def _test_convergence_bb_cxx(N,
D,
kernel,
preprocess_data_fn=None,
nonconj=False,
burnin_niters=10000,
skip=10,
ntries=50,
nsamples=1000,
kl_places=2):
r = rng()
cluster_hp = {'alpha': 2.0}
feature_hps = [{'alpha': 1.0, 'beta': 1.0}] * D
defn = model_definition(N, [bb] * D)
nonconj_defn = model_definition(N, [bbnc] * D)
Y, posterior = data_with_posterior(
defn, cluster_hp, feature_hps, preprocess_data_fn)
data = numpy_dataview(Y)
s = initialize(nonconj_defn if nonconj else defn,
data,
cluster_hp=cluster_hp,
feature_hps=feature_hps,
r=r)
bs = bind(s, data)
wrapped_kernel = lambda s: kernel(s, r)
_test_convergence(bs,
posterior,
wrapped_kernel,
burnin_niters,
skip,
ntries,
nsamples,
kl_places)
def _test_convergence_bb_cxx(N,
D,
kernel,
preprocess_data_fn=None,
nonconj=False,
burnin_niters=10000,
skip=10,
ntries=50,
nsamples=1000,
kl_places=2):
r = rng()
cluster_hp = {'alpha': 2.0}
feature_hps = [{'alpha': 1.0, 'beta': 1.0}] * D
defn = model_definition(N, [bb] * D)
nonconj_defn = model_definition(N, [bbnc] * D)
Y, posterior = data_with_posterior(defn, cluster_hp, feature_hps,
preprocess_data_fn)
data = numpy_dataview(Y)
s = initialize(nonconj_defn if nonconj else defn,
data,
cluster_hp=cluster_hp,
feature_hps=feature_hps,
r=r)
bs = bind(s, data)
wrapped_kernel = lambda s: kernel(s, r)
_test_convergence(bs, posterior, wrapped_kernel, burnin_niters, skip,
ntries, nsamples, kl_places)
def test_get_set_params():
defn = model_definition(1, [bb, bnb, gp, nich])
data = np.array([
(True, 3, 5, 10.),
],
dtype=[('', bool), ('', int), ('', int), ('', float)])
s = initialize(defn=defn, data=numpy_dataview(data), r=rng())
s.set_cluster_hp({'alpha': 3.0})
assert_dict_almost_equals(s.get_cluster_hp(), {'alpha': 3.0})
hyperparams = [
{
'alpha': 1.2,
'beta': 4.3
},
{
'alpha': 1.,
'beta': 1.,
'r': 1
},
{
'alpha': 1.,
'inv_beta': 1.
},
{
'mu': 30.,
'kappa': 1.,
'sigmasq': 1.,
'nu': 1.
},
]
for i, hp in enumerate(hyperparams):
s.set_feature_hp(i, hp)
assert_dict_almost_equals(s.get_feature_hp(i), hp)
def _make_one_feature_bb_mm(initialize_fn, dataview, Nk, K, alpha, beta, r):
# XXX: the rng parameter passed does not get threaded through the
# random *data* generation
# use the py_bb for sampling
py_bb = bb.py_desc()._model_module
shared = py_bb.Shared()
shared.load({'alpha': alpha, 'beta': beta})
def init_sampler():
samp = py_bb.Sampler()
samp.init(shared)
return samp
samplers = [init_sampler() for _ in xrange(K)]
def gen_cluster(samp):
data = [(samp.eval(shared),) for _ in xrange(Nk)]
return np.array(data, dtype=[('', bool)])
Y_clustered = tuple(map(gen_cluster, samplers))
Y, assignment = data_with_assignment(Y_clustered)
view = dataview(Y)
s = initialize_fn(model_definition(Y.shape[0], [bb]),
view,
cluster_hp={'alpha': 2.},
feature_hps=[{'alpha': alpha, 'beta': beta}],
r=r,
assignment=assignment)
return s, view
def test_runner_multiprocessing_convergence():
N, D = 4, 5
defn = model_definition(N, [bb] * D)
prng = rng()
Y, posterior = data_with_posterior(defn, r=prng)
view = numpy_dataview(Y)
latents = [model.initialize(defn, view, prng)
for _ in xrange(mp.cpu_count())]
runners = [runner.runner(defn, view, latent, ['assign'])
for latent in latents]
r = parallel.runner(runners)
r.run(r=prng, niters=1000) # burnin
idmap = {C: i for i, C in enumerate(permutation_iter(N))}
def sample_iter():
r.run(r=prng, niters=10)
for latent in r.get_latents():
yield idmap[tuple(permutation_canonical(latent.assignments()))]
ref = [None]
def sample_fn():
if ref[0] is None:
ref[0] = sample_iter()
try:
return next(ref[0])
except StopIteration:
ref[0] = None
return sample_fn()
assert_discrete_dist_approx(sample_fn, posterior, ntries=100, kl_places=2)
def test_posterior_predictive_statistic():
N, D = 10, 4 # D needs to be even
defn = model_definition(N, [bb] * D)
Y = toy_dataset(defn)
prng = rng()
view = numpy_dataview(Y)
latents = [model.initialize(defn, view, prng) for _ in xrange(10)]
q = ma.masked_array(
np.array([(False,) * D], dtype=[('', bool)] * D),
mask=[(False,) * (D / 2) + (True,) * (D / 2)])
statistic = query.posterior_predictive_statistic(q, latents, prng)
assert_equals(statistic.shape, (1,))
assert_equals(len(statistic.dtype), D)
statistic = query.posterior_predictive_statistic(
