Release PyFixest 0.28 (#787)

* add jax benchmark notebook

* prepare 0.28 release, add experimental status of feglm and demean_jax to docs

* update release notes

* update lock file

Author: s3alfisc

docs/changelog.qmd

@@ -2,11 +2,16 @@
 
 ## PyFixest 0.28.0 (In Development, can be installed from github)
 
-- Adds a `pf.feglm()` function that supports GLMs with normal and binomial families (gaussian, logit, probit) without fixed effects. Fixed effects support is work in progress.
+### New features and bug fixes
 - Adds a function argument `context`, that allows to pass information / context to the `formulaic.Formulaic.get_model_matrix()` call that creates the model matrix.
 - Fix a bug that caused reindexing of `LPDID._coeftable` when calling `LPDID.iplot()`. As a result, a second call of `LPDID.iplot()` would fail.
 - Bumps the required `formulaic` version to `1.1.0` and fixes errors that arose when a) the ref argument was used for i() syntax, which led to a silent failure under formulaic >= 1.1.0, and fixef() / predict() with fixed effects, which led to a loud error.
 
+### New experimental Features
+- Adds a `pf.feglm()` function that supports GLMs with normal and binomial families (gaussian, logit, probit) without fixed effects. Fixed effects support is work in progress.
+- Adds options to run the demean function via JAX. This might speed up the model fit if GPU is available.
+
+
 ## PyFixest 0.27.0
 
 - Adds support for Gelbach's (JoLe 2016) Regression Decomposition method using a `decompose()` method for `Feols`.

docs/quickstart.qmd

@@ -411,7 +411,7 @@ fit.wald_test(R=R1, q=q1)
 
 # Other GLMs (without fixed effects)
 
-`PyFixest` supports a range of other GLMs without fixed effects (adding fixed effect support is WIP) via the `pf.feglm()` function.
+`PyFixest` experimentally supports a range of other GLMs without fixed effects (adding fixed effect support is WIP) via the `pf.feglm()` function. Full support with all bells and whistles (in particular, fixed effects demeaning) is planned for PyFixest 0.29.
 
 ```{python}
 data_glm = pf.get_data(N=100, seed = 170)

pixi.lock

@@ -123,6 +123,7 @@ def feols(
 
     demeaner_backend: DemeanerBackendOptions, optional
         The backend to use for demeaning. Can be either "numba" or "jax". Defaults to "numba".
+        The "jax" backend is experimental. A full release is planned for pyfixest 0.29.
 
     use_compression: bool
         Whether to use sufficient statistics to losslessly fit the regression model
@@ -573,7 +574,9 @@ def fepois(
 
     demeaner_backend: DemeanerBackendOptions, optional
         The backend to use for demeaning. Can be either "numba" or "jax".
-        Defaults to "numba".
+        Defaults to "numba". The "jax" backend is experimental.
+        A full release is planned for pyfixest 0.29.
+
 
     drop_intercept : bool, optional
         Whether to drop the intercept from the model, by default False.
@@ -746,6 +749,7 @@ def feglm(
 ) -> Union[Feols, Fepois, FixestMulti]:
     """
     Estimate GLM regression models (currently without fixed effects, this is work in progress).
+    This feature is currently experimental, full support will be released with pyfixest 0.29.
 
     Parameters
     ----------

pyfixest/estimation/estimation.py

@@ -2,7 +2,7 @@
 name = "pyfixest"
 description = "Fast high dimensional fixed effect estimation following syntax of the fixest R package."
 requires-python = ">= 3.9"
-version = "0.27.0"
+version = "0.28.0"
 authors = [
   { name = "Alexander Fischer", email = "[email protected]" },
   { name = "Styfen Schär" },