https://github.com/GPflow/GPflow
Tip revision: 512007d6a980e1b9f6ded12b0351b6c3f88f33e2 authored by Alexander G. de G. Matthews on 27 October 2016, 15:06:06 UTC
v0.3.3 paper submission. (#243)
v0.3.3 paper submission. (#243)
Tip revision: 512007d
test_param.py
# Copyright 2016 the GPflow authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.from __future__ import print_function
from functools import reduce
import unittest
import GPflow
import tensorflow as tf
import numpy as np
from GPflow import settings
float_type = settings.dtypes.float_type
np_float_type = np.float32 if float_type is tf.float32 else np.float64
try:
import cPickle as pickle
except ImportError:
import pickle
class NamingTests(unittest.TestCase):
def test_unnamed(self):
p = GPflow.param.Param(1)
self.assertTrue(p.name == 'unnamed')
def test_bad_parent(self):
p = GPflow.param.Param(1)
m = GPflow.model.Model()
p._parent = m # do not do this.
with self.assertRaises(ValueError):
print(p.name)
def test_two_parents(self):
m = GPflow.model.Model()
m.p = GPflow.param.Param(1)
m.p2 = m.p # do not do this!
with self.assertRaises(ValueError):
print(m.p.name)
class ParamTestsScalar(unittest.TestCase):
def setUp(self):
tf.reset_default_graph()
self.m = GPflow.param.Parameterized()
self.m.p = GPflow.param.Param(1.0)
def testAssign(self):
self.m.p = 2.0
self.assertTrue(isinstance(self.m.p, GPflow.param.Param))
self.assertTrue(self.m.get_free_state() == 2.0)
def testValue(self):
# make sure the correct value is returned
self.m.p = 3.0
self.assertTrue(isinstance(self.m.p.value, np.ndarray))
# make sure assignment does not work
with self.assertRaises(AttributeError):
self.m.p.value = 2.53
# make sure we get a copy
self.assertFalse(self.m.p.value is self.m.p._array)
def testReplacement(self):
old_p = self.m.p
new_p = GPflow.param.Param(3.0)
self.m.p = new_p
# Parameterized instances should not have _needs_recompile
self.assertFalse(hasattr(self.m, '_needs_recompile'))
self.assertFalse(old_p.highest_parent is self.m)
def testHighestParent(self):
self.assertTrue(self.m.p.highest_parent is self.m)
def testName(self):
self.assertTrue(self.m.p.name == 'p')
def testFixing(self):
self.m.p.fixed = False
self.m.fixed = True
self.assertTrue(self.m.p.fixed)
self.m.p.fixed = False
self.assertFalse(self.m.fixed)
def testFixedFreeState(self):
self.assertTrue(len(self.m.get_free_state()) == 1)
self.m.set_state(np.ones(1))
self.m.fixed = True
self.assertTrue(len(self.m.get_free_state()) == 0)
self.m.set_state(np.ones(0))
def testMakeTF(self):
x = tf.placeholder('float64')
l = self.m.make_tf_array(x)
self.assertTrue(l == 1)
l = self.m.p.make_tf_array(x)
self.assertTrue(l == 1)
def testFreeState(self):
xx = self.m.get_free_state()
self.assertTrue(np.allclose(xx, np.ones(1)))
y = np.array([34.0], np_float_type)
self.m.set_state(y)
self.assertTrue(np.allclose(self.m.get_free_state(), y))
def testFixed(self):
self.m.p.fixed = True
self.assertTrue(len(self.m.get_free_state()) == 0)
self.assertTrue(self.m.make_tf_array(tf.placeholder(float_type)) == 0)
def testRecompile(self):
self.m._needs_recompile = False
self.m.p.fixed = True
self.assertTrue(self.m._needs_recompile)