q, latents, prng, merge='mode')
assert_equals(statistic.shape, (1,))
assert_equals(len(statistic.dtype), D)
statistic = query.posterior_predictive_statistic(
q, latents, prng, merge=['mode', 'mode', 'avg', 'avg'])
assert_equals(statistic.shape, (1,))
assert_equals(len(statistic.dtype), D)
q = ma.masked_array(
np.array([(False,) * D] * 3, dtype=[('', bool)] * D),
mask=[(False,) * (D / 2) + (True,) * (D / 2)] * 3)
statistic = query.posterior_predictive_statistic(q, latents, prng)
assert_equals(statistic.shape, (3,))
assert_equals(len(statistic.dtype), D)
def test_slice_theta_mm():
N = 100
data = np.array(
[(np.random.random() < 0.8,) for _ in xrange(N)],
dtype=[('', bool)])
defn = model_definition(N, [bbnc])
r = rng()
prior = {'alpha': 1.0, 'beta': 9.0}
view = numpy_dataview(data)
s = initialize(
defn,
view,
cluster_hp={'alpha': 1., 'beta': 9.},
feature_hps=[prior],
r=r,
assignment=[0] * N)
heads = len([1 for y in data if y[0]])
tails = N - heads
alpha1 = prior['alpha'] + heads
beta1 = prior['beta'] + tails
bs = bind(s, view)
params = {0: {'p': 0.05}}
def sample_fn():
theta(bs, r, tparams=params)
return s.get_suffstats(0, 0)['p']
rv = beta(alpha1, beta1)
assert_1d_cont_dist_approx_sps(sample_fn, rv, nsamples=50000)
def _test_stress(initialize_fn, dataview, R):
N = 20
D = 2
data = np.random.random(size=(N, D)) < 0.8
Y = np.array([tuple(y) for y in data], dtype=[('', bool)] * D)
view = dataview(Y)
defn = model_definition(N, [bb] * D)
s = initialize_fn(defn, view, cluster_hp={'alpha': 2.0}, r=R)
CHANGE_GROUP = 1
CHANGE_VALUE = 2
nops = 100
while nops:
assert len(s.groups()) >= 1
choice = np.random.choice([CHANGE_GROUP, CHANGE_VALUE])
if choice == CHANGE_GROUP:
# remove any empty groups. otherwise, add a new group
egroups = s.empty_groups()
if len(egroups) > 1:
s.delete_group(egroups[0])
else:
s.create_group(R)
else:
eid = np.random.randint(N)
if s.assignments()[eid] == -1:
# add to random group
egid = np.random.choice(s.groups())
s.add_value(egid, eid, Y[eid], R)
else:
s.remove_value(eid, Y[eid], R)
s.dcheck_consistency()
nops -= 1
def test_crp_empirical():
N = 4
alpha = 2.5
defn = model_definition(N, [bb])
Y = np.array([(True, )] * N, dtype=[('', bool)])
view = numpy_dataview(Y)
r = rng()
def crp_score(assignment):
latent = initialize(defn,
view,
r=r,
cluster_hp={'alpha': alpha},
assignment=assignment)
return latent.score_assignment()
scores = np.array(list(map(crp_score, permutation_iter(N))))
dist = scores_to_probs(scores)
idmap = {C: i for i, C in enumerate(permutation_iter(N))}
def sample_fn():
sample = permutation_canonical(_sample_crp(N, alpha))
return idmap[tuple(sample)]
assert_discrete_dist_approx(sample_fn, dist, ntries=100)
def test_slice_theta_mm():
N = 100
data = np.array([(np.random.random() < 0.8, ) for _ in xrange(N)],
dtype=[('', bool)])
defn = model_definition(N, [bbnc])
r = rng()
prior = {'alpha': 1.0, 'beta': 9.0}
view = numpy_dataview(data)
s = initialize(defn,
view,
cluster_hp={
'alpha': 1.,
'beta': 9.
},
feature_hps=[prior],
r=r,
assignment=[0] * N)
heads = len([1 for y in data if y[0]])
tails = N - heads
alpha1 = prior['alpha'] + heads
beta1 = prior['beta'] + tails
bs = bind(s, view)
params = {0: {'p': 0.05}}
def sample_fn():
theta(bs, r, tparams=params)
return s.get_suffstats(0, 0)['p']
rv = beta(alpha1, beta1)
assert_1d_cont_dist_approx_sps(sample_fn, rv, nsamples=50000)
def _make_one_feature_bb_mm(initialize_fn, dataview, Nk, K, alpha, beta, r):
# XXX: the rng parameter passed does not get threaded through the
# random *data* generation
# use the py_bb for sampling
py_bb = bb.py_desc()._model_module
shared = py_bb.Shared()
shared.load({'alpha': alpha, 'beta': beta})
def init_sampler():
samp = py_bb.Sampler()
samp.init(shared)
return samp
samplers = [init_sampler() for _ in xrange(K)]
def gen_cluster(samp):
data = [(samp.eval(shared), ) for _ in xrange(Nk)]
return np.array(data, dtype=[('', bool)])
Y_clustered = tuple(map(gen_cluster, samplers))
Y, assignment = data_with_assignment(Y_clustered)
view = dataview(Y)
s = initialize_fn(model_definition(Y.shape[0], [bb]),
view,
cluster_hp={'alpha': 2.},
feature_hps=[{
'alpha': alpha,
'beta': beta
}],
r=r,
assignment=assignment)
return s, view
def _test_stress(initialize_fn, dataview, R):
