class TestGM_Q1Lognormal(unittest.TestCase): # Tests lognormal
def setUp(self):
self.experiment = SerialExperiment(
bandit_algo=GM_BanditAlgo(
bandit=hyperopt.bandits.Q1Lognormal(),
good_estimator=IndependentAdaptiveParzenEstimator(),
bad_estimator=IndependentAdaptiveParzenEstimator()))
def test_optimize_20(self):
self.experiment.run(50)
import matplotlib.pyplot as plt
plt.subplot(1,2,1)
plt.plot(self.experiment.losses())
plt.subplot(1,2,2)
if 0:
plt.hist(
[t['x'] for t in self.experiment.trials],
bins=20)
else:
plt.scatter(
[t['x'] for t in self.experiment.trials],
range(len(self.experiment.trials)))
print self.experiment.losses()
print 'MIN', min(self.experiment.losses())
assert min(self.experiment.losses()) < .01
if 0:
plt.show()
def test_fit_quantized_lognormal(self):
bandit_algo = GPAlgo(self.qln_bandit)
bandit_algo.n_startup_jobs = 5
serial_exp = SerialExperiment(bandit_algo)
serial_exp.run(bandit_algo.n_startup_jobs)
# check that the Lognormal kernel has been
# identified as refinable
k = bandit_algo.kernels[0]
assert bandit_algo.is_refinable[k]
assert bandit_algo.bounds[k][0] > 0
serial_exp.run(25)
xvec = numpy.asarray([t['x'] for t in serial_exp.trials])
if 0:
show_bandit_algo(bandit_algo,
serial_exp.trials,
serial_exp.results,
xlim_low=1,
xlim_high=xvec.max() + 1,
)
assert min(serial_exp.losses()) == 0, (
serial_exp.losses(), min(serial_exp.losses()))
# check that all points were positive
assert xvec.min() > 0
# assert that the step size was respected
assert numpy.all(numpy.fmod(xvec, 1) == 0)
# the lenscale is about 1.8 Is that about right? What's right?
print bandit_algo.kernels[0].lenscale()
def test_2var_equal(self):
algo = GPAlgo(TestGaussian2D.Bandit(1, 1))
algo.n_startup_jobs = 5
se = SerialExperiment(algo)
se.run(25)
l0 = algo.kernels[0].lenscale()
l1 = algo.kernels[1].lenscale()
assert .85 < l0 / l1 < 1.15
# XXX: consider using this tighter bound
# when the mean and std are estimated from the
# startup jobs.
#assert min(se.losses()) < .005, min(se.losses())
assert min(se.losses()) < .05, min(se.losses())
def test_basic(self):
self.algo.n_startup_jobs = 7
n_iter = 40
serial_exp = SerialExperiment(self.algo)
serial_exp.run(self.algo.n_startup_jobs)
serial_exp.run(n_iter)
assert min(serial_exp.losses()) < 1e-2
def test_2var_unequal(self):
algo = GPAlgo(TestGaussian2D.Bandit(1, 0))
algo.n_startup_jobs = 25
se = SerialExperiment(algo)
se.run(50)
l0 = algo.kernels[0].lenscale()
l1 = algo.kernels[1].lenscale()
#N.B. a ratio in log-length scales is a big difference!
assert l1 / l0 > 3
assert min(se.losses()) < .005
class TestGM_DummyDBN(unittest.TestCase):
def setUp(self):
self.experiment = SerialExperiment(
bandit_algo=GM_BanditAlgo(
bandit=Dummy_DBN_Base(),
good_estimator=IndependentAdaptiveParzenEstimator(),
bad_estimator=IndependentAdaptiveParzenEstimator()))
self._old = theano.gof.link.raise_with_op.print_thunk_trace
theano.gof.link.raise_with_op.print_thunk_trace = True
def tearDown(self):
theano.gof.link.raise_with_op.print_thunk_trace = self._old
def test_optimize_20(self):
def callback(node, thunk, storage_map, compute_map):
numeric_outputs = [storage_map[v][0]
for v in node.outputs
