{{ message }}
Repository navigation
`CompoundModel.n_left_params counts parameter values, not parameters, so array-valued parameters and model sets break evaluate() and every parallel_fit_dask fit #20554
Description
Description
In my glorious attempt to make the lives of my users a nightmare, I have come up with fixed compound models for them to use as starting points for their spectral fitting.
This means having parameters with array and I hit this bug.
parallel_fit_dask documents that a model's parameters may be arrays shaped like the iterating axes, so each spectrum gets its own start. With a compound model such as Const1D + Gaussian1D, every fit raises TypeError: Const1D.evaluate() takes 2 positional arguments but 5 were given and every parameter comes back NaN. The same error hits any compound model with array-valued parameters or a model set (n_models > 1) whenever evaluate(*inputs, *parameters) is called, which is what the fitters do.
Expected behavior
n_left_params should be the number of parameters of the left model, len(self.left.param_names), regardless of the shape of their values.
Suggested patch:
--- a/astropy/modeling/core.py
+++ b/astropy/modeling/core.py
@@ CompoundModel.__init__ @@
- self.n_left_params = len(self.left.parameters)
+ self.n_left_params = len(self.left.param_names)How to Reproduce
LLM code example warning
import tempfile
from pathlib import Path
import numpy as np
from astropy.modeling import models
from astropy.modeling.fitting import TRFLSQFitter, parallel_fit_dask
scalar = models.Const1D(1.0) + models.Gaussian1D(1.0, 0.0, 1.0)
array = models.Const1D(np.ones((3, 4))) + models.Gaussian1D(np.ones((3, 4)), np.zeros((3, 4)), np.ones((3, 4)))
print("scalar parameters: n_left_params =", scalar.n_left_params)
print("(3, 4) parameters: n_left_params =", array.n_left_params, "for", len(array.left.param_names), "parameter")
# __call__ works; evaluate(*inputs, *parameters), which the fitters use, does not.
x = np.linspace(-1, 1, 5)
print("array(x).shape =", array(x[:, None, None]).shape)
try:
array.evaluate(x[:, None, None], *(getattr(array, name).value for name in array.param_names))
except TypeError as error:
print("array.evaluate(...) ->", error)
# Model sets are affected too.
model_set = models.Const1D([1.0, 2.0], n_models=2) + models.Gaussian1D([1.0, 1.0], [0.0, 0.0], [1.0, 1.0], n_models=2)
try:
model_set.evaluate(x[:, None], *(getattr(model_set, name).value for name in model_set.param_names))
except TypeError as error:
print("model_set.evaluate(...) ->", error)
# parallel_fit_dask with one start per pixel: every fit raises and comes back NaN.
rng = np.random.default_rng(0)
xs = np.linspace(-1, 1, 50)
data = 2 + 5 * np.exp(-0.5 * ((xs - rng.normal(0, 0.05, (3, 4))[..., None]) / 0.2) ** 2) + rng.normal(0, 0.1, (3, 4, 50))
start = models.Const1D(np.ones((3, 4))) + models.Gaussian1D(np.full((3, 4), 4.0), np.zeros((3, 4)), np.full((3, 4), 0.3))
with tempfile.TemporaryDirectory() as path:
fitted = parallel_fit_dask(model=start, fitter=TRFLSQFitter(), data=data, world=(xs,), fitting_axes=2,
scheduler="single-threaded", diagnostics="error", diagnostics_path=path)
log = sorted(Path(path).rglob("error.log"))[0].read_text().strip().splitlines()[-1]
print(f"parallel_fit_dask: {np.isnan(fitted.mean_1.value).sum()} of 12 fits NaN; error.log ends with: {log}")
# The proposed fix, applied by hand to this instance:
start.n_left_params = len(start.left.param_names)
fitted = parallel_fit_dask(model=start, fitter=TRFLSQFitter(), data=data, world=(xs,), fitting_axes=2, scheduler="single-threaded")
print(f"with n_left_params = len(left.param_names): {np.isnan(fitted.mean_1.value).sum()} of 12 fits NaN")Output:
scalar parameters: n_left_params = 1
(3, 4) parameters: n_left_params = 12 for 1 parameter
array(x).shape = (5, 3, 4)
array.evaluate(...) -> Const1D.evaluate() takes 2 positional arguments but 5 were given
model_set.evaluate(...) -> Const1D.evaluate() takes 2 positional arguments but 3 were given
parallel_fit_dask: 12 of 12 fits NaN; error.log ends with: TypeError: Const1D.evaluate() takes 2 positional arguments but 5 were given
with n_left_params = len(left.param_names): 0 of 12 fits NaN
Versions
platform.platform() = 'Linux-7.2.9-cachyos1.fc44.x86_64-x86_64-with-glibc2.43'
platform.python_version() = '3.14.5'
astropy 8.1.0.dev684+g2b5b20b6d (also reproduced on 8.0.1)
numpy 2.4.6
scipy 1.17.0
pyerfa 2.0.1.5
dask 2026.1.2
Activity
[-]CompoundModel.n_left_params counts parameter values, not parameters, so array-valued parameters and model sets break evaluate() and every parallel_fit_dask fit[/-][+]``CompoundModel.n_left_params` counts parameter values, not parameters, so array-valued parameters and model sets break `evaluate()` and every `parallel_fit_dask` fit[/+]Count parameters, not parameter values, in CompoundModel.n_left_params