N = 20
D = 2
data = np.random.random(size=(N, D)) < 0.8
Y = np.array([tuple(y) for y in data], dtype=[('', bool)] * D)
view = dataview(Y)
defn = model_definition(N, [bb] * D)
s = initialize_fn(defn, view, cluster_hp={'alpha': 2.0}, r=R)
CHANGE_GROUP = 1
CHANGE_VALUE = 2
nops = 100
while nops:
assert len(s.groups()) >= 1
choice = np.random.choice([CHANGE_GROUP, CHANGE_VALUE])
if choice == CHANGE_GROUP:
# remove any empty groups. otherwise, add a new group
egroups = s.empty_groups()
if len(egroups) > 1:
s.delete_group(egroups[0])
else:
s.create_group(R)
else:
eid = np.random.randint(N)
if s.assignments()[eid] == -1:
# add to random group
egid = np.random.choice(s.groups())
s.add_value(egid, eid, Y[eid], R)
else:
s.remove_value(eid, Y[eid], R)
s.dcheck_consistency()
nops -= 1
def test_operations():
N = 10
R = rng(12)
def mkrow():
return (np.random.choice([False,
True]), np.random.choice([False, True]),
np.random.random(), np.random.choice([False, True]))
dtype = [('', bool), ('', bool), ('', float), ('', bool)]
# non-masked data
data = [mkrow() for _ in xrange(N)]
data = np.array(data, dtype=dtype)
defn = model_definition(N, [bb, bb, nich, bb])
init_args = {
'defn':
defn,
'cluster_hp': {
'alpha': 2.0
},
'feature_hps': [
dist_bb.EXAMPLES[0]['shared'],
dist_bb.EXAMPLES[0]['shared'],
dist_nich.EXAMPLES[0]['shared'],
dist_bb.EXAMPLES[0]['shared'],
],
'r':
R,
}
cxx_s = cxx_initialize(data=cxx_numpy_dataview(data), **init_args)
# *_initialize() randomly assigns all entities to a group, so we'll have to
# unset this assignment for this test
unset(cxx_s, data, R)
ensure_k_groups(cxx_s, 3, R)
assert cxx_s.nentities() == N
cxx_s.dcheck_consistency()
assert cxx_s.ngroups() == 3 and set(cxx_s.empty_groups()) == set([0, 1, 2])
for i, yi in enumerate(data):
egid = i % 2
cxx_s.add_value(egid, i, yi, R)
cxx_s.dcheck_consistency()
for i, yi in it.islice(enumerate(data), 2):
cxx_s.remove_value(i, yi, R)
cxx_s.dcheck_consistency()
newrow = mkrow()
newdata = np.array([newrow], dtype=dtype)
cxx_score = cxx_s.score_value(newdata[0], R)
assert cxx_score is not None
cxx_s.dcheck_consistency()
def test_model_definition_pickle():
defn = model_definition(10, [bb, niw(3)])
bstr = pickle.dumps(defn)
defn1 = pickle.loads(bstr)
assert_equals(defn.n(), defn1.n())
assert_equals(len(defn.models()), len(defn1.models()))
for a, b in zip(defn.models(), defn1.models()):
assert_equals(a.name(), b.name())
def test_model_definition_copy():
defn = model_definition(10, [bb, niw(3)])
defn_shallow = copy.copy(defn)
defn_deep = copy.deepcopy(defn)
assert_is_not(defn, defn_shallow)
assert_is_not(defn, defn_deep)
assert_is_not(defn._models, defn_deep._models)
assert_equals(defn.n(), defn_shallow.n())
assert_equals(defn.n(), defn_deep.n())
def test_sample_sanity():
# just a sanity check
defn = model_definition(10, [bb, bnb, gp, nich, dd(5), niw(4)])
clusters, samplers = sample(defn)
assert_equals(len(clusters), len(samplers))
for cluster in clusters:
assert_true(len(cluster) > 0)
for v in cluster:
assert_equals(len(v), len(defn.models()))
def test_zmatrix():
N, D = 10, 4
defn = model_definition(N, [bb] * D)
Y = toy_dataset(defn)
prng = rng()
view = numpy_dataview(Y)
latents = [model.initialize(defn, view, prng) for _ in xrange(10)]
zmat = query.zmatrix(latents)
assert_equals(zmat.shape, (N, N))
def test_sample_post_pred():
N = 10
R = rng(5483932)
D = 4
def randombool():
return np.random.choice([False, True])
def mkrow():
return tuple(randombool() for _ in xrange(D))
dtype = [('', bool)] * D
data = [mkrow() for _ in xrange(N)]
data = np.array(data, dtype=dtype)
defn = model_definition(N, [bb] * D)
init_args = {
'defn': defn,
'cluster_hp': {
'alpha': 2.0
},
'feature_hps': [dist_bb.EXAMPLES[0]['shared']] * D,
'r': R,
}
cxx_s = cxx_initialize(data=cxx_numpy_dataview(data), **init_args)
G = 3
unset(cxx_s, data, R)
ensure_k_groups(cxx_s, 3, R)
for i, yi in enumerate(data):
egid = i % G
cxx_s.add_value(egid, i, yi, R)
# sample
y_new_data = mkrow()
y_new_mask = tuple(randombool() for _ in xrange(D))
y_new = ma.masked_array(np.array([y_new_data], dtype=dtype),
mask=[y_new_mask])[0]
n_samples = 1000
cxx_samples = np.hstack(
[cxx_s.sample_post_pred(y_new, R)[1] for _ in xrange(n_samples)])