if isinstance(v.type, theano.tensor.TensorType)]
numeric_inputs = [storage_map[v][0]
for v in node.inputs
if isinstance(v.type, theano.tensor.TensorType)]
if not all([numpy.all(numpy.isfinite(n)) for n in numeric_outputs]):
theano.printing.debugprint(node, depth=8)
print 'inputs'
print numeric_inputs
print 'outputs'
print numeric_outputs
raise ValueError('non-finite created in', node)
mode = theano.Mode(
optimizer='fast_compile',
linker=theano.gof.vm.VM_Linker(callback=callback))
self.experiment.bandit_algo.build_helpers(mode=mode)
_helper = self.experiment.bandit_algo._helper
theano.printing.debugprint(_helper)
for i in range(50):
print 'ITER', i
try:
self.experiment.run(1)
except:
raise
if 0:
import matplotlib.pyplot as plt
plt.subplot(1,2,1)
plt.plot(self.experiment.losses())
plt.subplot(1,2,2)
plt.scatter(
[t['x'] for t in self.experiment.trials],
range(len(self.experiment.trials)))
plt.show()
def test_4var_some_irrelevant(self):
return # XXX enable when compilation is faster
bandit_algo = GPAlgo(TestGaussian4D.Bandit(1, 0, 0, 1))
serial_exp = SerialExperiment(bandit_algo)
bandit_algo.n_startup_jobs = 10
serial_exp.run(50)
l0 = bandit_algo.kernels[0].lenscale()
l1 = bandit_algo.kernels[1].lenscale()
l2 = bandit_algo.kernels[2].lenscale()
l3 = bandit_algo.kernels[3].lenscale()
l4 = bandit_algo.kernels[4].lenscale()
for k in bandit_algo.kernels:
print 'last kernel fit', k, k.lenscale()
assert min(serial_exp.losses()) < .05
assert max(l1, l4) * 3 < min(l2, l3)
def test_4var_all_relevant(self):
bandit_algo = GPAlgo(TestGaussian4D.Bandit(1, .5, 2, 1))
serial_exp = SerialExperiment(bandit_algo)
bandit_algo.n_startup_jobs = 10
serial_exp.run(50)
l0 = bandit_algo.kernels[0].lenscale()
l1 = bandit_algo.kernels[1].lenscale()
l2 = bandit_algo.kernels[2].lenscale()
l3 = bandit_algo.kernels[3].lenscale()
l4 = bandit_algo.kernels[4].lenscale()
for k in bandit_algo.kernels:
print 'last kernel fit', k, k.lenscale()
assert min(serial_exp.losses()) < .05
gauss_scales = numpy.asarray([l1, l2, l3, l4])
assert gauss_scales.min() * 3 > gauss_scales.max()
def test_fit_uniform(self):
bandit_algo = GPAlgo(self.bandit)
bandit_algo.n_startup_jobs = 5
serial_exp = SerialExperiment(bandit_algo)
k = bandit_algo.kernels[0]
assert bandit_algo.is_refinable[k]
assert bandit_algo.bounds[k] == (self.xlim_low, self.xlim_high)
serial_exp.run(bandit_algo.n_startup_jobs)
serial_exp.run(20)
# a grid spacing would have used 25 points to cover 5 units of
# distance
# so be no more than 1/5**2 == .04. Here we test that the GP gets the
# error below .005
assert min(serial_exp.losses()) < 5e-3, serial_exp.results
# assert that the sampler has not exceeded the boundaries
assert min([t['x'] for t in serial_exp.trials]) >= self.xlim_low
assert min([t['x'] for t in serial_exp.trials]) <= self.xlim_high
def test_fit_lognormal(self):
bandit_algo = GPAlgo(self.ln_bandit)
bandit_algo.n_startup_jobs = 5
serial_exp = SerialExperiment(bandit_algo)
serial_exp.run(bandit_algo.n_startup_jobs)
# check that the Lognormal kernel has been
# identified as refinable
k = bandit_algo.kernels[0]
assert bandit_algo.is_refinable[k]
assert bandit_algo.bounds[k][0] > 0
serial_exp.run(25)
