upgrade API to torchlambertw=0.0.3; 10x faster MLE

Author: gmgeorg

pylambertw/__init__.py

@@ -0,0 +1,4 @@
+"""Init for module"""
+from ._version import __version__
+
+__all__ = ["__version__"]

pylambertw/_version.py

@@ -0,0 +1,3 @@
+"""Version."""
+
+__version__ = "0.0.2"

pylambertw/base.py

@@ -158,12 +158,12 @@ def __repr__(self):
     def tau(self):
         """Converts Theta (distribution dependent) to Tau (transformation only)."""
 
-        distr_constr = lwd.get_distribution_constructor(self.distribution_name)
+        distr_constr = lwd.utils.get_distribution_constructor(self.distribution_name)
         distr = distr_constr(**self.beta)
 
         return Tau(
             loc=distr.mean.numpy()
-            if lwd.is_location_family(self.distribution_name)
+            if lwd.utils.is_location_family(self.distribution_name)
             else 0.0,
             scale=distr.stddev.numpy(),
             lambertw_params=self.lambertw_params,

pylambertw/mle.py

@@ -50,7 +50,7 @@ def __init__(
         self.distribution_name = distribution_name
         self.distribution_constructor = (
             distribution_constructor
-            or lwd.get_distribution_constructor(self.distribution_name)
+            or lwd.utils.get_distribution_constructor(self.distribution_name)
         )
         self.lambertw_type = base.LambertWType(lambertw_type)
         self.max_iter = max_iter
@@ -73,14 +73,14 @@ def _initialize_params(self, data: np.ndarray):
         if self.lambertw_type == base.LambertWType.H:
             self.igmm = igmm.IGMM(
                 lambertw_type=self.lambertw_type,
-                location_family=lwd.is_location_family(self.distribution_name),
+                location_family=lwd.utils.is_location_family(self.distribution_name),
             )
             self.igmm.fit(data)
             x_init = self.igmm.transform(data)
 
             lambertw_params_init = self.igmm.tau.lambertw_params
         else:
-            if lwd.is_location_family(self.distribution_name):
+            if lwd.utils.is_location_family(self.distribution_name):
                 # Default to Normal distriubtion for location family.
                 params_data = ud.estimate_params(data, "Normal")
                 loc_init = params_data["loc"]
@@ -92,7 +92,7 @@ def _initialize_params(self, data: np.ndarray):
                 scale_init = 1.0 / params_data["rate"]
 
             z_init = (data - loc_init) / scale_init
-            if lwd.is_location_family(self.distribution_name):
+            if lwd.utils.is_location_family(self.distribution_name):
                 gamma_init = igmm.gamma_taylor(z_init)
             else:
                 gamma_init = 0.01

pylambertw/tests/test_utils.py

@@ -25,6 +25,6 @@ def test_estimate_params(dist_name):
     rng = np.random.RandomState(42)
     x = rng.normal(100)
     params = ud.estimate_params(x, dist_name)
-    constr = lwd.get_distribution_constructor(dist_name)
-    param_names = lwd.get_distribution_args(constr)
+    constr = lwd.utils.get_distribution_constructor(dist_name)
+    param_names = lwd.utils.get_distribution_args(constr)
     assert set(params.keys()) == set(param_names)

pylambertw/utils/distributions.py

@@ -36,8 +36,8 @@ def torch_sigmoid(x: torch.tensor) -> torch.tensor:
 def get_params_activations(distribution_name: str) -> Dict[str, Callable]:
     """Get activation functions for each distribution parameters."""
     assert isinstance(distribution_name, str)
-    distr_constr = lwd.get_distribution_constructor(distribution_name)
-    param_names = lwd.get_distribution_args(distr_constr)
+    distr_constr = lwd.utils.get_distribution_constructor(distribution_name)
+    param_names = lwd.utils.get_distribution_args(distr_constr)
 
     act_fns = {p: (torch_linear, linear_inverse) for p in param_names}
 

setup.py

@@ -1,14 +1,23 @@
 from setuptools import find_packages, setup
+import re
+
+_VERSION_FILE = "pylambertw/_version.py"
+verstrline = open(_VERSION_FILE, "rt").read()
+_VERSION = r"^__version__ = ['\"]([^'\"]*)['\"]"
+mo = re.search(_VERSION, verstrline, re.M)
+if mo:
+    verstr = mo.group(1)
+else:
+    raise RuntimeError("Unable to find version string in %s." % (_VERSION_FILE,))
 
 pkg_descr = """
 Python implementation of the Lambert W x F framework for analyzing skewed, heavy-tailed distribution
 with an sklearn interface and torch based maximum likelihood estimation (MLE).
 """
 
-
 setup(
     name="pylambertw",
-    version="0.0.1",
+    version=verstr,
     url="https://github.com/gmgeorg/pylambertw.git",
     author="Georg M. Goerg",
     author_email="[email protected]",
@@ -25,6 +34,6 @@
         "tqdm>=4.46.1",
         "dataclasses>=0.6",
         "scikit-learn>=1.0.1",
-        "torchlambertw @ git+ssh://[email protected]/gmgeorg/torchlambertw.git#egg=torchlambertw-0.0.1",
+        "torchlambertw @ git+ssh://[email protected]/gmgeorg/torchlambertw.git#egg=torchlambertw-0.0.3",
     ],
 )