idmap = {C: i for i, C in enumerate(it.product([False, True], repeat=D))}
def todist(samples):
dist = np.zeros(len(idmap))
for s in samples:
dist[idmap[tuple(s)]] += 1.0
dist /= dist.sum()
return dist
cxx_dist = todist(cxx_samples)
assert cxx_dist is not None
def _test_cluster_hp_inference(initialize_fn,
prior_fn,
grid_min,
grid_max,
grid_n,
dataview,
bind_fn,
init_inf_kernel_state_fn,
inf_kernel_fn,
map_actual_postprocess_fn,
prng,
burnin=1000,
nsamples=1000,
skip=10,
trials=100,
places=2):
print '_test_cluster_hp_inference: burnin', burnin, 'nsamples', nsamples, \
'skip', skip, 'trials', trials, 'places', places
N = 1000
D = 5
# create random binary data, doesn't really matter what the values are
Y = np.random.random(size=(N, D)) < 0.5
Y = np.array([tuple(y) for y in Y], dtype=[('', np.bool)] * D)
view = dataview(Y)
defn = model_definition(N, [bb] * D)
latent = initialize_fn(defn, view, r=prng)
model = bind_fn(latent, view)
def score_alpha(alpha):
prev_alpha = latent.get_cluster_hp()['alpha']
latent.set_cluster_hp({'alpha': alpha})
score = prior_fn(alpha) + latent.score_assignment()
latent.set_cluster_hp({'alpha': prev_alpha})
return score
def sample_fn():
for _ in xrange(skip - 1):
inf_kernel_fn(model, opaque, prng)
inf_kernel_fn(model, opaque, prng)
return latent.get_cluster_hp()['alpha']
alpha0 = np.random.uniform(grid_min, grid_max)
print 'start alpha:', alpha0
latent.set_cluster_hp({'alpha': alpha0})
opaque = init_inf_kernel_state_fn(latent)
for _ in xrange(burnin):
inf_kernel_fn(model, opaque, prng)
print 'finished burnin of', burnin, 'iterations'
print 'grid_min', grid_min, 'grid_max', grid_max
assert_1d_cont_dist_approx_emp(sample_fn, score_alpha, grid_min, grid_max,
grid_n, trials, nsamples, places)
def test_operations():
N = 10
R = rng(12)
def mkrow():
return (np.random.choice([False, True]),
np.random.choice([False, True]),
np.random.random(),
np.random.choice([False, True]))
dtype = [('', bool), ('', bool), ('', float), ('', bool)]
# non-masked data
data = [mkrow() for _ in xrange(N)]
data = np.array(data, dtype=dtype)
defn = model_definition(N, [bb, bb, nich, bb])
init_args = {
'defn': defn,
'cluster_hp': {'alpha': 2.0},
'feature_hps': [
dist_bb.EXAMPLES[0]['shared'],
dist_bb.EXAMPLES[0]['shared'],
dist_nich.EXAMPLES[0]['shared'],
dist_bb.EXAMPLES[0]['shared'],
],
'r': R,
}
cxx_s = cxx_initialize(data=cxx_numpy_dataview(data), **init_args)
# *_initialize() randomly assigns all entities to a group, so we'll have to
# unset this assignment for this test
unset(cxx_s, data, R)
ensure_k_groups(cxx_s, 3, R)
assert cxx_s.nentities() == N
cxx_s.dcheck_consistency()
assert cxx_s.ngroups() == 3 and set(cxx_s.empty_groups()) == set([0, 1, 2])
for i, yi in enumerate(data):
egid = i % 2
cxx_s.add_value(egid, i, yi, R)
cxx_s.dcheck_consistency()
for i, yi in it.islice(enumerate(data), 2):
cxx_s.remove_value(i, yi, R)
cxx_s.dcheck_consistency()
newrow = mkrow()
newdata = np.array([newrow], dtype=dtype)
cxx_score = cxx_s.score_value(newdata[0], R)
assert cxx_score is not None
cxx_s.dcheck_consistency()
def _test_sample_post_pred(initialize_fn, dataview, y_new, r):
defn = model_definition(N, [bb] * D)
data = [tuple(row) for row in (np.random.random(size=(N, D)) < 0.8)]
data = np.array(data, dtype=[('', bool)] * D)
s = initialize_fn(defn=defn,
data=dataview(data),
cluster_hp={'alpha': 2.},
feature_hps=[{
'alpha': 1.,
'beta': 1.
}] * D,
r=r)
n_samples = 10000
Y_samples = [s.sample_post_pred(None, r)[1] for _ in xrange(n_samples)]
Y_samples = np.hstack(Y_samples)
empty_groups = list(s.empty_groups())
if len(empty_groups):
for egid in empty_groups[1:]:
s.delete_group(egid)
else:
s.create_group(r)
assert len(s.empty_groups()) == 1
def score_post_pred(y):
# XXX: the C++ API can only handle structural arrays for now
y = np.array([y], dtype=[('', bool)] * D)[0]
_, scores = s.score_value(y, r)
return logsumexp(scores)
scores = np.array(
list(map(score_post_pred, it.product([False, True], repeat=D))))
scores = np.exp(scores)
assert_almost_equals(scores.sum(), 1.0, places=3)
# lazy man
idmap = {y: i for i, y in enumerate(it.product([False, True], repeat=D))}
smoothing = 1e-5
sample_hist = np.zeros(len(idmap), dtype=np.int)
for y in Y_samples:
sample_hist[idmap[tuple(y)]] += 1.
