Skip to content

GPU accelerated curve fitting - #76

Open
cypranowska wants to merge 48 commits into
devfrom
pytorch_fits
Open

GPU accelerated curve fitting#76
cypranowska wants to merge 48 commits into
devfrom
pytorch_fits

Conversation

@cypranowska

@cypranowska cypranowska commented Sep 10, 2026

Copy link
Copy Markdown
Member

This feature branch contains a PyTorch-based RS/OF Gaussian fitting implementation for the existing tuning models, with GPU support and fit batching in torch_fit_gaussian_blob.py. The goal is to mirror the existing SciPy fitting workflow while enabling faster fitting.

Features

  • Adds a PyTorch implementation of RS/OF Gaussian tuning models in torch_fit_gaussian_blob.py
  • Handles closedloop, openloop and treadmill protocols
  • Preserves legacy boundary condition semantics and data-informed parameter ranges
  • Adds fit-quality summaries and cross-validation
  • Integrates into pipeline_utils.py
  • Writes results to a dedicated torch parquet output path
    Testing
  • pytest --cov=v1_depth_seq tests/tests_pipelines/test_analysis_pipeline_torch.py

@cypranowska
[feature] GPU accelerated curve fitting for models in fit_gaussian_blob
f86bf48
@cypranowska
[chore] formatting
e6fe729
@cypranowska
[feature] allow upper only boundary conditions
1bdf295
@cypranowska
[feature] utilising upper only boundary on log_amplitude
49a34a1
@cypranowska
[feature] option to enforce boundary conditions on AdamW fit
1f3e562
@cypranowska
[debug] correcting per column loss evaluation
2eca3df
@cypranowska
[debug] matching the distribution of random values for log_amplitude …
a24e656 
…and offset to production code
@cypranowska
[feature] dataclasses for fit configurations, using different loss fu…
a74f64c 
…nctions
@cypranowska
[chore] formatting
81b027a
@cypranowska
[feature] demonstration notebook of pytorch fit convergence with prod…
9168b76 
…uction code
@cypranowska
[feature] pipeline function for running TRF only fit
adb30a8
@cypranowska
[refactor] clarifying class and function names
35a5973
@cypranowska
[refactor] updating callsites
f144bab
@cypranowska
[debug] updating pipeline configuration to use TRF only
a865a6e
@cypranowska
[debug] generate_n_inits_all_rois already returns initial params in d…
4e4f711 
…ata space with apply_bounds=False
@cypranowska
[minor] updating docstrings
313ada7
@cypranowska
[feature] shell script for torch fits
ee9e302
@cypranowska
[feature] adding optional installation of torch
70688dd
@cypranowska
[debug] correct activation of conda env
e766d49
@cypranowska
[refactor] using same model string names as fit_gaussian_blob
f954496
@cypranowska
[feature] using the same inits as fit_gaussian_blob
c676300
@cypranowska
[refactor] changing name for clarity
2434fc6
@cypranowska
[refactor] removing _torch suffix from dataframe column names
deb565d
@cypranowska
[feature] calculating tuning properties from fits in k_folds==1 loop
4bb4991
@cypranowska
[feature] filtering nans and resetting seed to match initial params t…
e3cc52a 
…o scipy curve_fit exactly
@cypranowska
[feature] increasing the number of iterations to 1000 if k_folds > 1
bc3a265
@cypranowska
[feature] matching min_valid_frames condition on k_fold crossval arms
e4313ff
@cypranowska
[feature] get exact same p0s passed to fit_gaussian_blob
e2af22e
@cypranowska
[performance] increasing chunk size to use more available GPU memory
11f2161
@cypranowska
[feature] seeding initial trust radius
6098b7e
@cypranowska
[minor] cleaning up demonstration notebook
67b6c91
@cypranowska
[debug] ensuring rs, of array filtering is the same between new modul…
d10be43 
…e and production code
@cypranowska
[feature] matching p0 init in fit_gaussian_blob
284ad07
@cypranowska
[feature] calculating chunk size for amount of memory on GPU node
a87257e
@cypranowska
[minor] deprecating unused functions
9edf62a
@cypranowska
[minor] deprecating adamW fits
4099e9a
@cypranowska
[feature] improving initial radius guess and step sizes close to boun…
596decc 
…dary conditions
@cypranowska
[feature] using data-driven guess for fit boundaries, one ln unit bey…
8171d6b 
…ond RS and OF data range
@cypranowska
[debug] default overhead too low for some models
05fae85
@cypranowska
[feature] implementing data-driven calculation of RS, OF boundaries
810286d
@cypranowska
[feature] helper functions for loading and fitting a single model
138b26d
@cypranowska
[refactor] consolidating things
6bba52c
@cypranowska
[feature] adjust K parameter for GPU chunk size for each model
c1d1e3f
@cypranowska
[debug] fixing dimensionality issue on trial_average arm of process_r…
a66f7a5 
…s_of_for_fit in torch module
@cypranowska
[tests] unit tests for torch fitting
645c2d9
@cypranowska
[tests] integration tests for torch pipeline
3ee1976
@cypranowska
[feature] adding trial_average=True handling for fitting on treadmill…
9465431 
… sessions
@cypranowska
[tests] updated to match pipeline defaults
61b4231
@cypranowska
cypranowska requested a review from ablot September 10, 2026 10:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@ablot ablot Awaiting requested review from ablot

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@cypranowska