python-fedci 0.1.3__tar.gz → 0.1.5__tar.gz

{python_fedci-0.1.3 → python_fedci-0.1.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-fedci
-Version: 0.1.3
+Version: 0.1.5
 Summary: A small package for federated independence tests
 Author-email: Maximilian Hahn <max.hahn@gmx.de>
 License: SPDX-License-Identifier: AGPL-3.0-or-later

{python_fedci-0.1.3 → python_fedci-0.1.5}/fedci/client.py

@@ -7,13 +7,17 @@ import statsmodels.api as sm
 import statsmodels.genmod.families.family as fam
 from scipy.special import softmax
 
-from .env import get_env_additive_masking, get_env_client_heterogeniety, get_env_fit_intercept
+from .env import (
+    get_env_additive_masking,
+    get_env_client_heterogeniety,
+    get_env_fit_intercept,
+)
 from .utils import (
     BetaUpdateData,
     VariableType,
     categorical_separator,
-    constant_colname,
     client_colname,
+    constant_colname,
     ordinal_separator,
     polars_dtype_map,
 )
@@ -99,7 +103,7 @@ class ComputationHelper:
         family: DistributionalFamily,
     ):
         eta = np.zeros_like(y)
-        if get_env_client_heterogeniety()==1:
+        if get_env_client_heterogeniety() == 1:
             alpha = ComputationHelper.fit_local_alpha(y=y, offset=eta, family=family)
             eta += alpha
         mu: np.ndarray = family.inverse_link(eta)
@@ -115,11 +119,11 @@ class ComputationHelper:
             "xwx": xwx,
             "xwz": xwz,
             "rss": 0,
-            "n": 0,
+            "n": y.shape[0],
         }
         if family == Gaussian:
             result["rss"] = np.sum((y - mu) ** 2).item()
-            result["n"] = y.shape[0]
+            # result["n"] = y.shape[0]
         return result
 
     @staticmethod
@@ -128,7 +132,7 @@ class ComputationHelper:
         family: DistributionalFamily,
     ):
         eta = np.zeros_like(y)
-        if get_env_client_heterogeniety()==1:
+        if get_env_client_heterogeniety() == 1:
             alpha = ComputationHelper.fit_local_alpha(y=y, offset=eta, family=family)
             eta += alpha
         else:
@@ -146,7 +150,7 @@ class ComputationHelper:
     ):
         eta: np.ndarray = X @ beta
 
-        if get_env_client_heterogeniety()==1:
+        if get_env_client_heterogeniety() == 1:
             alpha = ComputationHelper.fit_local_alpha(y=y, offset=eta, family=family)
         else:
             alpha = np.zeros_like(eta)
@@ -188,10 +192,10 @@ class ComputationHelper:
         xwx, xwz = ComputationHelper.get_irls_step(y, X, eta, mu, alpha, family)
         llf: float = family.loglik(y, mu)
         # only use non-dummy values in rss and n for gaussian regression
-        result = {"llf": llf, "xwx": xwx, "xwz": xwz, "rss": 0, "n": 0}
+        result = {"llf": llf, "xwx": xwx, "xwz": xwz, "rss": 0, "n": y.shape[0]}
         if family == Gaussian:
             result["rss"] = np.sum((y - mu) ** 2).item()
-            result["n"] = y.shape[0]
+            # result["n"] = y.shape[0]
         return result
 
     @staticmethod
@@ -357,7 +361,7 @@ class CategoricalComputationUnit(ComputationUnit):
             # Reshape beta (K x (J-1))
             beta = beta.reshape(num_features, -1, order="F")
 
-        if get_env_client_heterogeniety()==1:
+        if get_env_client_heterogeniety() == 1:
             if X is None:
                 offset = np.zeros_like(y)
             else:
@@ -385,7 +389,7 @@ class CategoricalComputationUnit(ComputationUnit):
             xwx = np.empty((0, 0))
             xwz = np.empty((0, 0))
 
-            return {"llf": llf, "xwx": xwx, "xwz": xwz, "rss": 0, "n": 0}
+            return {"llf": llf, "xwx": xwx, "xwz": xwz, "rss": 0, "n": y.shape[0]}
 
         # Compute eta and mu
         eta = X @ beta  # N x (J-1)
@@ -424,11 +428,17 @@ class CategoricalComputationUnit(ComputationUnit):
             XWX += Xi.T @ Wi @ Xi
             XWz += Xi.T @ Wi @ z_i
 
-        logprob = np.log(np.clip(mu, 1e-8, 1-1e-8))
+        logprob = np.log(np.clip(mu, 1e-8, 1 - 1e-8))
         llf = np.sum(y_full * logprob)
 
         # only use non-dummy values in rss and n for gaussian regression
-        return {"llf": llf, "xwx": XWX, "xwz": XWz.reshape(-1, 1), "rss": 0, "n": 0}
+        return {
+            "llf": llf,
+            "xwx": XWX,
+            "xwz": XWz.reshape(-1, 1),
+            "rss": 0,
+            "n": y.shape[0],
+        }
 
 
 class OrdinalComputationUnit:  # (ComputationUnit):
@@ -534,7 +544,7 @@ class OrdinalComputationUnit:  # (ComputationUnit):
             llf += np.sum(np.log(np.take(mu_diff, current_level_indices.nonzero()[0])))
         mu_diff = mus_diff[-1]
         llf += np.sum(np.log(np.take(mu_diff, reference_level_indices.nonzero()[0])))
-        result = {"llf": llf, "xwx": xwx, "xwz": xwz, "rss": 0, "n": 0}
+        result = {"llf": llf, "xwx": xwx, "xwz": xwz, "rss": 0, "n": y.shape[0]}
         return result
 
