UPDATE: bo

bayeso/bo.py

@@ -1,6 +1,6 @@
 # bo
 # author: Jungtaek Kim (jtkim@postech.ac.kr)
-# last updated: June 20, 2018
+# last updated: June 23, 2018
 
 import numpy as np
 from scipy.optimize import minimize
@@ -13,18 +13,24 @@
 from bayeso import constants
 
 
+# TODO: I am not sure, but flatten() should be replaced.
 class BO():
     def __init__(self, arr_range,
         str_cov=constants.STR_GP_COV,
         str_acq=constants.STR_BO_ACQ,
         is_ard=True,
         prior_mu=None,
+        verbose=False,
+        debug=False,
     ):
-        # TODO: use is_ard
+        # TODO: use is_ard.
+        # TODO: add debug cases.
         assert isinstance(arr_range, np.ndarray)
         assert isinstance(str_cov, str)
         assert isinstance(str_acq, str)
         assert isinstance(is_ard, bool)
+        assert isinstance(verbose, bool)
+        assert isinstance(debug, bool)
         assert callable(prior_mu) or prior_mu is None
         assert len(arr_range.shape) == 2
         assert arr_range.shape[1] == 2
@@ -34,13 +40,19 @@ def __init__(self, arr_range,
         self.str_cov = str_cov
         self.str_acq = str_acq
         self.is_ard = is_ard
+        self.verbose = verbose
+        self.debug = debug
         self.prior_mu = prior_mu
 
     def _get_initial_grid(self, int_grid=constants.NUM_BO_GRID):
+        assert isinstance(int_grid, int)
+
         arr_initials = utils_bo.get_grid(self.arr_range, int_grid)
         return arr_initials
 
     def _get_initial_uniform(self, int_samples, int_seed=None):
+        assert isinstance(int_seed, int) or int_seed is None
+
         if int_seed is not None:
             np.random.seed(int_seed)
         list_initials = []
@@ -53,6 +65,8 @@ def _get_initial_uniform(self, int_samples, int_seed=None):
         return arr_initials
 
     def _get_initial_sobol(self, int_samples, int_seed=None):
+        assert isinstance(int_seed, int) or int_seed is None
+
         if int_seed is None:
             int_seed = np.random.randint(0, 10000)
         arr_samples = sobol_seq.i4_sobol_generate(self.num_dim, int_samples, int_seed)
@@ -64,10 +78,15 @@ def _get_initial_latin(self, int_samples):
 
     def _get_initial(self, str_initial_method,
         fun_objective=None,
-        int_samples=10,
-        int_seed=None
+        int_samples=constants.NUM_ACQ_SAMPLES,
+        int_seed=None,
     ):
+        assert callable(fun_objective) or fun_objective is None
+        assert isinstance(int_samples, int)
+        assert isinstance(int_seed, int) or int_seed is None
+
         if str_initial_method == 'grid':
+            assert fun_objective is not None
             arr_initials = self._get_initial_grid()
             arr_initials = utils_bo.get_best_acquisition(arr_initials, fun_objective)
         elif str_initial_method == 'uniform':
@@ -78,6 +97,9 @@ def _get_initial(self, str_initial_method,
             raise NotImplementedError('_get_initial: latin')
         else:
             raise ValueError('_get_initial: missing condition for str_initial_method')
+        if self.debug:
+            print('DEBUG: _get_initial: arr_initials')
+            print(arr_initials)
         return arr_initials
 
     def _optimize_objective(self, fun_acquisition, X_train, Y_train, X_test, cov_X_X, inv_cov_X_X, hyps):
@@ -86,7 +108,7 @@ def _optimize_objective(self, fun_acquisition, X_train, Y_train, X_test, cov_X_X
         acquisitions = fun_acquisition(pred_mean.flatten(), pred_std.flatten(), Y_train=Y_train)
         return acquisitions
 
-    def _optimize(self, fun_objective, str_initial_method='sobol', verbose=False):
+    def _optimize(self, fun_objective, str_initial_method):
         list_bounds = []
         for elem in self.arr_range:
             list_bounds.append(tuple(elem))
@@ -98,18 +120,25 @@ def _optimize(self, fun_objective, str_initial_method='sobol', verbose=False):
                 x0=arr_initial,
                 bounds=list_bounds,
                 method=constants.STR_OPTIMIZER_METHOD_BO,
-                options={'disp': verbose}
+                options={'disp': self.verbose}
             )
             list_next_point.append(next_point.x)
-            if verbose:
-                print('INFORM: optimized result for acq. ', next_point.x)
+            if self.verbose:
+                print('INFORM: _optimize: optimized result for acq. ', next_point.x)
         next_point = utils_bo.get_best_acquisition(np.array(list_next_point), fun_objective)
         return next_point.flatten()
 
