ValueError: Unknown solver highs-ds

[anilsh]

In ExponentiatedGradient, the following error is coming. The error goes away after making run_linprog_step=False but this slows down the code and doesn't converge in even 100 iterations.

Also, the arguments doesn't support providing another solver.

The stack trace is:

Cell In[13], line 25, in reduction_exp_gradient(base_model)
     23     # train model 
---> 25     model_exp_grad.fit(gfr.X_train, gfr.y_train, sensitive_features=gfr.S_train)


File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\fairlearn\reductions\_exponentiated_gradient\exponentiated_gradient.py:200, in ExponentiatedGradient.fit(self, X, y, **kwargs)
    196     gap_LP = np.PINF
    197 else:
    198     # saddle point optimization over the convex hull of
    199     # classifiers returned so far
--> 200     Q_LP, self.lambda_vecs_LP_[t], result_LP = lagrangian.solve_linprog(
    201         self.nu
    202     )
    203     gap_LP = result_LP.gap()
    205 # keep values from exponentiated gradient or linear programming

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\fairlearn\reductions\_exponentiated_gradient\_lagrangian.py:168, in _Lagrangian.solve_linprog(self, nu)
    166 A_eq = np.concatenate((np.ones((1, n_hs)), np.zeros((1, 1))), axis=1)
    167 b_eq = np.ones(1)
--> 168 result = opt.linprog(
    169     c, A_ub=A_ub, b_ub=b_ub, A_eq=A_eq, b_eq=b_eq, method="highs-ds"
    170 )
    171 Q = pd.Series(result.x[:-1], self.hs.index)
    172 dual_c = np.concatenate((b_ub, -b_eq))

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\scipy\optimize\_linprog.py:554, in linprog(c, A_ub, b_ub, A_eq, b_eq, bounds, method, callback, options, x0)
    550         x, status, message, iteration = _linprog_rs(
    551             c, c0=c0, A=A, b=b, x0=x0, callback=callback,
    552             postsolve_args=postsolve_args, **solver_options)
    553     else:
--> 554         raise ValueError('Unknown solver %s' % method)
    556 # Eliminate artificial variables, re-introduce presolved variables, etc.
    557 # need modified bounds here to translate variables appropriately
    558 disp = solver_options.get('disp', False)

ValueError: Unknown solver highs-ds

[anilsh]

It worked after upgrading scipy to '1.10.0' but it creates problem with LogisticRegression().

Cell In[19], line 28, in reduction_exp_gradient(base_model)
     26     # train model 
     27     print ('Training base model using specified constraints..')
---> 28     model_exp_grad.fit(gfr.X_train, gfr.y_train, sensitive_features=gfr.S_train)
     29 #     self.inprocess['model'] = model_exp_grad
     30 
     31     # make predictions
     32     y_pred = model_exp_grad.predict(gfr.X_test) 

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\fairlearn\reductions\_exponentiated_gradient\exponentiated_gradient.py:168, in ExponentiatedGradient.fit(self, X, y, **kwargs)
    165 lambda_EG = self.lambda_vecs_EG_.mean(axis=1)
    167 # select classifier according to best_h method
--> 168 h, h_idx = lagrangian.best_h(lambda_vec)
    170 if t == 0:
    171     if self.nu is None:

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\fairlearn\reductions\_exponentiated_gradient\_lagrangian.py:234, in _Lagrangian.best_h(self, lambda_vec)
    228 def best_h(self, lambda_vec):
    229     """Solve the best-response problem.
    230 
    231     Returns the classifier that solves the best-response problem for
    232     the vector of Lagrange multipliers `lambda_vec`.
    233     """
--> 234     classifier = self._call_oracle(lambda_vec)
    236     def h(X):
    237         pred = classifier.predict(X)

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\fairlearn\reductions\_exponentiated_gradient\_lagrangian.py:222, in _Lagrangian._call_oracle(self, lambda_vec)
    219     estimator = clone(estimator=self.estimator, safe=False)
    221 oracle_call_start_time = time()
--> 222 estimator.fit(self.constraints.X, redY, **{self.sample_weight_name: redW})
    223 self.oracle_execution_times.append(time() - oracle_call_start_time)
    224 self.n_oracle_calls += 1

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\sklearn\linear_model\_logistic.py:1407, in LogisticRegression.fit(self, X, y, sample_weight)
   1405 else:
   1406     prefer = 'processes'
-> 1407 fold_coefs_ = Parallel(n_jobs=self.n_jobs, verbose=self.verbose,
   1408                        **_joblib_parallel_args(prefer=prefer))(
   1409     path_func(X, y, pos_class=class_, Cs=[C_],
   1410               l1_ratio=self.l1_ratio, fit_intercept=self.fit_intercept,
   1411               tol=self.tol, verbose=self.verbose, solver=solver,
   1412               multi_class=multi_class, max_iter=self.max_iter,
   1413               class_weight=self.class_weight, check_input=False,
   1414               random_state=self.random_state, coef=warm_start_coef_,
   1415               penalty=penalty, max_squared_sum=max_squared_sum,
   1416               sample_weight=sample_weight)
   1417     for class_, warm_start_coef_ in zip(classes_, warm_start_coef))
   1419 fold_coefs_, _, n_iter_ = zip(*fold_coefs_)
   1420 self.n_iter_ = np.asarray(n_iter_, dtype=np.int32)[:, 0]

