Better support for symbolic (e.g. jax tracer) values in non-index coordinates
Los mantenedores suelen responder en 1 día
Nadie ha tomado este issue todavía.
Evaluación
- Dificultad
- 5/5
- Tiempo estimado
- Más de una semana
- Aptitud para principiantes
- 42/100
Línea de trabajo
Comienza con xarray/core/variable.py, alrededor de la comparación de compatibilidad mencionada en el issue, y luego sigue el manejo de coordenadas aritméticas, xr.align y las configuraciones existentes de xarray.set_options. Revisa la discusión para determinar cuál de los tres enfoques de compatibilidad se acepta; la tarea estará terminada cuando el comportamiento elegido esté implementado de forma coherente y cubierto para valores de coordenadas simbólicos.
Escrito por el modelo de indexación a partir del texto del issue.
Descripción
Is your feature request related to a problem?
Hello! xarray_jax maintainer here.
I am trying to decide whether it is feasible to rely on using symbolic arrays (such as jax tracers used during JIT compilation) as coordinate data in xarray.
For index coordinates this doesn't work at all, because xarray requires dynamic access to the data in the index to do alignment. So our approach in xarray_jax is to treat index coordinates as static data. I think this is acceptable: even if we could do index alignment with tracers, the result would be variable-shape which jax.jit can't handle without re-tracing anyway.
For non-index coordinates though, it is possible to use jax tracers! However these coordinates have a habit of disappearing due to failed alignment checks, for example in arithmetic operations:
import xarray as xr
import jax
import jax.numpy as jnp
def foo(coord):
a = xr.DataArray(data=jnp.ones(3), dims=['x'], coords={'foo': (['x'], coord)})
b = xr.DataArray(data=jnp.ones(3), dims=['x'], coords={'foo': (['x'], coord)})
print((a + b).coords)
foo(jnp.arange(3))
# prints:
# Coordinates:
# foo (x) int32 12B ...
jax.jit(foo)(jnp.arange(3))
# prints:
# Coordinates:
# *empty*
The coordinate is (correctly) preserved when not JIT-ing, because xarray is able to do an equality comparison to check that the coordinate is compatible.
When JITing however, the coordinate array will be a jax Tracer, and the equality comparison will fail with a jax.errors.TracerBoolConversionError (or similar). xarray catches this under the hood here since TracerBoolConversionError subclasses TypeError. xarray then treats the failed comparison as not-equal resulting in the coordinate being dropped.
This isn't entirely consistent with behaviour elsewhere in xarray.
For example when doing an explicit merge it is possible to set compat='override' which ensures that coordinates aren't tested for equality and so jax tracers survive. And I believe compat='override' is to become the default soon.
xr.align also lets jax Tracers survive:
def foo(coord):
a = xr.DataArray(data=jnp.ones(3), dims=['x'], coords={'foo': (['x'], coord)})
b = xr.DataArray(data=jnp.ones(3), dims=['x'], coords={'foo': (['x'], coord)})
a, b = xr.align(a, b)
print(a.coords)
print(b.coords)
jax.jit(foo)(jnp.arange(3))
# prints:
# Coordinates:
# foo (x) int32 12B ...
# Coordinates:
# foo (x) int32 12B ...
Overall, while I can see why some might want them, these compatibility checks on non-index coordinates seem a bit of an awkward niche behaviour which is configurable in some places but not others, and (at least after the compat='override' default takes over) will have different defaults in different places too IIUC, which is not ideal. And of course they throw a bit of a spanner in the works of using jax here.
Describe the solution you'd like
Three suggestions:
-
Add a
xarray.set_options(arithmetic_compat='override')setting which controls this, and make 'override' the default for it at the same timecompat='override'becomes the default more widely. Note there is already anarithmetic_joinsetting which controls how index coordinates are joined/aligned, but this doesn't give any control over these compatibility checks on non-index coordinates.