assert min(serial_exp.losses()) < .005
# check that all points were positive
assert min([t['x'] for t in serial_exp.trials]) > 0
# the lenscale is about 1.8 Is that about right? What's right?
print bandit_algo.kernels[0].lenscale()
class TestPickle(unittest.TestCase):
def setUp(self):
numpy.random.seed(555)
self.algo_a = GPAlgo(GaussWave2())
self.algo_a.n_startup_jobs = 10
self.algo_a.EI_ambition = 0.75
self.algo_b = GPAlgo(GaussWave2())
self.algo_b.n_startup_jobs = 10
self.algo_b.EI_ambition = 0.75
self.exp_a = SerialExperiment(self.algo_a)
self.exp_b = SerialExperiment(self.algo_b)
def test_reproducible(self):
self.exp_a.run(21)
self.exp_b.run(21)
for i, (ta, tb) in enumerate(zip(
self.exp_a.trials,
self.exp_b.trials)):
print i, ta, tb
print self.exp_a.losses()
print self.exp_b.losses()
# N.B. exact comparison, not approximate
assert numpy.all(self.exp_a.losses() == self.exp_b.losses())
def test_reproducible_w_recompiling(self):
for i in range(21):
self.exp_b.run(1)
if not i % 5:
todel = [k for k, v in self.algo_b.__dict__.items()
if isinstance(v, theano.compile.Function)]
for name in todel:
delattr(self.algo_b, name)
self.exp_a.run(21)
for i, (ta, tb) in enumerate(zip(
self.exp_a.trials,
self.exp_b.trials)):
print i, ta, tb
print self.exp_a.losses()
print self.exp_b.losses()
# N.B. exact comparison, not approximate
assert numpy.all(self.exp_a.losses() == self.exp_b.losses())
def test_reproducible_w_pickling(self):
self.exp_a.bandit_algo.trace_on = True
self.exp_b.bandit_algo.trace_on = True
ITERS = 12
for i in range(ITERS):
print 'running experiment b', i
self.exp_b.run(1)
if not i % 5:
# This knocks out the theano functions
# (see test_reproducible_w_recompiling)
# but also deep-copies the rest of the experiment
####print 'pickling'
pstr = cPickle.dumps(self.exp_b)
####print 'unpickling'
self.exp_b = cPickle.loads(pstr)
self.exp_a.run(ITERS)
trace_a = self.exp_a.bandit_algo._trace
trace_b = self.exp_b.bandit_algo._trace
for ta, tb in zip(trace_a, trace_b):
assert ta[0] == tb[0], (ta[0], tb[0])
print 'matching', ta[0]
na = numpy.asarray(ta[1])
nb = numpy.asarray(tb[1])
if not numpy.all(na == nb):
print ta[0]
print ''
print na.shape
print na
print ''
print nb.shape
print nb
print ''
print (na - nb)
assert 0
for i, (ta, tb) in enumerate(zip(
self.exp_a.trials,
self.exp_b.trials)):
###print 'trial', i
###print ' exp a', ta
###print ' exp b', tb
pass
print self.exp_a.losses()
print self.exp_b.losses()
assert numpy.allclose(self.exp_a.losses(), self.exp_b.losses())
class TestGM_Distractor(unittest.TestCase): # Tests normal
def setUp(self):
self.experiment = SerialExperiment(
bandit_algo=GM_BanditAlgo(
bandit=hyperopt.bandits.Distractor(),
good_estimator=IndependentAdaptiveParzenEstimator(),
bad_estimator=IndependentAdaptiveParzenEstimator()))
def test_op_counts(self):
# If everything is done right, there should be
# 2 adaptive parzen estimators in the algorithm
# - one for fitting the good examples
# - one for fitting the rest of the examples
# 1 GMM1 Op for drawing from the fit of good examples
def gmms(fn):
return [ap for ap in fn.maker.env.toposort()
if isinstance(ap.op, montetheano.distributions.GMM1)]
def adaptive_parzens(fn):
return [ap for ap in fn.maker.env.toposort()
if isinstance(ap.op, idxs_vals_rnd.AdaptiveParzen)]
self.experiment.bandit_algo.build_helpers(do_compile=True)
HL = self.experiment.bandit_algo.helper_locals
if 1:
f = theano.function(
[HL['n_to_draw'], HL['n_to_keep'], HL['y_thresh'], HL['yvals']]
+ HL['s_obs'].flatten(),
HL['G_ll'],
allow_input_downcast=True,
)
# theano.printing.debugprint(f)
assert len(gmms(f)) == 1
assert len(adaptive_parzens(f)) == 1
if 1:
f = theano.function(
[HL['n_to_draw'], HL['n_to_keep'], HL['y_thresh'], HL['yvals']]
+ HL['s_obs'].flatten(),
HL['G_ll'] - HL['B_ll'],
allow_input_downcast=True,
)
#print gmms(f)
#print adaptive_parzens(f)
assert len(gmms(f)) == 1
assert len(adaptive_parzens(f)) == 2
self.experiment.bandit_algo.build_helpers(do_compile=True)
_helper = self.experiment.bandit_algo._helper
assert len(gmms(_helper)) == 1
assert len(adaptive_parzens(_helper)) == 2
def test_optimize_20(self):
