rename ba_init, get_params, rfun (#95)

docs/source/examples/hoburgabbeel_ex6_1_output.txt

@@ -7,9 +7,9 @@ w**3.4411 = 0.422736 * (u_1)**-2.14843
     + 0.424169 * (u_1)**-2.14784
     + 0.15339 * (u_1)**0.584654
 ISMA fit from params
-1 = (0.994797/w**0.238961) * (u_1)**-0.138389
-    + (0.949519/w**0.0924504) * (u_1)**0.0165798
-    + (0.967646/w**0.115505) * (u_1)**-0.0131876
+1 = (0.992648/w**0.35353) * (u_1)**-0.204093
+    + (0.947302/w**0.0920266) * (u_1)**0.017725
+    + (0.961409/w**0.11673) * (u_1)**-0.011164
 MA RMS Error: 0.0023556
 SMA RMS Error: 2.3856e-05
-ISMA RMS Error: 1.0765e-06
+ISMA RMS Error: 8.0757e-07

gpfit/fit.py

@@ -2,35 +2,41 @@
 from numpy import ones, exp, sqrt, mean, square, hstack
 from .classes import max_affine, softmax_affine, implicit_softmax_affine
 from .levenberg_marquardt import levenberg_marquardt
-from .ba_init import ba_init
+from .initialize import get_initial_parameters
 from .print_fit import print_ISMA, print_SMA, print_MA
 from .constraint_set import FitConstraintSet
 
-ALPHA_INIT = 10
-RFUN = {"ISMA": implicit_softmax_affine,
-        "SMA": softmax_affine,
-        "MA": max_affine}
+ALPHA0 = 10
+CLASSES = {
+    "ISMA": implicit_softmax_affine,
+    "SMA": softmax_affine,
+    "MA": max_affine,
+}
+
 
 # pylint: disable=invalid-name
-def get_params(ftype, K, xdata, ydata):
+def get_parameters(ftype, K, xdata, ydata):
     "Perform least-squares fitting optimization."
-    def rfun(params):
+
+    ydata_col = ydata.reshape(ydata.size, 1)
+    ba = get_initial_parameters(xdata, ydata_col, K).flatten('F')
+
+    def residual(params):
         "A specific residual function."
-        [yhat, drdp] = RFUN[ftype](xdata, params)
+        [yhat, drdp] = CLASSES[ftype](xdata, params)
         r = yhat - ydata
         return r, drdp
 
-    ba = ba_init(xdata, ydata.reshape(ydata.size, 1), K).flatten('F')
-
     if ftype == "ISMA":
-        params, _ = levenberg_marquardt(rfun, hstack((ba, ALPHA_INIT*ones(K))))
+        params, _ = levenberg_marquardt(residual, hstack((ba, ALPHA0*ones(K))))
     elif ftype == "SMA":
-        params, _ = levenberg_marquardt(rfun, hstack((ba, ALPHA_INIT)))
+        params, _ = levenberg_marquardt(residual, hstack((ba, ALPHA0)))
     else:
-        params, _ = levenberg_marquardt(rfun, ba)
+        params, _ = levenberg_marquardt(residual, ba)
 
     return params
 
+
 # pylint: disable=too-many-locals
 # pylint: disable=too-many-branches
 # pylint: disable=import-error
@@ -75,7 +81,7 @@ def fit(xdata, ydata, K, ftype="ISMA"):
             fitdata["lb%d" % i] = exp(min(xdata.T[i]))
             fitdata["ub%d" % i] = exp(max(xdata.T[i]))
 
-    params = get_params(ftype, K, xdata, ydata)
+    params = get_parameters(ftype, K, xdata, ydata)
 
     # A: exponent parameters, B: coefficient parameters
     A = params[[i for i in range(K*(d+1)) if i % (d + 1) != 0]]
@@ -125,7 +131,7 @@ def evaluate(xdata):
             ydata:      Dependent variable data in 1D numpy array
         """
         xdata = xdata.reshape(xdata.size, 1) if xdata.ndim == 1 else xdata.T
-        return RFUN[ftype](xdata, params)[0]
+        return CLASSES[ftype](xdata, params)[0]
 
     # cstrt.evaluate = evaluate
     fitdata["rms_err"] = sqrt(mean(square(evaluate(xdata.T)-ydata)))

gpfit/ba_init.py → gpfit/initialize.py

@@ -1,11 +1,11 @@
-"Implements ba_init"
+"Implements get_initial_parameters"
 from numpy import ones, hstack, zeros, tile, argmin
 from numpy.linalg import lstsq, matrix_rank
 from numpy.random import permutation as randperm
 
 
 # pylint: disable=too-many-locals
-def ba_init(x, y, K):
+def get_initial_parameters(x, y, K):
     """
     Initializes max-affine fit to data (y, x)
     ensures that initialization has at least K+1 points per partition (i.e.
@@ -49,7 +49,8 @@ def ba_init(x, y, K):
     # loop through each partition, making local fits
     # note we expand partitions that result in singular least squares problems
     # why this way? some points will be shared by multiple partitions, but
-    # resulting max-affine fit will tend to be good. (as opposed to solving least-norm version)
+    # resulting max-affine fit will tend to be good. (as opposed to solving
+    # least-norm version)
     for k in range(K):
         inds = mindistind == k
 
@@ -73,11 +74,10 @@ def ba_init(x, y, K):
                 inds[sortdistind[i]] = 1
 
             if options['bverbose']:
-                print("ba_init: Added %s points to partition %s to maintain"
-                      "full rank for local fitting." % (i-iinit, k))
+                print("Initialization: Added %s points to partition %s to "
+                      "maintain full rank for local fitting." % (i-iinit, k))
         # now create the local fit
         b[:, k] = lstsq(X[inds.nonzero()], y[inds.nonzero()], rcond=-1)[0][:, 0]
         # Rank condition specified to default for python upgrades
 
-
     return b

gpfit/tests/run_tests.py

@@ -8,8 +8,8 @@
 from gpfit.tests import t_logsumexp
 TESTS += t_logsumexp.TESTS
 
-from gpfit.tests import t_ba_init
-TESTS += t_ba_init.TESTS
+from gpfit.tests import t_initialize
+TESTS += t_initialize.TESTS
 
 from gpfit.tests import t_classes
 TESTS += t_classes.TESTS

gpfit/tests/t_ba_init.py → gpfit/tests/t_initialize.py

@@ -1,9 +1,9 @@
-"unit tests for ba_init function"
+"unit tests for get_initial_parameters function"
 import unittest
 import numpy as np
 from numpy import arange, newaxis, vstack, log, exp
 from numpy.random import random_sample
-from gpfit.ba_init import ba_init
+from gpfit.initialize import get_initial_parameters
 
 SEED = 33404
 
@@ -16,7 +16,7 @@ class TestMaxAffineInitK2(unittest.TestCase):
     x = arange(0., 16.)[:, newaxis]
     y = arange(0., 16.)[:, newaxis]
     K = 2
-    ba = ba_init(x, y, K)
+    ba = get_initial_parameters(x, y, K)
 
     def test_ba_ndim_k2(self):
         self.assertEqual(self.ba.ndim, 2)
@@ -43,7 +43,7 @@ class TestMaxAffineInitK4(unittest.TestCase):
     y = y.reshape(y.size, 1)
     K = 4
 
-    ba = ba_init(x, y, K)
+    ba = get_initial_parameters(x, y, K)
 
     def test_ba_shape_k4(self):
         self.assertEqual(self.ba.shape, (3, 4))