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\joblib\parallel.py:1048, in Parallel.__call__(self, iterable)
   1039 try:
   1040     # Only set self._iterating to True if at least a batch
   1041     # was dispatched. In particular this covers the edge
   (...)
   1045     # was very quick and its callback already dispatched all the
   1046     # remaining jobs.
   1047     self._iterating = False
-> 1048     if self.dispatch_one_batch(iterator):
   1049         self._iterating = self._original_iterator is not None
   1051     while self.dispatch_one_batch(iterator):

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\joblib\parallel.py:864, in Parallel.dispatch_one_batch(self, iterator)
    862     return False
    863 else:
--> 864     self._dispatch(tasks)
    865     return True

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\joblib\parallel.py:782, in Parallel._dispatch(self, batch)
    780 with self._lock:
    781     job_idx = len(self._jobs)
--> 782     job = self._backend.apply_async(batch, callback=cb)
    783     # A job can complete so quickly than its callback is
    784     # called before we get here, causing self._jobs to
    785     # grow. To ensure correct results ordering, .insert is
    786     # used (rather than .append) in the following line
    787     self._jobs.insert(job_idx, job)

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\joblib\_parallel_backends.py:208, in SequentialBackend.apply_async(self, func, callback)
    206 def apply_async(self, func, callback=None):
    207     """Schedule a func to be run"""
--> 208     result = ImmediateResult(func)
    209     if callback:
    210         callback(result)

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\joblib\_parallel_backends.py:572, in ImmediateResult.__init__(self, batch)
    569 def __init__(self, batch):
    570     # Don't delay the application, to avoid keeping the input
    571     # arguments in memory
--> 572     self.results = batch()

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\joblib\parallel.py:263, in BatchedCalls.__call__(self)
    259 def __call__(self):
    260     # Set the default nested backend to self._backend but do not set the
    261     # change the default number of processes to -1
    262     with parallel_backend(self._backend, n_jobs=self._n_jobs):
--> 263         return [func(*args, **kwargs)
    264                 for func, args, kwargs in self.items]

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\joblib\parallel.py:263, in <listcomp>(.0)
    259 def __call__(self):
    260     # Set the default nested backend to self._backend but do not set the
    261     # change the default number of processes to -1
    262     with parallel_backend(self._backend, n_jobs=self._n_jobs):
--> 263         return [func(*args, **kwargs)
    264                 for func, args, kwargs in self.items]

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\sklearn\linear_model\_logistic.py:762, in _logistic_regression_path(X, y, pos_class, Cs, fit_intercept, max_iter, tol, verbose, solver, coef, class_weight, dual, penalty, intercept_scaling, multi_class, random_state, check_input, max_squared_sum, sample_weight, l1_ratio)
    755     iprint = [-1, 50, 1, 100, 101][
    756         np.searchsorted(np.array([0, 1, 2, 3]), verbose)]
    757     opt_res = optimize.minimize(
    758         func, w0, method="L-BFGS-B", jac=True,
    759         args=(X, target, 1. / C, sample_weight),
    760         options={"iprint": iprint, "gtol": tol, "maxiter": max_iter}
    761     )
--> 762     n_iter_i = _check_optimize_result(
    763         solver, opt_res, max_iter,
    764         extra_warning_msg=_LOGISTIC_SOLVER_CONVERGENCE_MSG)
    765     w0, loss = opt_res.x, opt_res.fun
    766 elif solver == 'newton-cg':

File ~\Anaconda3\envs\sheyzam-fairness-env\lib\site-packages\sklearn\utils\optimize.py:243, in _check_optimize_result(solver, result, max_iter, extra_warning_msg)
    235 if solver == "lbfgs":
    236     if result.status != 0:
    237         warning_msg = (
    238             "{} failed to converge (status={}):\n{}.\n\n"
    239             "Increase the number of iterations (max_iter) "
    240             "or scale the data as shown in:\n"
    241             "    https://scikit-learn.org/stable/modules/"
    242             "preprocessing.html"
--> 243         ).format(solver, result.status, result.message.decode("latin1"))
    244         if extra_warning_msg is not None:
    245             warning_msg += "\n" + extra_warning_msg

AttributeError: 'str' object has no attribute 'decode'

[romanlutz]

@anilsh we have a bug template that provides instructions such as providing the versions of several key packages. Without that it's hard to impossible to help. Can you try adding this information?

[anilsh]

Hi @romanlutz , the code works file when I use Scipy==1.6.0.
In 1.10.0, the solver throws an error str object has no member decode.

Fairlearn allow any version of Scipy above 1.4.1 but higher ones have changed significantly.

[romanlutz]

@anilsh can you please follow the instructions from the bug template to print all dependencies? This sort of simple thing saves us several back-and-forth questions and speeds things up considerably. Thanks!

[anilsh]

@romanlutz Can you please share the link to bug template? Couldn't find it.

[romanlutz]

When you open an issue there's a template for "bug report". Just follow the instructions it contains.