Merge pull request #12 from jungtaekkim/add-pures

Add pures

bayeso/acquisition.py

@@ -72,3 +72,31 @@ def aei(pred_mean, pred_std, Y_train, noise,
     ei = (np.min(Y_train) - pred_mean) * scipy.stats.norm.cdf(val_z) + pred_std * scipy.stats.norm.pdf(val_z)
     aei = ei * (1.0 - noise / np.sqrt(pred_std**2 + noise**2))
     return aei
+
+# pred_std and Y_train are ignored.
+def pure_exploit(pred_mean, pred_std=None, Y_train=None):
+    assert isinstance(pred_mean, np.ndarray)
+    assert isinstance(pred_std, np.ndarray) or pred_std is None
+    assert isinstance(Y_train, np.ndarray) or Y_train is None
+    assert len(pred_mean.shape) == 1
+    if pred_std is not None:
+        assert len(pred_std.shape) == 1
+        assert pred_mean.shape[0] == pred_std.shape[0]
+    if Y_train is not None:
+        assert len(Y_train.shape) == 2
+
+    return -pred_mean
+
+# pred_mean and Y_train are ignored.
+def pure_explore(pred_std, pred_mean=None, Y_train=None):
+    assert isinstance(pred_mean, np.ndarray) or pred_mean is None
+    assert isinstance(pred_std, np.ndarray)
+    assert isinstance(Y_train, np.ndarray) or Y_train is None
+    assert len(pred_std.shape) == 1
+    if pred_mean is not None:
+        assert len(pred_mean.shape) == 1
+        assert pred_mean.shape[0] == pred_std.shape[0]
+    if Y_train is not None:
+        assert len(Y_train.shape) == 2
+
+    return pred_std

bayeso/bo.py

@@ -113,7 +113,8 @@ def get_initial(self, str_initial_method,
     def _optimize_objective(self, fun_acquisition, X_train, Y_train, X_test, cov_X_X, inv_cov_X_X, hyps):
         X_test = np.atleast_2d(X_test)
         pred_mean, pred_std = gp.predict_test_(X_train, Y_train, X_test, cov_X_X, inv_cov_X_X, hyps, self.str_cov, self.prior_mu)
-        acquisitions = fun_acquisition(pred_mean.flatten(), pred_std.flatten(), Y_train=Y_train)
+        # no matter which acquisition functions are given, we input pred_mean, pred_std, and Y_train.
+        acquisitions = fun_acquisition(pred_mean=pred_mean.flatten(), pred_std=pred_std.flatten(), Y_train=Y_train)
         return acquisitions
 
     def _optimize(self, fun_negative_acquisition, str_initial_method, int_samples):
@@ -171,6 +172,10 @@ def optimize(self, X_train, Y_train,
             fun_acquisition = acquisition.ucb
         elif self.str_acq == 'aei':
             fun_acquisition = lambda pred_mean, pred_std, Y_train: acquisition.aei(pred_mean, pred_std, Y_train, hyps['noise'])
+        elif self.str_acq == 'pure_exploit':
+            fun_acquisition = acquisition.pure_exploit
+        elif self.str_acq == 'pure_explore':
+            fun_acquisition = acquisition.pure_explore
         else:
             raise NotImplementedError('optimize: allowed str_acq, but it is not implemented.')
 

bayeso/constants.py

@@ -32,7 +32,7 @@
 RANGE_SHADE = 1.96
 
 ALLOWED_GP_COV = ['se', 'matern32', 'matern52']
-ALLOWED_BO_ACQ = ['pi', 'ei', 'ucb', 'aei']
+ALLOWED_BO_ACQ = ['pi', 'ei', 'ucb', 'aei', 'pure_exploit', 'pure_explore']
 ALLOWED_INITIALIZATIONS_BO = ['sobol', 'uniform', 'latin']
 ALLOWED_INITIALIZATIONS_AO = ALLOWED_INITIALIZATIONS_BO + ['grid']
 

examples/example_basics_bo.py

@@ -0,0 +1,47 @@
+# example_basics_bo
+# author: Jungtaek Kim (jtkim@postech.ac.kr)
+# last updated: July 12, 2018
+
+import numpy as np
+import os
+
+from bayeso import gp
+from bayeso import bo
+from bayeso import acquisition
+from bayeso.utils import utils_plotting
+
+
+PATH_SAVE = './figures/bo/'
+
+def fun_target(X):
+    return 4.0 * np.cos(X) + 0.1 * X + 2.0 * np.sin(X) + 0.4 * (X - 0.5)**2
+
+def main():
+    num_iter = 10
+    X_train = np.array([
+        [-5],
+        [-1],
+        [1],
+        [2],
+    ])
+    num_init = X_train.shape[0]
+    model_bo = bo.BO(np.array([[-6., 6.]]))
+    X_test = np.linspace(-6, 6, 400)
+    X_test = np.reshape(X_test, (400, 1))
+    for ind_ in range(1, num_iter + 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), str_initial_method='uniform')
+        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.expand_dims(acq_test, axis=1)
+        X_train = np.vstack((X_train, next_x))
+        Y_train = fun_target(X_train)
+        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='basics_bo_' + str(ind_), int_init=num_init)
+        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='basics_bo_' + str(ind_), int_init=num_init, is_acq_axis_small=True)
+
+
+if __name__ == '__main__':
+    if not os.path.isdir(PATH_SAVE):
+        os.makedirs(PATH_SAVE)
+    main()
+

examples/example_basics_bo_aei.py → examples/example_bo_aei.py

@@ -1,4 +1,4 @@
-# example_basics_bo_aei
+# example_bo_aei
 # author: Jungtaek Kim (jtkim@postech.ac.kr)
 # last updated: July 12, 2018
 
@@ -37,8 +37,8 @@ def main():
         acq_test = np.expand_dims(acq_test, axis=1)
         X_train = np.vstack((X_train, next_x))
         Y_train = fun_target(X_train)
-        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bascis_{}_'.format(str_acq) + str(ind_), int_init=num_init)
-        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='basics_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
+        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init)
+        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
 
 
 if __name__ == '__main__':

examples/example_basics_bo_ei.py → examples/example_bo_ei.py

@@ -1,4 +1,4 @@
-# example_basics_bo_ei
+# example_bo_ei
 # author: Jungtaek Kim (jtkim@postech.ac.kr)
 # last updated: July 12, 2018
 
@@ -37,8 +37,8 @@ def main():
         acq_test = np.expand_dims(acq_test, axis=1)
         X_train = np.vstack((X_train, next_x))
         Y_train = fun_target(X_train)
-        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bascis_{}_'.format(str_acq) + str(ind_), int_init=num_init)
-        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='basics_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
+        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init)
+        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
 
 
 if __name__ == '__main__':

examples/example_basics_bo_pi.py → examples/example_bo_pi.py

@@ -1,4 +1,4 @@
-# example_basics_bo_pi
+# example_bo_pi
 # author: Jungtaek Kim (jtkim@postech.ac.kr)
 # last updated: July 12, 2018
 
@@ -37,8 +37,8 @@ def main():
         acq_test = np.expand_dims(acq_test, axis=1)
         X_train = np.vstack((X_train, next_x))
         Y_train = fun_target(X_train)
-        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bascis_{}_'.format(str_acq) + str(ind_), int_init=num_init)
-        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='basics_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
+        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init)
+        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
 
 
 if __name__ == '__main__':