sample_hist = np.array(sample_hist, dtype=np.float) + smoothing
sample_hist /= sample_hist.sum()
#print 'actual', scores
#print 'emp', sample_hist
kldiv = KL_discrete(scores, sample_hist)
print 'KL:', kldiv
assert kldiv <= 0.005
def test_sample_post_pred():
N = 10
R = rng(5483932)
D = 4
def randombool():
return np.random.choice([False, True])
def mkrow():
return tuple(randombool() for _ in xrange(D))
dtype = [('', bool)] * D
data = [mkrow() for _ in xrange(N)]
data = np.array(data, dtype=dtype)
defn = model_definition(N, [bb] * D)
init_args = {
'defn': defn,
'cluster_hp': {'alpha': 2.0},
'feature_hps': [dist_bb.EXAMPLES[0]['shared']] * D,
'r': R,
}
cxx_s = cxx_initialize(data=cxx_numpy_dataview(data), **init_args)
G = 3
unset(cxx_s, data, R)
ensure_k_groups(cxx_s, 3, R)
for i, yi in enumerate(data):
egid = i % G
cxx_s.add_value(egid, i, yi, R)
# sample
y_new_data = mkrow()
y_new_mask = tuple(randombool() for _ in xrange(D))
y_new = ma.masked_array(
np.array([y_new_data], dtype=dtype),
mask=[y_new_mask])[0]
n_samples = 1000
cxx_samples = np.hstack(
[cxx_s.sample_post_pred(y_new, R)[1] for _ in xrange(n_samples)])
idmap = {C: i for i, C in enumerate(it.product([False, True], repeat=D))}
def todist(samples):
dist = np.zeros(len(idmap))
for s in samples:
dist[idmap[tuple(s)]] += 1.0
dist /= dist.sum()
return dist
cxx_dist = todist(cxx_samples)
assert cxx_dist is not None
def test_runner_multyvac():
defn = model_definition(10, [bb, nich, niw(3)])
Y = toy_dataset(defn)
view = numpy_dataview(Y)
kc = runner.default_kernel_config(defn)
prng = rng()
latents = [model.initialize(defn, view, prng)
for _ in xrange(2)]
runners = [runner.runner(defn, view, latent, kc) for latent in latents]
r = parallel.runner(runners, backend='multyvac', layer='perf', core='f2')
r.run(r=prng, niters=1000)
r.run(r=prng, niters=1000)
def _test_sample_post_pred(initialize_fn, dataview, y_new, r):
defn = model_definition(N, [bb] * D)
data = [tuple(row) for row in (np.random.random(size=(N, D)) < 0.8)]
data = np.array(data, dtype=[('', bool)] * D)
s = initialize_fn(
defn=defn,
data=dataview(data),
cluster_hp={'alpha': 2.},
feature_hps=[{'alpha': 1., 'beta': 1.}] * D,
r=r)
n_samples = 10000
Y_samples = [s.sample_post_pred(None, r)[1] for _ in xrange(n_samples)]
Y_samples = np.hstack(Y_samples)
empty_groups = list(s.empty_groups())
if len(empty_groups):
for egid in empty_groups[1:]:
s.delete_group(egid)
else:
s.create_group(r)
assert len(s.empty_groups()) == 1
def score_post_pred(y):
# XXX: the C++ API can only handle structural arrays for now
y = np.array([y], dtype=[('', bool)] * D)[0]
_, scores = s.score_value(y, r)
return logsumexp(scores)
scores = np.array(
list(map(score_post_pred, it.product([False, True], repeat=D))))
scores = np.exp(scores)
assert_almost_equals(scores.sum(), 1.0, places=3)
# lazy man
idmap = {y: i for i, y in enumerate(it.product([False, True], repeat=D))}
smoothing = 1e-5
sample_hist = np.zeros(len(idmap), dtype=np.int)
for y in Y_samples:
sample_hist[idmap[tuple(y)]] += 1.
sample_hist = np.array(sample_hist, dtype=np.float) + smoothing
sample_hist /= sample_hist.sum()
#print 'actual', scores
#print 'emp', sample_hist
kldiv = KL_discrete(scores, sample_hist)
print 'KL:', kldiv
assert kldiv <= 0.005
def _test_scalar_hp_inference(view,
prior_fn,
w,
grid_min,
grid_max,
grid_n,
likelihood_model,
scalar_hp_key,
burnin=1000,
nsamples=1000,
every=10,
trials=100,
places=2):
"""
view must be 1D
"""
r = rng()
hparams = {0: {scalar_hp_key: (prior_fn, w)}}
def score_fn(scalar):
d = latent.get_feature_hp(0)
prev_scalar = d[scalar_hp_key]
d[scalar_hp_key] = scalar
latent.set_feature_hp(0, d)
score = prior_fn(scalar) + latent.score_data(0, None, r)
d[scalar_hp_key] = prev_scalar
latent.set_feature_hp(0, d)
return score
defn = model_definition(len(view), [likelihood_model])
latent = initialize(defn, view, r=r)
model = bind(latent, view)
def sample_fn():
for _ in xrange(every):
slice_hp(model, r, hparams=hparams)
return latent.get_feature_hp(0)[scalar_hp_key]
for _ in xrange(burnin):
slice_hp(model, r, hparams=hparams)
print 'finished burnin of', burnin, 'iterations'
print 'grid_min', grid_min, 'grid_max', grid_max
assert_1d_cont_dist_approx_emp(sample_fn,
score_fn,
grid_min,
grid_max,
grid_n,
trials,
nsamples,
places)
def test_runner_multiprocessing():
defn = model_definition(10, [bb, nich, niw(3)])
Y = toy_dataset(defn)
view = numpy_dataview(Y)
kc = runner.default_kernel_config(defn)
prng = rng()
latents = [model.initialize(defn, view, prng)
for _ in xrange(mp.cpu_count())]
runners = [runner.runner(defn, view, latent, kc) for latent in latents]
