Merge pull request #1 from tpospisi/dev

Merging `dev` branch: implementing a `fit` method in NNKCDE
lee-group-cmu · Jun 27, 2019 · ed2df5e · ed2df5e
2 parents 3ec9762 + f1d51a8
commit ed2df5e
Show file tree

Hide file tree

Showing 9 changed files with 16 additions and 15 deletions.
diff --git a/python/setup.py b/python/setup.py
@@ -6,7 +6,7 @@
     README_TEXT = f.read()
 
 setup(name="nnkcde",
-      version="0.2",
+      version="0.3",
       license="MIT",
       description="Fits nearest neighbor kernel conditional density estimates",
       long_description = README_TEXT,

diff --git a/python/src/nnkcde/core.py b/python/src/nnkcde/core.py
@@ -5,9 +5,10 @@
 from .kde import kde
 
 class NNKCDE(object):
-    def __init__(self, x_train, z_train, k=None):
+    def __init__(self, k=None):
         self.k = k
 
+    def fit(self, x_train, z_train):
         if len(z_train.shape) == 1:
             z_train = z_train.reshape(-1, 1)
         if len(x_train.shape) == 1:

diff --git a/python/src/nnkcde/kde.py b/python/src/nnkcde/kde.py
@@ -1,8 +1,6 @@
 """Functions for kernel density estimation."""
 
-import numpy as np
 import statsmodels.api as sm
-import scipy.stats
 
 def kde(responses, grid, bandwidth):
     """Calculates the kernel density estimate.
@@ -35,15 +33,14 @@ def kde(responses, grid, bandwidth):
 
     n_grid, n_dim = grid.shape
     n_obs, _ = responses.shape
-    density = np.zeros(n_grid)
 
     if n_dim == 1:
         kde = sm.nonparametric.KDEUnivariate(responses[:, 0])
-        kde.fit(bw = bandwidth, fft = False)
+        kde.fit(bw=bandwidth, fft=False)
         return kde.evaluate(grid[:, 0])
     else:
         if isinstance(bandwidth, (float, int)):
             bandwidth = [bandwidth] * n_dim
-        kde = sm.nonparametric.KDEMultivariate(responses, var_type = "c" * n_dim,
-                                               bw = bandwidth)
+        kde = sm.nonparametric.KDEMultivariate(responses, var_type="c" * n_dim,
+                                               bw=bandwidth)
         return kde.pdf(grid)
diff --git a/python/tests/__pycache__/test_loss_estimation.cpython-35-PYTEST.pyc b/python/tests/__pycache__/test_loss_estimation.cpython-35-PYTEST.pyc
diff --git a/python/tests/__pycache__/test_performance.cpython-35-PYTEST.pyc b/python/tests/__pycache__/test_performance.cpython-35-PYTEST.pyc
diff --git a/python/tests/__pycache__/test_univariate.cpython-35-PYTEST.pyc b/python/tests/__pycache__/test_univariate.cpython-35-PYTEST.pyc
diff --git a/python/tests/test_loss_estimation.py b/python/tests/test_loss_estimation.py
@@ -19,11 +19,11 @@ def generate_data(n):
     x_train, z_train = generate_data(1000)
     x_test, z_test = generate_data(1000)
 
-    obj = nnkcde.NNKCDE(x_train, z_train)
+    obj = nnkcde.NNKCDE()
+    obj.fit(x_train, z_train)
     n_grid = 1000
     z_grid = np.linspace(-5.0, 5.0, n_grid)
 
-#    for bandwidth in (0.1, 1.0, 3.0):
     for bandwidth in (1.0, 1.0):
         for k in (2, 5, 10, 100):
             cde = obj.predict(x_test, z_grid, k=k, bandwidth=bandwidth)

diff --git a/python/tests/test_performance.py b/python/tests/test_performance.py
@@ -22,7 +22,8 @@ def generate_data(n):
     k = 100
     bandwidth = 0.1
 
-    obj = nnkcde.NNKCDE(x_train, z_train, k=k)
+    obj = nnkcde.NNKCDE(k=k)
+    obj.fit(x_train, z_train)
     n_grid = 1000
     z_grid = np.linspace(0, 1, n_grid)
     density = obj.predict(x_test, z_grid, bandwidth=bandwidth)

diff --git a/python/tests/test_univariate.py b/python/tests/test_univariate.py
@@ -12,13 +12,14 @@ def test_k_equals_1():
     x_train = np.random.uniform(0, 1, n_train)
     z_train = np.random.normal(x_train, 1, n_train)
 
-    obj = nnkcde.NNKCDE(x_train, z_train)
+    obj = nnkcde.NNKCDE()
+    obj.fit(x_train, z_train)
 
     for bandwidth in (0.1, 0.3, 1.0, 2.0, 3.0):
         x_obs = np.random.uniform(0, 1, 1)
         nearest = np.argmin(np.abs(x_train - x_obs))
         expected = scipy.stats.norm.pdf(z_grid, z_train[nearest], bandwidth).reshape(1, -1)
-        preds = obj.predict(x_obs, z_grid, k = 1, bandwidth = bandwidth)
+        preds = obj.predict(x_obs, z_grid, k=1, bandwidth=bandwidth)
         np.testing.assert_almost_equal(expected, preds)
 
 def test_k_equals_n():
@@ -30,10 +31,11 @@ def test_k_equals_n():
     x_train = np.random.uniform(0, 1, n_train)
     z_train = np.random.normal(x_train, 1, n_train)
 
-    obj = nnkcde.NNKCDE(x_train, z_train)
+    obj = nnkcde.NNKCDE()
+    obj.fit(x_train, z_train)
 
     for bandwidth in (0.1, 0.3, 1.0, 2.0, 3.0):
         x_obs = np.random.uniform(0, 1, 1)
         expected = nnkcde.kde.kde(z_train, z_grid, bandwidth).reshape(1, -1)
-        preds = obj.predict(x_obs, z_grid, k = n_train, bandwidth = bandwidth)
+        preds = obj.predict(x_obs, z_grid, k=n_train, bandwidth=bandwidth)
         np.testing.assert_almost_equal(expected, preds)