examples/example_bo_pure_exploit.py

@@ -0,0 +1,48 @@
+# example_bo_pure_exploit
+# author: Jungtaek Kim (jtkim@postech.ac.kr)
+# last updated: July 12, 2018
+
+import numpy as np
+import os
+
+from bayeso import gp
+from bayeso import bo
+from bayeso import acquisition
+from bayeso.utils import utils_plotting
+
+
+PATH_SAVE = './figures/bo/'
+
+def fun_target(X):
+    return 4.0 * np.cos(X) + 0.1 * X + 2.0 * np.sin(X) + 0.4 * (X - 0.5)**2
+
+def main():
+    str_acq = 'pure_exploit'
+    num_iter = 10
+    X_train = np.array([
+        [-5],
+        [-1],
+        [1],
+        [2],
+    ])
+    num_init = X_train.shape[0]
+    model_bo = bo.BO(np.array([[-6., 6.]]), str_acq=str_acq)
+    X_test = np.linspace(-6, 6, 400)
+    X_test = np.reshape(X_test, (400, 1))
+    for ind_ in range(1, num_iter + 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), str_initial_method='uniform')
+        mu_test, sigma_test = gp.predict_test_(X_train, Y_train, X_test, cov_X_X, inv_cov_X_X, hyps)
+        acq_test = acquisition.pure_exploit(mu_test.flatten())
+        acq_test = np.expand_dims(acq_test, axis=1)
+        X_train = np.vstack((X_train, next_x))
+        Y_train = fun_target(X_train)
+        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init)
+        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
+
+
+if __name__ == '__main__':
+    if not os.path.isdir(PATH_SAVE):
+        os.makedirs(PATH_SAVE)
+    main()
+

examples/example_bo_pure_explore.py

@@ -0,0 +1,48 @@
+# example_bo_pure_explore
+# author: Jungtaek Kim (jtkim@postech.ac.kr)
+# last updated: July 12, 2018
+
+import numpy as np
+import os
+
+from bayeso import gp
+from bayeso import bo
+from bayeso import acquisition
+from bayeso.utils import utils_plotting
+
+
+PATH_SAVE = './figures/bo/'
+
+def fun_target(X):
+    return 4.0 * np.cos(X) + 0.1 * X + 2.0 * np.sin(X) + 0.4 * (X - 0.5)**2
+
+def main():
+    str_acq = 'pure_explore'
+    num_iter = 10
+    X_train = np.array([
+        [-5],
+        [-1],
+        [1],
+        [2],
+    ])
+    num_init = X_train.shape[0]
+    model_bo = bo.BO(np.array([[-6., 6.]]), str_acq=str_acq)
+    X_test = np.linspace(-6, 6, 400)
+    X_test = np.reshape(X_test, (400, 1))
+    for ind_ in range(1, num_iter + 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), str_initial_method='uniform')
+        mu_test, sigma_test = gp.predict_test_(X_train, Y_train, X_test, cov_X_X, inv_cov_X_X, hyps)
+        acq_test = acquisition.pure_explore(sigma_test.flatten())
+        acq_test = np.expand_dims(acq_test, axis=1)
+        X_train = np.vstack((X_train, next_x))
+        Y_train = fun_target(X_train)
+        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init)
+        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
+
+
+if __name__ == '__main__':
+    if not os.path.isdir(PATH_SAVE):
+        os.makedirs(PATH_SAVE)
+    main()
+

examples/example_basics_bo_ucb.py → examples/example_bo_ucb.py

@@ -1,4 +1,4 @@
-# example_basics_bo_ucb
+# example_bo_ucb
 # author: Jungtaek Kim (jtkim@postech.ac.kr)
 # last updated: July 12, 2018
 
@@ -37,8 +37,8 @@ def main():
         acq_test = np.expand_dims(acq_test, axis=1)
         X_train = np.vstack((X_train, next_x))
         Y_train = fun_target(X_train)
-        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bascis_{}_'.format(str_acq) + str(ind_), int_init=num_init)
-        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='basics_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
+        utils_plotting.plot_bo_step(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init)
+        utils_plotting.plot_bo_step_acq(X_train, Y_train, X_test, fun_target(X_test), mu_test, sigma_test, acq_test, path_save=PATH_SAVE, str_postfix='bo_{}_'.format(str_acq) + str(ind_), int_init=num_init, is_acq_axis_small=True)
 
 
 if __name__ == '__main__':

setup.py

@@ -10,7 +10,7 @@
 
 setup(
     name='bayeso',
-    version='0.1.0',
+    version='0.2.3',
     author='Jungtaek Kim',
     author_email='jtkim@postech.ac.kr',
     url='https://github.com/jungtaekkim/bayeso',

tests/test_acquisition.py

@@ -106,3 +106,42 @@ def test_aei():
     truth_val_acq = np.array([1.16847489e-01, 2.44025364e-02, 2.48686922e-03, 1.11930407e-04, 2.09279771e-06, 1.56585558e-08, 4.57958958e-11, 5.15587486e-14, 2.21142019e-17, 3.58729395e-21])
     assert (np.abs(val_acq - truth_val_acq) < TEST_EPSILON).all()
 
+def test_pure_exploit():
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_exploit('abc')
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_exploit(np.ones((10, 1)))
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_exploit(np.ones(10), pred_std='abc')
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_exploit(np.ones(10), Y_train='abc')
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_exploit(np.ones(10), pred_std=np.ones((10, 1)))
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_exploit(np.ones(10), pred_std=np.ones(20))
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_exploit(np.ones(10), Y_train=np.ones(20))
+
+    val_acq = acquisition.pure_exploit(np.arange(0, 10))
+    truth_val_acq = -np.arange(0, 10)
+    assert (np.abs(val_acq - truth_val_acq) < TEST_EPSILON).all()
+
+def test_pure_explore():
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_explore('abc')
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_explore(np.ones((10, 1)))
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_explore(np.ones(10), pred_mean='abc')
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_explore(np.ones(10), Y_train='abc')
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_explore(np.ones(10), pred_mean=np.ones((10, 1)))
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_explore(np.ones(10), pred_mean=np.ones(20))
+    with pytest.raises(AssertionError) as error:
+        acquisition.pure_explore(np.ones(10), Y_train=np.ones(20))
+
+    val_acq = acquisition.pure_explore(np.arange(0, 10))
+    truth_val_acq = np.arange(0, 10)
+    assert (np.abs(val_acq - truth_val_acq) < TEST_EPSILON).all()

tests/test_bo.py

@@ -204,3 +204,33 @@ def test_optimize():
     assert next_point.shape[0] == dim_X
     assert next_points.shape[1] == dim_X
     assert next_points.shape[0] == acquisitions.shape[0]
+
+    model_bo = bo.BO(arr_range_1, str_acq='pure_exploit')
+    next_point, next_points, acquisitions, cov_X_X, inv_cov_X_X, hyps = model_bo.optimize(X, Y)
+    assert isinstance(next_point, np.ndarray)
+    assert isinstance(next_points, np.ndarray)
+    assert isinstance(acquisitions, np.ndarray)
+    assert isinstance(cov_X_X, np.ndarray)
+    assert isinstance(inv_cov_X_X, np.ndarray)
+    assert isinstance(hyps, dict)
+    assert len(next_point.shape) == 1
+    assert len(next_points.shape) == 2
+    assert len(acquisitions.shape) == 1
+    assert next_point.shape[0] == dim_X
+    assert next_points.shape[1] == dim_X
+    assert next_points.shape[0] == acquisitions.shape[0]
+
+    model_bo = bo.BO(arr_range_1, str_acq='pure_explore')
+    next_point, next_points, acquisitions, cov_X_X, inv_cov_X_X, hyps = model_bo.optimize(X, Y)
+    assert isinstance(next_point, np.ndarray)
+    assert isinstance(next_points, np.ndarray)
+    assert isinstance(acquisitions, np.ndarray)
+    assert isinstance(cov_X_X, np.ndarray)
+    assert isinstance(inv_cov_X_X, np.ndarray)
+    assert isinstance(hyps, dict)
+    assert len(next_point.shape) == 1
+    assert len(next_points.shape) == 2
+    assert len(acquisitions.shape) == 1
+    assert next_point.shape[0] == dim_X
+    assert next_points.shape[1] == dim_X
+    assert next_points.shape[0] == acquisitions.shape[0]