r = parallel.runner(runners)
# check it is restartable
r.run(r=prng, niters=10)
r.run(r=prng, niters=10)
def run_dpgmm(niter=1000, datadir="../../", nfeatures=13):
ranking = [10, 6, 7, 26, 5, 8, 4, 19, 12, 23, 24, 33, 28, 25,
14, 3, 0, 1, 21, 30, 11, 31, 13, 9, 22, 2, 27, 29,
32, 17, 18, 20, 16, 15]
features, labels, lc, hr, tstart, \
features_lb, labels_lb, lc_lb, hr_lb, \
fscaled, fscaled_lb, fscaled_full, labels_all = \
load_data(datadir, tseg=1024.0, log_features=None,
ranking=ranking)
labels_phys = feature_engineering.convert_labels_to_physical(labels)
labels_phys_lb = feature_engineering.convert_labels_to_physical(labels_lb)
labels_all_phys = np.hstack([labels_phys["train"], labels_phys["val"],
labels_phys["test"]])
fscaled_small = fscaled_full[:, :13]
nchains = 8
# The random state object
prng = rng()
# Define a DP-GMM where the Gaussian is 2D
defn = model_definition(fscaled_small.shape[0],
[normal_inverse_wishart(fscaled_small.shape[1])])
fscaled_rec = np.array([(list(f),) for f in fscaled_small],
dtype=[('', np.float32, fscaled_small.shape[1])])
# Create a wrapper around the numpy recarray which
# data-microscopes understands
view = numpy_dataview(fscaled_rec)
# Initialize nchains start points randomly in the state space
latents = [model.initialize(defn, view, prng) for _ in xrange(nchains)]
# Create a runner for each chain
runners = [runner.runner(defn, view, latent,
kernel_config=['assign']) for latent in latents]
r = parallel.runner(runners)
r.run(r=prng, niters=niter)
with open(datadir+"grs1915_dpgmm.pkl", "w") as f:
pickle.dump(r, f)
return
def score_dataset(counts):
M, K = counts.shape
Y = np.array([(y, ) for y in counts], dtype=[('', np.int, (K, ))])
view = cxx_numpy_dataview(Y)
r = rng()
defn = model_definition(M, [dm(K)])
prior = {'alphas': [1.] * K}
s = cxx_initialize(defn,
view,
r,
feature_hps=[prior],
assignment=[0] * M)
assert_equals(s.groups(), [0])
return s.score_data(None, None, r)
def _test_crp(initialize_fn, dataview, alpha, r):
N = 6
defn = model_definition(N, [bb])
Y = np.array([(True,)] * N, dtype=[('', bool)])
view = dataview(Y)
def crp_score(assignment):
latent = initialize_fn(
defn, view, r=r,
cluster_hp={'alpha': alpha}, assignment=assignment)
return latent.score_assignment()
dist = np.array(list(map(crp_score, permutation_iter(N))))
dist = np.exp(dist)
assert_almost_equals(dist.sum(), 1.0, places=3)
def test_masked_operations():
N = 10
R = rng(2347785)
dtype = [('', bool), ('', int), ('', float)]
def randombool():
return np.random.choice([False, True])
def mkrow():
return (randombool(), np.random.randint(1, 10), np.random.random())
def mkmask():
return (randombool(), randombool(), randombool())
data = [mkrow() for _ in xrange(N)]
data = np.array(data, dtype=dtype)
mask = [mkmask() for _ in xrange(N)]
data = ma.masked_array(data, mask=mask)
defn = model_definition(N, [bb, bnb, nich])
init_args = {
'defn':
defn,
'cluster_hp': {
'alpha': 10.0
},
'feature_hps': [
dist_bb.EXAMPLES[0]['shared'],
dist_bnb.EXAMPLES[0]['shared'],
dist_nich.EXAMPLES[0]['shared'],
],
'r':
R,
}
cxx_s = cxx_initialize(data=cxx_numpy_dataview(data), **init_args)
# see comment above
unset(cxx_s, data, R)
ensure_k_groups(cxx_s, 3, R)
for i, yi in enumerate(data):
egid = i % 2
cxx_s.add_value(egid, i, yi, R)
cxx_s.dcheck_consistency()
for i, yi in enumerate(data):
cxx_s.remove_value(i, yi, R)
cxx_s.dcheck_consistency()
def score_dataset(counts):
M, K = counts.shape
Y = np.array([(y,) for y in counts], dtype=[('', np.int, (K,))])
view = cxx_numpy_dataview(Y)
r = rng()
defn = model_definition(M, [dm(K)])
prior = {'alphas': [1.] * K}
s = cxx_initialize(
defn,
view,
r,
feature_hps=[prior],
assignment=[0] * M)
assert_equals(s.groups(), [0])
return s.score_data(None, None, r)
def test_get_set_params():
defn = model_definition(1, [bb, bnb, gp, nich])
data = np.array([(True, 3, 5, 10.), ],
dtype=[('', bool), ('', int), ('', int), ('', float)])
s = initialize(defn=defn, data=numpy_dataview(data), r=rng())
s.set_cluster_hp({'alpha': 3.0})
assert_dict_almost_equals(s.get_cluster_hp(), {'alpha': 3.0})
hyperparams = [
{'alpha': 1.2, 'beta': 4.3},
{'alpha': 1., 'beta': 1., 'r': 1},
{'alpha': 1., 'inv_beta': 1.},
{'mu': 30., 'kappa': 1., 'sigmasq': 1., 'nu': 1.},
]
for i, hp in enumerate(hyperparams):
s.set_feature_hp(i, hp)
assert_dict_almost_equals(s.get_feature_hp(i), hp)
def _test_crp(initialize_fn, dataview, alpha, r):
N = 6
defn = model_definition(N, [bb])
Y = np.array([(True, )] * N, dtype=[('', bool)])
view = dataview(Y)
def crp_score(assignment):
latent = initialize_fn(defn,
view,
r=r,
cluster_hp={'alpha': alpha},