self.experiment.run(50)
import matplotlib.pyplot as plt
plt.subplot(1,2,1)
plt.plot(self.experiment.losses())
plt.subplot(1,2,2)
plt.hist(
[t['x'] for t in self.experiment.trials],
bins=20)
print self.experiment.losses()
print 'MIN', min(self.experiment.losses())
assert min(self.experiment.losses()) < -1.85
if 0:
plt.show()
class TestGaussWave2(unittest.TestCase): # Tests nested search
def setUp(self):
self.experiment = SerialExperiment(
bandit_algo=GM_BanditAlgo(
bandit=hyperopt.bandits.GaussWave2(),
good_estimator=IndependentAdaptiveParzenEstimator(),
bad_estimator=IndependentAdaptiveParzenEstimator()))
def test_op_counts_in_llik(self):
self.experiment.bandit_algo.build_helpers(do_compile=True, mode='FAST_RUN')
HL = self.experiment.bandit_algo.helper_locals
f = theano.function(
[HL['n_to_draw'], HL['n_to_keep'], HL['y_thresh'], HL['yvals']]
+ HL['s_obs'].flatten(),
HL['log_EI'],
no_default_updates=True,
mode='FAST_RUN') # required for shape inference
try:
assert len(gmms(f)) == 0
assert len(bgmms(f)) == 2 # sampling from good
assert len(categoricals(f)) == 1 # sampling from good
assert len(adaptive_parzens(f)) == 4 # fitting both good and bad
except:
theano.printing.debugprint(f)
raise
def test_op_counts_in_Gsamples(self):
self.experiment.bandit_algo.build_helpers(do_compile=True, mode='FAST_RUN')
HL = self.experiment.bandit_algo.helper_locals
f = theano.function(
[HL['n_to_draw'], HL['n_to_keep'], HL['y_thresh'], HL['yvals']]
+ HL['s_obs'].flatten(),
HL['Gsamples'].flatten(),
no_default_updates=True, # allow prune priors
mode='FAST_RUN') # required for shape inference
try:
assert len(gmms(f)) == 0
assert len(bgmms(f)) == 2 # sampling from good
assert len(categoricals(f)) == 1 # sampling from good
assert len(adaptive_parzens(f)) == 2 # fitting both good and bad
except:
theano.printing.debugprint(f)
raise
def test_optimize_20(self):
self.experiment.run(50)
import matplotlib.pyplot as plt
plt.subplot(1,2,1)
plt.plot(self.experiment.losses())
plt.subplot(1,2,2)
plt.scatter(
[t['x'] for t in self.experiment.trials],
range(len(self.experiment.trials)))
print self.experiment.losses()
print 'MIN', min(self.experiment.losses())
assert min(self.experiment.losses()) < -1.75
if 0:
plt.show()
def test_fit(self):
self.experiment.run(150)
plt.plot(
range(len(self.experiment.losses())),
self.experiment.losses())
plt.figure()
hyperopt.plotting.main_plot_vars(self.experiment,
end_with_show=True)
class TestGM_TwoArms(unittest.TestCase): # Tests one_of
def setUp(self):
self.experiment = SerialExperiment(
bandit_algo=GM_BanditAlgo(
bandit=hyperopt.bandits.TwoArms(),
good_estimator=IndependentAdaptiveParzenEstimator(),
bad_estimator=IndependentAdaptiveParzenEstimator()))
def test_optimize_20(self):
self.experiment.bandit_algo.build_helpers()
HL = self.experiment.bandit_algo.helper_locals
assert len(HL['Gsamples']) == 1
Gpseudocounts = HL['Gsamples'][0].vals.owner.inputs[1]
Bpseudocounts = HL['Bsamples'][0].vals.owner.inputs[1]
f = self.experiment.bandit_algo._helper
debug = theano.function(
[HL['n_to_draw'], HL['n_to_keep'], HL['y_thresh'], HL['yvals']]
+ HL['s_obs'].flatten(),
(HL['Gobs'].flatten()
+ [Gpseudocounts]
+ [Bpseudocounts]
+ [HL['yvals'][where(HL['yvals'] < HL['y_thresh'])]]
+ [HL['yvals'][where(HL['yvals'] >= HL['y_thresh'])]]
),
allow_input_downcast=True,
)
debug_rval = [None]
def _helper(*args):
rval = f(*args)
debug_rval[0] = debug(*args)
return rval
self.experiment.bandit_algo._helper = _helper
self.experiment.run(200)
gobs_idxs, gobs_vals, Gpseudo, Bpseudo, Gyvals, Byvals = debug_rval[0]
print gobs_idxs
print 'Gpseudo', Gpseudo
print 'Bpseudo', Bpseudo
import matplotlib.pyplot as plt
plt.subplot(1,4,1)
Xs = [t['x'] for t in self.experiment.trials]
Ys = self.experiment.losses()
plt.plot(Ys)
plt.xlabel('time')
plt.ylabel('loss')
plt.subplot(1,4,2)
plt.scatter(Xs,Ys )
plt.xlabel('X')
plt.ylabel('loss')
plt.subplot(1,4,3)
plt.hist(Xs )
plt.xlabel('X')
plt.ylabel('freq')
plt.subplot(1,4,4)
plt.hist(Gyvals, bins=20)
plt.hist(Byvals, bins=20)
print self.experiment.losses()
print 'MIN', min(self.experiment.losses())
assert min(self.experiment.losses()) < -3.00
if 0:
plt.show()