One might also consider exposing thecompatoption onxr.aligntoo and any other places where compatibility checks are happening or for consistency ought to happen, to give control over this. -
Allow
jax.errors.TracerBoolConversionErrorto bubble up from generic equality-testing code (or perhaps catch and reraise it as something less jax-specific likexr.SymbolicComparisonError) rather than catching it and treating it as unequal. Then in coordinate compatibility-checking code, catch this error and fall back oncompat='override'behavior in this specific case. This has the advantage that non-index-coord compatibility checks don't have to be disabled across the board, only for jax Tracers. It would require some jax-specific special case in xarray though. -
Just to get rid of alignment checks on non-index coordinates entirely, if noone feels strongly about them. They are already somewhat inconsistent both in configurability and in default behaviour. In many cases it's a safe assumption that any non-index coordinates are a fixed function of the index coordinates and so it's sufficient to check for alignment of the index coordinates. And perhaps other cases it would be OK to make any checking the user's responsibility. I realise this would be backwards-incompatible though.
Describe alternatives you've considered
See above.
Additional context
No response
- Lenguaje dominante
- Python
- Estrellas
- 4.2k
- Forks
- 1.5k
- Merge medio
- 2 d 21 h
- PR fusionados (30 d)
- 74
Preparar el entorno
Inicia el contenedor de desarrollo del proyecto en tu navegador, con tu propia cuenta de GitHub.
- Sin Dockerfile ni archivo de Docker Compose
- Tiene una plantilla de pull request
- Leer la guía de contribución
Primeros pasos
- Lee el issue completo y luego la guía de contribución del proyecto.
- Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
- Haz un fork del repositorio y trabaja en una rama.
- Abre un pull request que haga referencia al número del issue.
Más de pydata/xarray
-
xarray removes axis grid from pcolormeshPosiblemente ocupada @Sealdot la tomó hoy. Abiertobug needs triage
Dificultad 2/5 1-3 horas Aptitud para principiantes 70/100
Los mantenedores suelen responder en 1 día
-
`DataTree.map_over_datasets` behavior can be obscure sometimesPosiblemente ocupada @mohansree14 la tomó hace 24 días. Abiertoneeds triage
Dificultad 2/5 1-3 horas Aptitud para principiantes 78/100
pydata/xarray#11813 · 1 comentario ·
Los mantenedores suelen responder en 1 día
-
Comparing attrs that contain a dict of NumPy arrays raises ValueError in identical / assert_identicalPosiblemente ocupada @SohamXYZDev la tomó hace 29 días. Abiertobug needs triage
Dificultad 2/5 1-3 horas Aptitud para principiantes 73/100
pydata/xarray#11777 · 1 comentario ·
Los mantenedores suelen responder en 1 día
-
RangeIndex` with datetime64 bounds cannot be printed (`Unknown format code 'g'`)Posiblemente ocupada @d-chambers la tomó hace 60 días. Abiertobug
Dificultad 2/5 1-3 horas Aptitud para principiantes 68/100
pydata/xarray#11598 · 3 comentarios ·
Los mantenedores suelen responder en 1 día
-
DataTree constructor error message says dict even though dict is not acceptedPosiblemente ocupada @aryansk la tomó hace 100 días. Abiertobug contrib-good-first-issue topic-DataTree topic-error reporting
Dificultad 2/5 1-3 horas Aptitud para principiantes 72/100
pydata/xarray#11514 · 2 comentarios ·
Los mantenedores suelen responder en 1 día
Todos los issues de pydata/xarray
Issues similares
-
Dificultad 2/5 1-3 horas Aptitud para principiantes 78/100
abetlen/llama-cpp-python#2424 ·
Los mantenedores suelen responder en 1 día
-
Bug IronPython Python 3
Dificultad 2/5 1-3 horas Aptitud para principiantes 72/100
pyrevitlabs/pyRevit#4184 ·
Los mantenedores suelen responder en 1 día
-
CLAW (Agents & Orchestration)Abierto
Dificultad 1/5 Menos de una hora Aptitud para principiantes 74/100
jqueryscript/awesome-claude-code#1009 ·
-
tests
Dificultad 2/5 1-3 horas Aptitud para principiantes 66/100
Los mantenedores suelen responder en 3 días
-
Dificultad 2/5 1-3 horas Aptitud para principiantes 72/100
Los mantenedores suelen responder en 1 día