assignment=assignment)
return latent.score_assignment()
dist = np.array(list(map(crp_score, permutation_iter(N))))
dist = np.exp(dist)
assert_almost_equals(dist.sum(), 1.0, places=3)
def latent(groups, entities_per_group, features, r):
N = groups * entities_per_group
defn = model_definition(N, [bb] * features)
# generate fake data
Y = np.random.random(size=(N, features)) <= 0.5
view = numpy_dataview(
np.array([tuple(y) for y in Y], dtype=[('', bool)] * features))
# assign entities to their respective groups
assignment = [[g] * entities_per_group for g in xrange(groups)]
assignment = list(it.chain.from_iterable(assignment))
latent = bind(initialize(defn, view, r, assignment=assignment), view)
latent.create_group(r) # perftest() doesnt modify group assignments
return latent
def _test_scalar_hp_inference(view,
prior_fn,
w,
grid_min,
grid_max,
grid_n,
likelihood_model,
scalar_hp_key,
burnin=1000,
nsamples=1000,
every=10,
trials=100,
places=2):
"""
view must be 1D
"""
r = rng()
hparams = {0: {scalar_hp_key: (prior_fn, w)}}
def score_fn(scalar):
d = latent.get_feature_hp(0)
prev_scalar = d[scalar_hp_key]
d[scalar_hp_key] = scalar
latent.set_feature_hp(0, d)
score = prior_fn(scalar) + latent.score_data(0, None, r)
d[scalar_hp_key] = prev_scalar
latent.set_feature_hp(0, d)
return score
defn = model_definition(len(view), [likelihood_model])
latent = initialize(defn, view, r=r)
model = bind(latent, view)
def sample_fn():
for _ in xrange(every):
slice_hp(model, r, hparams=hparams)
return latent.get_feature_hp(0)[scalar_hp_key]
for _ in xrange(burnin):
slice_hp(model, r, hparams=hparams)
print 'finished burnin of', burnin, 'iterations'
print 'grid_min', grid_min, 'grid_max', grid_max
assert_1d_cont_dist_approx_emp(sample_fn, score_fn, grid_min, grid_max,
grid_n, trials, nsamples, places)
def test_runner_convergence():
N, D = 4, 5
defn = model_definition(N, [bb] * D)
prng = rng()
Y, posterior = data_with_posterior(defn, r=prng)
view = numpy_dataview(Y)
latent = model.initialize(defn, view, prng)
r = runner.runner(defn, view, latent, ['assign'])
r.run(r=prng, niters=1000) # burnin
idmap = {C: i for i, C in enumerate(permutation_iter(N))}
def sample_fn():
r.run(r=prng, niters=10)
new_latent = r.get_latent()
return idmap[tuple(permutation_canonical(new_latent.assignments()))]
assert_discrete_dist_approx(sample_fn, posterior, ntries=100)
def _test_serializer(initialize_fn, deserialize_fn, dataview):
N = 10
R = rng()
dtype = [('', bool), ('', int), ('', float)]
def randombool():
return np.random.choice([False, True])
def mkrow():
return (randombool(), np.random.randint(1, 10), np.random.random())
def mkmask():
return (randombool(), randombool(), randombool())
data = [mkrow() for _ in xrange(N)]
data = np.array(data, dtype=dtype)
defn = model_definition(N, [bb, bnb, nich])
init_args = {
'defn':
defn,
'data':
dataview(data),
'cluster_hp': {
'alpha': 10.0
},
'feature_hps': [
dist_bb.EXAMPLES[0]['shared'],
dist_bnb.EXAMPLES[0]['shared'],
dist_nich.EXAMPLES[0]['shared'],
],
'r':
R,
}
state = initialize_fn(**init_args)
raw = state.serialize()
state1 = deserialize_fn(defn, raw)
assert state1 is not None
bstr = pickle.dumps(state)
state2 = pickle.loads(bstr)
assert state2 is not None
def test_masked_operations():
N = 10
R = rng(2347785)
dtype = [('', bool), ('', int), ('', float)]
def randombool():
return np.random.choice([False, True])
def mkrow():
return (randombool(), np.random.randint(1, 10), np.random.random())
def mkmask():
return (randombool(), randombool(), randombool())
data = [mkrow() for _ in xrange(N)]
data = np.array(data, dtype=dtype)
mask = [mkmask() for _ in xrange(N)]
data = ma.masked_array(data, mask=mask)
defn = model_definition(N, [bb, bnb, nich])
init_args = {
'defn': defn,
'cluster_hp': {'alpha': 10.0},
'feature_hps': [
dist_bb.EXAMPLES[0]['shared'],
dist_bnb.EXAMPLES[0]['shared'],
dist_nich.EXAMPLES[0]['shared'],
],
'r': R,
}
cxx_s = cxx_initialize(data=cxx_numpy_dataview(data), **init_args)
# see comment above
unset(cxx_s, data, R)
ensure_k_groups(cxx_s, 3, R)
for i, yi in enumerate(data):
egid = i % 2
cxx_s.add_value(egid, i, yi, R)
cxx_s.dcheck_consistency()
for i, yi in enumerate(data):
cxx_s.remove_value(i, yi, R)
cxx_s.dcheck_consistency()
def test_posterior_predictive():
N, D = 10, 4 # D needs to be even
defn = model_definition(N, [bb] * D)
Y = toy_dataset(defn)
prng = rng()
view = numpy_dataview(Y)
latents = [model.initialize(defn, view, prng) for _ in xrange(10)]
q = ma.masked_array(
np.array([(False,) * D], dtype=[('', bool)] * D),
mask=[(False,) * (D / 2) + (True,) * (D / 2)])