-    def optimize(self, X_train, Y_train, is_grid_optimized=False, verbose=False):
-        cov_X_X, inv_cov_X_X, hyps = gp.get_optimized_kernel(X_train, Y_train, self.prior_mu, self.str_cov, verbose=verbose)
+    def optimize(self, X_train, Y_train, str_initial_method='sobol'):
+        assert isinstance(X_train, np.ndarray)
+        assert isinstance(Y_train, np.ndarray)
+        assert isinstance(str_initial_method, str)
+        assert len(X_train.shape) == 2
+        assert len(Y_train.shape) == 2
+        assert Y_train.shape[1] == 1
+        assert X_train.shape[0] == Y_train.shape[0]
+
+        cov_X_X, inv_cov_X_X, hyps = gp.get_optimized_kernel(X_train, Y_train, self.prior_mu, self.str_cov, verbose=self.verbose)
 
-        # NEED: to add acquisition function
         if self.str_acq == 'pi':
             fun_acquisition = acquisition.pi
         elif self.str_acq == 'ei':
@@ -120,5 +149,5 @@ def optimize(self, X_train, Y_train, is_grid_optimized=False, verbose=False):
             raise ValueError('optimize: missing condition for self.str_acq.')
 
         fun_objective = lambda X_test: -1.0 * constants.MULTIPLIER_ACQ * self._optimize_objective(fun_acquisition, X_train, Y_train, X_test, cov_X_X, inv_cov_X_X, hyps)
-        next_point = self._optimize(fun_objective, verbose=verbose)
+        next_point = self._optimize(fun_objective, str_initial_method=str_initial_method)
         return next_point, cov_X_X, inv_cov_X_X, hyps

bayeso/constants.py

@@ -12,8 +12,9 @@
 
 NUM_BO_GRID = 50
 NUM_BO_RANDOM = 1000
+NUM_ACQ_SAMPLES = 50
 
-MULTIPLIER_ACQ = 1000.0
+MULTIPLIER_ACQ = 10.0
 
 TIME_PAUSE = 2.0
 RANGE_SHADE = 1.96

bayeso/gp.py

@@ -63,7 +63,6 @@ def get_optimized_kernel(X_train, Y_train, prior_mu, str_cov, str_optimizer_meth
     assert len(X_train.shape) == 2
     assert len(Y_train.shape) == 2
     assert X_train.shape[0] == Y_train.shape[0]
-    assert str_optimizer_method == 'L-BFGS-B'
 
     prior_mu_train = get_prior_mu(prior_mu, X_train)
     num_dim = X_train.shape[1]
@@ -79,7 +78,7 @@ def get_optimized_kernel(X_train, Y_train, prior_mu, str_cov, str_optimizer_meth
     result_optimized = result_optimized.x
     hyps = utils_covariance.restore_hyps(str_cov, result_optimized)
     if verbose:
-        print('INFORM: optimized result for gpr ', hyps)
+        print('INFORM: get_optimized_kernel: optimized result for gpr ', hyps)
     cov_X_X, inv_cov_X_X = get_kernels(X_train, hyps, str_cov)
     return cov_X_X, inv_cov_X_X, hyps
 

examples/example_bo_ei.py

@@ -22,7 +22,7 @@ def main():
     X_test = np.reshape(X_test, (400, 1))
     for ind_ in range(1, 20+1):
         Y_train = fun_target(X_train)
-        next_x, cov_X_X, inv_cov_X_X, hyps = model_bo.optimize(X_train, fun_target(X_train))
+        next_x, cov_X_X, inv_cov_X_X, hyps = model_bo.optimize(X_train, fun_target(X_train), str_initial_method='grid')
         mu_test, sigma_test = gp.predict_test_(X_train, Y_train, X_test, cov_X_X, inv_cov_X_X, hyps)
         acq_test = acquisition.ei(mu_test.flatten(), sigma_test.flatten(), Y_train)
         acq_test = np.reshape(acq_test, (acq_test.shape[0], 1))

tests/test_gp.py

@@ -123,8 +123,6 @@ def test_get_optimized_kernel():
         gp.get_optimized_kernel(X, Y, prior_mu, 'abc')
     with pytest.raises(AssertionError) as error:
         gp.get_optimized_kernel(X, Y, prior_mu, 'se', str_optimizer_method=1)
-    with pytest.raises(AssertionError) as error:
-        gp.get_optimized_kernel(X, Y, prior_mu, 'se', str_optimizer_method='abc')
 
 def test_predict_test_():
     np.random.seed(42)