 
@@ -554,8 +564,10 @@ class Client:
         self.id = id
         self.data: pl.DataFrame = data
 
-        if get_env_client_heterogeniety()==2:
-            self.data = self.data.with_columns(pl.lit(str(self.id)).alias(client_colname))
+        if get_env_client_heterogeniety() == 2:
+            self.data = self.data.with_columns(
+                pl.lit(str(self.id)).alias(client_colname)
+            )
 
         self.schema: Dict[str, VariableType] = {
             column: polars_dtype_map[dtype]
@@ -792,10 +804,7 @@ class Client:
                 new_cond_vars.append(constant_colname)
             cond_vars = new_cond_vars
             result = regression_computation_map[self.schema[resp_var]].compute(
-                self.expanded_data,
-                resp_var,
-                cond_vars,
-                beta
+                self.expanded_data, resp_var, cond_vars, beta
             )
             if len(self.contributing_clients) > 0:
                 result = self.apply_masks(test_key, result)

{python_fedci-0.1.3 → python_fedci-0.1.5}/fedci/testing.py

@@ -30,6 +30,7 @@ class RegressionTest:
         self.num_classes, self.num_parameters = params
         self.dof = self.num_classes * self.num_parameters
         self.beta = np.zeros((self.dof, 1))
+        self.num_samples = None
         # self.beta = np.random.randn(self.dof, 1)
         self.alpha = 1.0
 
@@ -132,6 +133,10 @@ class RegressionTest:
         xwx = sum([_update.xwx for _update in update])
         xwz = sum([_update.xwz for _update in update])
         n = int(np.sum([_update.n for _update in update]).item())
+        assert self.num_samples is None or self.num_samples == n, (
+            "Number of supporting samples changed during test"
+        )
+        self.num_samples = n
 
         if abs(llf) < 1e-10:
             self.early_stop = True
@@ -232,8 +237,9 @@ class LikelihoodRatioTest:
         self.bad_fit = False
 
         self.p_value: Optional[float] = None
-        self.chi2stat = None
-        self.dof = None
+        self.chi2stat: Optional[float] = None
+        self.dof: Optional[int] = None
+        self.num_samples: Optional[int] = None
 
     def __repr__(self):
         if self.p_value is None:
@@ -324,6 +330,11 @@ class LikelihoodRatioTest:
                     for _update in update
                 ]
             )
+        if self.num_samples is None:
+            assert (
+                self.restricted_test.num_samples == self.unrestricted_test.num_samples
+            ), "Restricted and Unrestricted Test mustn't differ in supporting samples"
+            self.num_samples = self.restricted_test.num_samples
 
     def get_iterations(self):
         return max([self.restricted_test.iterations, self.unrestricted_test.iterations])
@@ -336,8 +347,8 @@ class LikelihoodRatioTest:
         t1_dof = self.unrestricted_test.dof
 
         self.chi2stat = 2 * (t1_llf - t0_llf)
-        self.dofs = t1_dof - t0_dof
-        self.p_value = scipy.stats.chi2.sf(self.chi2stat, self.dofs).item()
+        self.dof = t1_dof - t0_dof
+        self.p_value = scipy.stats.chi2.sf(self.chi2stat, self.dof).item()
 
         if get_env_debug() >= 2:
             print(
@@ -371,6 +382,7 @@ class SymmetricLikelihoodRatioTest:
             self.lrt2: LikelihoodRatioTest = lrt1
 
         self.p_value: Optional[float] = None
+        self.num_samples: Optional[int] = None
 
     def __repr__(self):
         if self.p_value is None:
@@ -418,6 +430,11 @@ class SymmetricLikelihoodRatioTest:
             self.lrt2.update_parameters(
                 [_update[self.lrt2.response] for _update in update]
             )
+        if self.num_samples is None:
+            assert self.lrt1.num_samples == self.lrt2.num_samples, (
+                "Restricted and Unrestricted Test mustn't differ in supporting samples"
+            )
+            self.num_samples = self.lrt1.num_samples
 
     def get_iterations(self):
         return max([self.lrt1.get_iterations(), self.lrt2.get_iterations()])

{python_fedci-0.1.3 → python_fedci-0.1.5}/pyproject.toml

@@ -5,7 +5,7 @@ build-backend = "setuptools.build_meta"
 [project]
 name = "python-fedci"
 authors = [{ name = "Maximilian Hahn", email = "max.hahn@gmx.de" }]
-version = "0.1.3"
+version = "0.1.5"
 license = { file = "LICENSE" }
 description = "A small package for federated independence tests"
 readme = "README.md"

{python_fedci-0.1.3 → python_fedci-0.1.5}/python_fedci.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-fedci
-Version: 0.1.3
+Version: 0.1.5
 Summary: A small package for federated independence tests
 Author-email: Maximilian Hahn <max.hahn@gmx.de>
 License: SPDX-License-Identifier: AGPL-3.0-or-later