Pass `fixef_tol` parameter to `scipy` and `cupy` estimators

[s3alfisc]

At the moment, the fixef_tol flag is not passed to both estimators.

Probably suffices to set tolerance here if self.fixef_tol is passed?

pyfixest/pyfixest/estimation/cupy/demean_cupy_.py

Line 171 in e5eb9ff

if self.solver_maxiter is None:

New issue to partially fix / address #1042 (comment).

FYI @rhstanton.

[rhstanton]

Editing the code to this:

        if self.solver_maxiter is None:
            self.solver_maxiter = maxiter
        if tol is not None:
            print(f"DEBUG: Setting LSMR tolerances to {tol}")
            self.solver_atol = tol
            self.solver_btol = tol

causes the test file to generate the following output:

Numba backend:
/home/stanton/projects/pyfixest/pyfixest/estimation/feols_.py:2756: UserWarning: 
            20 variables dropped due to multicollinearity.
            The following variables are dropped: 
    C(relative_year, contr.treatment(base=-1))[T.-10]
    C(relative_year, contr.treatment(base=-1))[T.-9]
    C(relative_year, contr.treatment(base=-1))[T.-8]
    C(relative_year, contr.treatment(base=-1))[T.-7]
    C(relative_year, contr.treatment(base=-1))[T.-6]
    ....
            
  warnings.warn(
  Coefficients: 1, Dropped: 20
CuPy backend:
DEBUG: Setting LSMR tolerances to 1e-06
  Coefficients: 21, Dropped: 0
CuPy backend stric fixef_tol:
DEBUG: Setting LSMR tolerances to 1e-12
/home/stanton/projects/pyfixest/pyfixest/estimation/feols_.py:2756: UserWarning: 
            20 variables dropped due to multicollinearity.
            The following variables are dropped: 
    C(relative_year, contr.treatment(base=-1))[T.-10]
    C(relative_year, contr.treatment(base=-1))[T.-9]
    C(relative_year, contr.treatment(base=-1))[T.-8]
    C(relative_year, contr.treatment(base=-1))[T.-7]
    C(relative_year, contr.treatment(base=-1))[T.-6]
    ....
            
  warnings.warn(
  Coefficients: 1, Dropped: 20

This looks pretty hopeful, but there are a lot of similar-sounding variables. Does this sound about right?

[rhstanton]

When I use the code with this edit in my own example where the results were wrong, passing fixef_to=1e-12, it runs for several hours (compared with a few minutes total for my Julia code) before giving up, and the output from my code looks like this:

Regression formula = est_sqft ~ i(relative_year, ref=-1) + i(relative_year, control0, ref=-1) + control0 | sa_property_id + year^fire_id
Using pyfixest v0.40.1
Demeaning backend: CUPY (GPU)
/home/stanton/projects/pyfixest/pyfixest/estimation/model_matrix_fixest_.py:221: UserWarning: 375 singleton fixed effect(s) detected. These observations are dropped from the model.
  warnings.warn(
DEBUG: Setting LSMR tolerances to 1e-12
CUPY failed (ValueError: Demeaning failed after 500000 iterations.); trying Rust...
Demeaning backend: Rust (CPU)
/home/stanton/projects/pyfixest/pyfixest/estimation/model_matrix_fixest_.py:221: UserWarning: 375 singleton fixed effect(s) detected. These observations are dropped from the model.
  warnings.warn(