self.m._needs_recompile = False
self.m.p.prior = GPflow.priors.Gaussian(0, 1)
self.assertTrue(self.m._needs_recompile)
def testTFMode(self):
x = tf.placeholder('float64')
self.m.make_tf_array(x)
self.assertTrue(isinstance(self.m.p, GPflow.param.Param))
with self.m.tf_mode():
self.assertTrue(isinstance(self.m.p, tf.Tensor))
class ParamTestsDeeper(unittest.TestCase):
def setUp(self):
tf.reset_default_graph()
self.m = GPflow.param.Parameterized()
self.m.foo = GPflow.param.Parameterized()
self.m.foo.bar = GPflow.param.Parameterized()
self.m.foo.bar.baz = GPflow.param.Param(1.0)
def testHighestParent(self):
self.assertTrue(self.m.foo.highest_parent is self.m)
self.assertTrue(self.m.foo.bar.highest_parent is self.m)
self.assertTrue(self.m.foo.bar.baz.highest_parent is self.m)
def testReplacement(self):
old_p = self.m.foo.bar.baz
new_p = GPflow.param.Param(3.0)
self.m.foo.bar.baz = new_p
# Parameterized instances should not have _needs_recompile
self.assertFalse(hasattr(self.m, '_needs_recompile'))
self.assertFalse(old_p.highest_parent is self.m)
def testName(self):
self.assertTrue(self.m.foo.name == 'foo')
self.assertTrue(self.m.foo.bar.name == 'bar')
self.assertTrue(self.m.foo.bar.baz.name == 'baz')
def testMakeTF(self):
x = tf.placeholder('float64')
l = self.m.make_tf_array(x)
self.assertTrue(l == 1)
l = self.m.foo.make_tf_array(x)
self.assertTrue(l == 1)
l = self.m.foo.bar.make_tf_array(x)
self.assertTrue(l == 1)
l = self.m.foo.bar.baz.make_tf_array(x)
self.assertTrue(l == 1)
def testFreeState(self):
xx = self.m.get_free_state()
self.assertTrue(np.allclose(xx, np.ones(1)))
y = np.array([34.0], np_float_type)
self.m.set_state(y)
self.assertTrue(np.allclose(self.m.get_free_state(), y))
def testFixed(self):
self.m.foo.bar.baz.fixed = True
self.assertTrue(len(self.m.get_free_state()) == 0)
def testFixing(self):
self.m.fixed = False
self.m.foo.bar.fixed = True
self.assertTrue(self.m.fixed)
self.assertTrue(self.m.foo.fixed)
self.assertTrue(self.m.foo.bar.fixed)
self.assertTrue(self.m.foo.bar.baz.fixed)
self.m.foo.bar.baz.fixed = False
self.assertFalse(self.m.fixed)
self.assertFalse(self.m.foo.fixed)
self.assertFalse(self.m.foo.bar.fixed)
self.assertFalse(self.m.foo.bar.baz.fixed)
def testRecompile(self):
self.m._needs_recompile = False
self.m.foo.bar.baz.fixed = True
self.assertTrue(self.m._needs_recompile)
self.m._needs_recompile = False
self.m.foo.bar.baz.prior = GPflow.priors.Gaussian(0, 1)
self.assertTrue(self.m._needs_recompile)
def testTFMode(self):
x = tf.placeholder('float64')
self.m.make_tf_array(x)
self.assertTrue(isinstance(self.m.foo.bar.baz, GPflow.param.Param))
with self.m.tf_mode():
self.assertTrue(isinstance(self.m.foo.bar.baz, tf.Tensor))
class ParamTestsWider(unittest.TestCase):
def setUp(self):
tf.reset_default_graph()
self.m = GPflow.param.Parameterized()
self.m.foo = GPflow.param.Param(1.0)
self.m.bar = GPflow.param.Param(np.arange(10))
self.m.baz = GPflow.param.Param(np.random.randn(3, 3))
def testHighestParent(self):
self.assertTrue(self.m.foo.highest_parent is self.m)
self.assertTrue(self.m.bar.highest_parent is self.m)
self.assertTrue(self.m.baz.highest_parent is self.m)
def testName(self):
