluation
+        x = torch.tensor([[0.8, -0.02], [0.8, 0.0]])
+        output1 = mean_fn(x, polarization='plus')


Variable output1 is not used.

Suggested change

output1 = mean_fn(x, polarization='plus')

mean_fn(x, polarization='plus')

Copilot · 2026-02-03T21:45:28Z

scripts/compare_kernels.py

+from heron.models.lalsimulation import IMRPhenomPv2
+
+# Import GP models directly to create custom approximants
+from heron.models.gpytorch import ExactGPModelKeOps, ExactGPModelMatern


Import of 'ExactGPModelKeOps' is not used.

Suggested change

from heron.models.gpytorch import ExactGPModelKeOps, ExactGPModelMatern

from heron.models.gpytorch import ExactGPModelMatern

Copilot · 2026-02-03T21:45:28Z

examples/example_gpr_with_imrphenomd_mean.py

+
+import torch
+import gpytorch
+import numpy as np


Import of 'np' is not used.

Suggested change

import numpy as np

Copilot · 2026-02-03T21:45:29Z

tests/test_mean_functions.py

+
+import pytest
+import torch
+import numpy as np


Import of 'np' is not used.

Suggested change

import numpy as np

transientlunatic and others added 3 commits February 3, 2026 21:31

Update .gitignore to exclude archive directories

486f4f6

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

Copilot AI review requested due to automatic review settings February 3, 2026 21:38

Copilot started reviewing on behalf of transientlunatic February 3, 2026 21:38 View session

Copilot AI reviewed Feb 3, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Comments

Update htcondor and some model code#60

Update htcondor and some model code#60
transientlunatic wants to merge 3 commits intomasterfrom
update-htcondor

transientlunatic commented Feb 3, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Copilot AI Feb 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant


		### 1. Mean Function: `IMRPhenomDMeanFunction`

		Located in: [`heron/models/mean_functions.py`](../heron/models/mean_functions.py)

	Located in: [`heron/models/mean_functions.py`](../heron/models/mean_functions.py)
	Implemented by the `IMRPhenomDMeanFunction` class in the mean-function module of the `heron.models` package.

	output1 = mean_fn(x, polarization='plus')
	mean_fn(x, polarization='plus')

	from heron.models.gpytorch import ExactGPModelKeOps, ExactGPModelMatern
	from heron.models.gpytorch import ExactGPModelMatern

Comments

Conversation

transientlunatic commented Feb 3, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 3, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant