PyAntiGen 2.1.2__tar.gz → 2.1.4__tar.gz

{pyantigen-2.1.2 → pyantigen-2.1.4}/LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2026 UW Elbert Lab
+Copyright (c) 2026 University of Washington - Elbert Lab
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

{pyantigen-2.1.2 → pyantigen-2.1.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PyAntiGen
-Version: 2.1.2
+Version: 2.1.4
 Summary: Declarative generation, simulation and identifiability analysis of compartmental Antimony/SBML models
 Author-email: Don Elbert <elbert5770@gmail.com>
 License-Expression: MIT

{pyantigen-2.1.2 → pyantigen-2.1.4}/PyAntiGen.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PyAntiGen
-Version: 2.1.2
+Version: 2.1.4
 Summary: Declarative generation, simulation and identifiability analysis of compartmental Antimony/SBML models
 Author-email: Don Elbert <elbert5770@gmail.com>
 License-Expression: MIT

{pyantigen-2.1.2 → pyantigen-2.1.4}/PyAntiGen.egg-info/SOURCES.txt

@@ -14,6 +14,7 @@ PyAntiGen.egg-info/requires.txt
 PyAntiGen.egg-info/scm_file_list.json
 PyAntiGen.egg-info/scm_version.json
 PyAntiGen.egg-info/top_level.txt
+docs/ANALYTIC_GRADIENT.md
 docs/MULTISTART.md
 docs/PROTOCOL_LAYER.md
 docs/V2_DESIGN.md
@@ -24,6 +25,7 @@ pyantigen/cli.py
 pyantigen/.agents/skills/module_generation/SKILL.md
 pyantigen/.agents/skills/ode_conversion_antimony/SKILL.md
 pyantigen/.agents/skills/ode_extraction_pipeline/SKILL.md
+pyantigen/engine/Analytic_gradient.py
 pyantigen/engine/Anchor_cache.py
 pyantigen/engine/Deadline.py
 pyantigen/engine/Differential_evolution.py
@@ -45,6 +47,7 @@ pyantigen/engine/Profile_checkpoint.py
 pyantigen/engine/Profile_plan.py
 pyantigen/engine/README.md
 pyantigen/engine/Results.py
+pyantigen/engine/Scaled_amici.py
 pyantigen/engine/Segmented.py
 pyantigen/engine/Sensitivity_analysis.py
 pyantigen/engine/Simulate.py
@@ -64,6 +67,7 @@ pyantigen/generate/module_base.py
 pyantigen/generate/pyantigen.py
 pyantigen/generate/rate_laws.py
 pyantigen/generate/reaction_creation.py
+pyantigen/generate/scaled_time.py
 pyantigen/study/__init__.py
 pyantigen/study/assay.py
 pyantigen/study/describe.py

{pyantigen-2.1.2 → pyantigen-2.1.4}/PyAntiGen.egg-info/scm_file_list.json

@@ -7,6 +7,7 @@
     "LICENSE",
     "MANIFEST.in",
     "README.md",
+    "docs/ANALYTIC_GRADIENT.md",
     "docs/MULTISTART.md",
     "docs/PROTOCOL_LAYER.md",
     "docs/V2_DESIGN.md",
@@ -16,6 +17,7 @@
     "pyantigen/.gitignore",
     "pyantigen/__init__.py",
     "pyantigen/cli.py",
+    "pyantigen/engine/Analytic_gradient.py",
     "pyantigen/engine/Anchor_cache.py",
     "pyantigen/engine/Deadline.py",
     "pyantigen/engine/Differential_evolution.py",
@@ -37,6 +39,7 @@
     "pyantigen/engine/Profile_plan.py",
     "pyantigen/engine/README.md",
     "pyantigen/engine/Results.py",
+    "pyantigen/engine/Scaled_amici.py",
     "pyantigen/engine/Segmented.py",
     "pyantigen/engine/Sensitivity_analysis.py",
     "pyantigen/engine/Simulate.py",
@@ -56,6 +59,7 @@
     "pyantigen/generate/pyantigen.py",
     "pyantigen/generate/rate_laws.py",
     "pyantigen/generate/reaction_creation.py",
+    "pyantigen/generate/scaled_time.py",
     "pyantigen/study/__init__.py",
     "pyantigen/study/assay.py",
     "pyantigen/study/describe.py",
@@ -99,6 +103,11 @@
     "tests/README.md",
     "tests/__init__.py",
     "tests/conftest.py",
+    "tests/engine/test_amici_flux_outputs.py",
+    "tests/engine/test_amici_jumps.py",
+    "tests/engine/test_amici_sensitivity.py",
+    "tests/engine/test_analytic_gradient.py",
+    "tests/engine/test_analytic_gradient_amici.py",
     "tests/engine/test_anchor_cache.py",
     "tests/engine/test_better_point.py",
     "tests/engine/test_de_fit.py",
@@ -123,7 +132,9 @@
     "tests/engine/test_profile_records.py",
     "tests/engine/test_profile_stream.py",
     "tests/engine/test_results_csv_policy.py",
+    "tests/engine/test_schedule_bounds.py",
     "tests/engine/test_search_mode.py",
+    "tests/engine/test_segment_jumps.py",
     "tests/engine/test_segment_sensitivity.py",
     "tests/engine/test_segmented.py",
     "tests/engine/test_sigma_floor.py",
@@ -139,6 +150,7 @@
     "tests/generate/test_gen_models.py",
     "tests/generate/test_gen_pyantigen_api.py",
     "tests/generate/test_gen_rxndict_to_antimony.py",
+    "tests/generate/test_gen_scaled_time.py",
     "tests/silk_fixtures/antimony_models/Antimony_Bloomingdale_2021_1a_all_reactions.txt",
     "tests/silk_fixtures/antimony_models/Antimony_Bloomingdale_2021_1a_rules.txt",
     "tests/silk_fixtures/antimony_models/Antimony_Elbert_2022_1a_all_reactions.txt",

pyantigen-2.1.4/PyAntiGen.egg-info/scm_version.json

@@ -0,0 +1,8 @@
+{
+  "tag": "2.1.4",
+  "distance": 0,
+  "node": "g2e65abb870fe7cb1136e01bcd8c30b6d0f712dd6",
+  "dirty": false,
+  "branch": "HEAD",
+  "node_date": "2026-10-14"
+}

pyantigen-2.1.4/docs/ANALYTIC_GRADIENT.md

@@ -0,0 +1,137 @@
+# Fitting with a schedule: the bounds check and the AMICI gradient
+
+Two options of `optimizer_kwargs` for a spec whose simulations take their clock-driven inputs from
+an `InputSchedule` (see `pyantigen/generate/input_schedule.py`). Both leave a spec without a
+schedule exactly as it was.
+
+## `schedule_bounds`: keep the box inside the region the schedule can be integrated in
+
+A schedule can only be integrated while its switches keep their order (`InputSchedule.validate`):
+in the SILK hour `0 < t_CSFdraw < V_LP/Q_CSF < 1`. A parameter set where they cross is a failed
+evaluation, and a gradient method reads the failure as a cliff. Before the first evaluation the
+optimizer checks the bounds of every fitted parameter that moves a switch time against each
+simulation's schedule (`InputSchedule.check_box`: the corners of the box, the starting point, and
+64 pseudo-random points inside it).
+
+```python
+optimizer_kwargs = {"schedule_bounds": "warn"}      # "warn" (default) | "error" | "clip" | "ignore"
+```
+
+* `"warn"` prints which corner is infeasible and goes on. A box wider than the region is not
+  necessarily wrong, but those evaluations will fail.
+* `"error"` raises `ScheduleError` with the same message.
+* `"clip"` cuts the bounds of the switch-moving parameters toward x0, by the largest common factor
+  for which the box is feasible (`InputSchedule.shrink_box`), and prints what it did.
+* A starting point that is itself infeasible stops the fit in every mode except `"ignore"`.
+
+Only simulations whose solver settings carry a `schedule` are checked, and the fitted names are taken
+to be the schedule's parameters: a parameter a hook derives from them is not followed.
+
+## `analytic_gradient`: the gradient from AMICI's forward sensitivities
+
+```python
+optimizer_kwargs = {
+    "method": "L-BFGS-B",
+    "analytic_gradient": True,                       # or a dict of the settings below
+}
+```
+
+Used by every gradient method (`L-BFGS-B`, `BFGS`, `CG`, `TNC`, `SLSQP`): in a single fit and in the
+local fits of the multi-start stage, in place of the finite differences. With a gradient-free method it
+is ignored (with a note). Needs AMICI >= 1.1 (on Windows, under WSL) and the simulations' inputs as an
+`InputSchedule`, or as `at (time >= ...)` event lines `InputSchedule.from_antimony_events` can read.
+A problem it cannot handle stops the run before any fit (`Multistart.GradientSetupError`).
+
+| key | default | meaning |
+| --- | --- | --- |
+| `rtol`, `atol` | 1e-10, 1e-12 | AMICI's tolerances (state and sensitivities) |
+| `max_steps` | 2 000 000 | AMICI step cap; the 10 000 h pre-equilibration of SILK needs it |
+| `rel_step` | 1e-5 | step of the loss differentiation (below) |
+| `cache_dir` | `<plot_path>/amici_models` | compiled models; one directory per distinct model |
+| `seed` | none | `{state: amount}` given to AMICI at the start of every run, on top of the model's own initial state (see "A model that starts at zero") |
+
+### What it computes
+
+The objective is the engine's concentrated NLL. Its derivative is not written out again; it is split by
+the chain rule into the part AMICI knows and the part the loss knows:
+
+1. AMICI integrates the event-free model of `generate/scaled_time.py` and differentiates every
+   observed column (concentrations as `species / compartment`, assigned variables, parameters) with
+   respect to the model parameters the fitted ones reach. That table is mapped to the fitted parameters by
+   the Jacobian of the engine's own parameter updates (`Update_parameters`, `parameter_hooks`,
+   `Update_opt_parameters`), obtained by replaying them on the RoadRunner model.
+2. `LossGradient` differentiates the engine's own loss (`collect_loss_blocks`, `concentrated_nll`:
+   groups, pooled sigma blocks, sigma floors, observable expressions such as
+   `np.log10(np.maximum(predicted_B, 1e-12))`) along that table, by a Richardson central difference of the
+   loss alone. It re-does arithmetic on tables in memory: no integration, no noise. A parameter the loss
+   names (`SF*predicted_B`) gets its explicit partial derivative too.
+
+The value in the fit is still the engine's (`Multistart.AnalyticObjective`: value from the objective,
+gradient from the provider), so the best-so-far tracking, caches and failure handling are unchanged.
+If AMICI fails at one point (a solver error, a non-finite sensitivity) that gradient is taken by forward
+differences of the objective, counted, and reported with the fit; a zero gradient there would read as
+convergence.
+
+### It differentiates the engine's objective, not the model's formulas
+
+Where the two differ, the gradient follows the engine:
+
+* **RoadRunner's `reset()` re-derives a species' initial assignment from the current parameters but not a
+  parameter's.** A parameter defined by an initial assignment from a fitted one (SILK's `Q_glymph = Q_CSF`,
+  `Q_SN = Q_CSF*f_SN`) keeps the value it had when the model was built while the optimizer moves `Q_CSF`.
+  The AMICI model gets those parameters as plain numbers at the engine's values
+  (`freeze_stale_assignments`); a schedule input among them keeps its stale starting value when it
+  differs from the formula. The stale value is a property of the in-place simulation, and the gradient
+  matches it.
+* **Every parameter with an initial assignment is frozen**, not only those that read a fitted one. A model
+  whose constants are written as expressions (`k = 9.56e-7*70`, `V = V_base`) has them as initial
+  assignments, and a project's `Update_parameters` or hooks may overwrite them on the RoadRunner model.
+  Each becomes a plain parameter at the value RoadRunner holds, so AMICI integrates what the engine
+  integrates, and a hook that moves it is seen by the replay. (Before, only those reading a fitted parameter
+  were frozen; the others stayed NaN placeholders that reached AMICI's parameter vector, and a value a hook
+  had overwritten was not the one AMICI used.) Variables the schedule assigns, and what their initial values
+  are built from, are left alone.
+* A species a hook sets directly is handed to AMICI as an initial value (without sensitivities; a hook
+  whose value depends on the fitted parameters is refused).
+
+### A model that starts at zero
+
+AMICI differentiates the right-hand side symbolically and evaluates it at the first instant. A rate law
+with a fractional power of an amount that is exactly zero there, `(Total/V)^0.8`, has an infinite derivative
+at 0, and `0 * inf` is NaN: the run fails at t = 0 with status -42 ("NaN value for sxdot"). RoadRunner does
+not differentiate, so the same model integrates without complaint. Give the states in question a negligible
+amount, in AMICI only:
+
+```python
+optimizer_kwargs = {"analytic_gradient": {"seed": {"AB40_BrainISF": 1e-12, "AB42_BrainISF": 1e-12}}}
+```
+
+Read the AMICI log for the names: it lists the first states and rates that were NaN. The effect on the
+trajectory is of the order of the seed, forgotten within the first steps of a stable system; check it for
+yours by comparing the AMICI values with the engine's, as `Checks/wsl/` does. Unguarded ratios of two
+zero amounts (`A / B` as an output) also print a NaN warning at t = 0 but do not stop the run unless the
+dynamics read them.
+
+### Limits
+
+* Time points: the loss is differentiated at the data times and the ends of the tracked window; the
+  engine interpolates a dense grid linearly between them. The two differ by the interpolation error
+  (a few 1e-4 of the gradient's scale on SILK).
+* The fitted parameters must reach the model through plain parameters (constants without rule or
+  initial assignment) or hooks onto them. A fitted species initial value, or a parameter with an initial
+  assignment, is refused with a message naming it.
+* Columns that are not species, parameters or compartments (reaction rates) are not available to the
+  loss; the gradient is then zero for anything that depends only on them.
+* The profile-likelihood stage keeps its own optimizers.
+  The gradient can differentiate what a profile point minimizes: `provider.freeze_sigmas(block_sigmas)` pins the floored blocks'
+  sigmas as `evaluate_nll_fixed(frozen_sigmas=...)` does (without it the gradient is off by 10 to 30% away from the fit);
+  `freeze_sigmas(None)` releases them. Nothing calls it yet.
+* Cost: the gradient costs one AMICI forward-sensitivity run, which grows with the number of
+  parameters it differentiates. On SILK, 8 parameters: about 1.2 s a gradient against 9 objective
+  evaluations (about 1.4 s) for forward differences; with all 48 parameters AMICI needs about 40 s a
+  gradient against about 8 s for 49 evaluations. The case for it is a gradient without the noise of the
+  integrator tolerance and without a step to choose, not speed at every size. Fit only the parameters
+  that need fitting: the model is compiled for exactly those.
+
+Checks on the real SILK model are in the examples repository,
+`Projects/Elbert_2022_PyAntiGen_v4/Checks/wsl/fit_gradient.py`.

{pyantigen-2.1.2 → pyantigen-2.1.4}/pyantigen/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '2.1.2'
-__version_tuple__ = version_tuple = (2, 1, 2)
+__version__ = version = '2.1.4'
+__version_tuple__ = version_tuple = (2, 1, 4)
 
-__commit_id__ = commit_id = 'g6b3300aa1'
+__commit_id__ = commit_id = 'g2e65abb87'