self.assertTrue(self.m.foo.name == 'foo')
self.assertTrue(self.m.bar.name == 'bar')
self.assertTrue(self.m.baz.name == 'baz')
def testMakeTF(self):
x = tf.placeholder('float64')
l = self.m.make_tf_array(x)
self.assertTrue(l == 20)
l = self.m.foo.make_tf_array(x)
self.assertTrue(l == 1)
l = self.m.bar.make_tf_array(x)
self.assertTrue(l == 10)
l = self.m.baz.make_tf_array(x)
self.assertTrue(l == 9)
def testFreeState(self):
xx = self.m.get_free_state()
self.assertTrue(len(xx) == 20)
y = np.random.randn(20)
self.m.set_state(y)
self.assertTrue(np.allclose(self.m.get_free_state(), y))
def testIndexParam(self):
fs = self.m.get_free_state()
for p in [self.m.foo, self.m.bar, self.m.baz]:
index, found = self.m.get_param_index(p)
self.assertTrue(found)
self.assertTrue(fs[index] == p.get_free_state()[0])
def testFixed(self):
self.m.foo.fixed = True
self.assertTrue(len(self.m.get_free_state()) == 19)
self.m.foo.fixed = False
self.m.bar.fixed = True
self.assertTrue(len(self.m.get_free_state()) == 10)
def testFixing(self):
self.m.fixed = False
self.m.foo.fixed = True
self.assertFalse(self.m.fixed)
self.assertTrue(self.m.foo.fixed)
self.assertFalse(self.m.bar.fixed)
self.assertFalse(self.m.baz.fixed)
self.m.bar.fixed = True
self.m.baz.fixed = True
self.assertTrue(self.m.fixed)
self.assertTrue(self.m.foo.fixed)
self.assertTrue(self.m.bar.fixed)
self.assertTrue(self.m.baz.fixed)
def testRecompile(self):
self.m._needs_recompile = False
self.m.foo.fixed = True
self.assertTrue(self.m._needs_recompile)
self.m._needs_recompile = False
self.m.bar.prior = GPflow.priors.Gaussian(0, 1)
self.assertTrue(self.m._needs_recompile)
def testTFMode(self):
x = tf.placeholder('float64')
self.m.make_tf_array(x)
self.assertTrue(all([isinstance(p, GPflow.param.Param) for p in (self.m.foo, self.m.bar, self.m.baz)]))
with self.m.tf_mode():
self.assertTrue(all([isinstance(p, tf.Tensor)
for p in (self.m.foo, self.m.bar, self.m.baz)]))
class TestParamList(unittest.TestCase):
def test_construction(self):
GPflow.param.ParamList([])
GPflow.param.ParamList([GPflow.param.Param(1)])
with self.assertRaises(AssertionError):
GPflow.param.ParamList([GPflow.param.Param(1), 'stringsnotallowed'])
def test_naming(self):
p1 = GPflow.param.Param(1.2)
p2 = GPflow.param.Param(np.array([3.4, 5.6], np_float_type))
GPflow.param.ParamList([p1, p2])
self.assertTrue(p1.name == 'item0')
self.assertTrue(p2.name == 'item1')
def test_connected(self):
p1 = GPflow.param.Param(1.2)
p2 = GPflow.param.Param(np.array([3.4, 5.6], np_float_type))
l = GPflow.param.ParamList([p1, p2])
x = l.get_free_state()
x.sort()
self.assertTrue(np.all(x == np.array([1.2, 3.4, 5.6], np_float_type)))
def test_setitem(self):
p1 = GPflow.param.Param(1.2)
p2 = GPflow.param.Param(np.array([3.4, 5.6], np_float_type))
l = GPflow.param.ParamList([p1, p2])
l[0] = 1.2
self.assertTrue(p1._array == 1.2)
l[1] = np.array([1.1, 2.2], np_float_type)
self.assertTrue(np.all(p2._array == np.array([1.1, 2.2], np_float_type)))
with self.assertRaises(TypeError):
l[0] = GPflow.param.Param(12)
def test_append(self):
p1 = GPflow.param.Param(1.2)
p2 = GPflow.param.Param(np.array([3.4, 5.6], np_float_type))
l = GPflow.param.ParamList([p1])
l.append(p2)
self.assertTrue(p2 in l.sorted_params)
with self.assertRaises(AssertionError):
l.append('foo')
def test_with_parameterized(self):
pzd = GPflow.param.Parameterized()
p = GPflow.param.Param(1.2)
pzd.p = p
l = GPflow.param.ParamList([pzd])
# test assignment:
l[0].p = 5
self.assertTrue(l.get_free_state() == 5)
# test to make sure tf_mode get turned on and off
self.assertFalse(pzd._tf_mode)
with l.tf_mode():
self.assertTrue(pzd._tf_mode)
self.assertFalse(pzd._tf_mode)
def test_in_model(self):
class Foo(GPflow.model.Model):
def __init__(self):
GPflow.model.Model.__init__(self)
self.l = GPflow.param.ParamList([
GPflow.param.Param(1), GPflow.param.Param(12)])
def build_likelihood(self):
return -reduce(tf.add, [tf.square(x) for x in self.l])
m = Foo()
self.assertTrue(m.get_free_state().size == 2)
m.optimize(disp=False)
atol = 1e-6 if np_float_type is np.float32 else 1e-8
self.assertTrue(np.allclose(m.get_free_state(), 0., atol=atol))
class TestPickleAndDict(unittest.TestCase):
def setUp(self):
rng = np.random.RandomState(0)
X = rng.randn(10, 1)
Y = rng.randn(10, 1)
self.m = GPflow.gpr.GPR(X, Y, kern=GPflow.kernels.RBF(1))
def test(self):
# pickle and reload the model
s1 = pickle.dumps(self.m)
m1 = pickle.loads(s1)
d1 = self.m.get_parameter_dict()
d2 = m1.get_parameter_dict()
for key, val in d1.items():
assert np.all(val == d2[key])
class TestDictEmpty(unittest.TestCase):
def setUp(self):
self.m = GPflow.model.Model()
def test(self):
d = self.m.get_parameter_dict()
self.assertTrue(len(d.keys()) == 0)
self.m.set_parameter_dict(d)
class TestDictSimple(unittest.TestCase):
def setUp(self):
self.m = GPflow.model.Model()
self.m.p1 = GPflow.param.Param(np.random.randn(3, 2))
self.m.p2 = GPflow.param.Param(np.random.randn(10))
def test(self):
d = self.m.get_parameter_dict()
self.assertTrue(len(d.keys()) == 2)
state1 = self.m.get_free_state().copy()
self.m.set_state(state1 * 0)
self.m.set_parameter_dict(d)
self.assertTrue(np.all(state1 == self.m.get_free_state()))
class TestDictSVGP(unittest.TestCase):
def setUp(self):
self.rng = np.random.RandomState(0)
X = self.rng.randn(10, 1)
Y = self.rng.randn(10, 1)
Z = self.rng.randn(5, 1)
self.m = GPflow.svgp.SVGP(X, Y, Z=Z, likelihood=GPflow.likelihoods.Gaussian(), kern=GPflow.kernels.RBF(1))
def test(self):
loglik1 = self.m.compute_log_likelihood()
d = self.m.get_parameter_dict()
# muck up the model
self.m.set_state(self.rng.randn(self.m.get_free_state().size))
loglik2 = self.m.compute_log_likelihood()
# reset the model
self.m.set_parameter_dict(d)
loglik3 = self.m.compute_log_likelihood()
self.assertFalse(np.allclose(loglik1, loglik2))
self.assertTrue(np.allclose(loglik1, loglik3))
class TestScopes(unittest.TestCase):
def setUp(self):
rng = np.random.RandomState(0)
X = rng.randn(10, 1)
k = GPflow.kernels.RBF(1)
Y = rng.randn(10, 1)
self.m = GPflow.gpr.GPR(X, Y, k)
self.m._compile()
def test_kern_name(self):
with self.m.tf_mode():
l = self.m.build_likelihood()
self.assertTrue('model.build_likelihood' in l.name)
def test_likelihood_name(self):
with self.m.tf_mode():
K = self.m.kern.K(self.m.X)
self.assertTrue('kern.K' in K.name)
if __name__ == "__main__":
unittest.main()