samples = query.posterior_predictive(q, latents, prng)
assert_equals(samples.shape, (1, len(latents)))
q = ma.masked_array(
np.array([(False,) * D] * 3, dtype=[('', bool)] * D),
mask=[(False,) * (D / 2) + (True,) * (D / 2)] * 3)
samples = query.posterior_predictive(q, latents, prng)
assert_equals(samples.shape, (3, len(latents)))
def test_crp_empirical():
N = 4
alpha = 2.5
defn = model_definition(N, [bb])
Y = np.array([(True,)] * N, dtype=[('', bool)])
view = numpy_dataview(Y)
r = rng()
def crp_score(assignment):
latent = initialize(
defn, view, r=r,
cluster_hp={'alpha': alpha}, assignment=assignment)
return latent.score_assignment()
scores = np.array(list(map(crp_score, permutation_iter(N))))
dist = scores_to_probs(scores)
idmap = {C: i for i, C in enumerate(permutation_iter(N))}
def sample_fn():
sample = permutation_canonical(_sample_crp(N, alpha))
return idmap[tuple(sample)]
assert_discrete_dist_approx(sample_fn, dist, ntries=100)
def test_dm_cxx():
K = 4
Y = np.array([
([0, 1, 2, 5],),
([1, 0, 1, 2],),
([0, 2, 9, 9],),
], dtype=[('', np.int, (K,))])
Y_np = np.vstack(y[0] for y in Y)
cxx_view = cxx_numpy_dataview(Y)
r = rng()
defn = model_definition(Y.shape[0], [dm(K)])
prior = {'alphas': [1.] * K}
cxx_s = cxx_initialize(
defn,
cxx_view,
r,
feature_hps=[prior],
assignment=[0] * Y.shape[0])
counts = cxx_s.get_suffstats(0, 0)['counts']
assert_sequence_equal(counts, list(Y_np.sum(axis=0)))
def _test_serializer(initialize_fn, deserialize_fn, dataview):
N = 10
R = rng()
dtype = [('', bool), ('', int), ('', float)]
def randombool():
return np.random.choice([False, True])
def mkrow():
return (randombool(), np.random.randint(1, 10), np.random.random())
def mkmask():
return (randombool(), randombool(), randombool())
data = [mkrow() for _ in xrange(N)]
data = np.array(data, dtype=dtype)
defn = model_definition(N, [bb, bnb, nich])
init_args = {
'defn': defn,
'data': dataview(data),
'cluster_hp': {'alpha': 10.0},
'feature_hps': [
dist_bb.EXAMPLES[0]['shared'],
dist_bnb.EXAMPLES[0]['shared'],
dist_nich.EXAMPLES[0]['shared'],
],
'r': R,
}
state = initialize_fn(**init_args)
raw = state.serialize()
state1 = deserialize_fn(defn, raw)
assert state1 is not None
bstr = pickle.dumps(state)
state2 = pickle.loads(bstr)
assert state2 is not None
def test_dm_cxx():
K = 4
Y = np.array([
([0, 1, 2, 5], ),
([1, 0, 1, 2], ),
([0, 2, 9, 9], ),
],
dtype=[('', np.int, (K, ))])
Y_np = np.vstack(y[0] for y in Y)
cxx_view = cxx_numpy_dataview(Y)
r = rng()
defn = model_definition(Y.shape[0], [dm(K)])
prior = {'alphas': [1.] * K}
cxx_s = cxx_initialize(defn,
cxx_view,
r,
feature_hps=[prior],
assignment=[0] * Y.shape[0])
counts = cxx_s.get_suffstats(0, 0)['counts']
assert_sequence_equal(counts, list(Y_np.sum(axis=0)))
def _test_runner_kernel_config(kc_fn, models):
defn = model_definition(10, models)
Y = toy_dataset(defn)
view = numpy_dataview(Y)
kc = kc_fn(defn)
prng = rng()
ntries = 5
while ntries:
latent = model.initialize(defn, view, prng)
assignments = latent.assignments()
r = runner.runner(defn, view, latent, kc)
r.run(r=prng, niters=10)
assignments1 = r.get_latent().assignments()
# XXX: it should be very unlikely the assignments are all equal
if assignments == assignments1:
ntries -= 1
else:
return # success
assert_true(False) # exceeded ntries
def _test_multivariate_models(initialize_fn, dataview, bind, gibbs_assign, R):
# XXX: this test only checks that the operations don't crash
mu = np.ones(3)
kappa = 0.3
Q = random_orthonormal_matrix(3)
psi = np.dot(Q, np.dot(np.diag([1.0, 0.5, 0.2]), Q.T))
nu = 6
N = 10
def genrow():
return tuple([
np.random.choice([False, True]),
[np.random.uniform(-3.0, 3.0) for _ in xrange(3)]
])
X = np.array([genrow() for _ in xrange(N)],
dtype=[('', bool), ('', float, (3, ))])
view = dataview(X)
defn = model_definition(N, [bb, niw(3)])
s = initialize_fn(defn,
view,
cluster_hp={'alpha': 2.},
feature_hps=[{
'alpha': 2.,
'beta': 2.
}, {
'mu': mu,
'kappa': kappa,
'psi': psi,
'nu': nu
}],
r=R)
bound_s = bind(s, view)
for _ in xrange(10):
gibbs_assign(bound_s, R)
def _test_multivariate_models(initialize_fn,
dataview,
bind,
gibbs_assign,
R):
# XXX: this test only checks that the operations don't crash
mu = np.ones(3)
kappa = 0.3
Q = random_orthonormal_matrix(3)
psi = np.dot(Q, np.dot(np.diag([1.0, 0.5, 0.2]), Q.T))
nu = 6
N = 10
def genrow():
return tuple(
[np.random.choice([False, True]),
[np.random.uniform(-3.0, 3.0) for _ in xrange(3)]])
X = np.array([genrow()
for _ in xrange(N)], dtype=[('', bool), ('', float, (3,))])
view = dataview(X)
defn = model_definition(N, [bb, niw(3)])
s = initialize_fn(
defn,
view,
cluster_hp={'alpha': 2.},
feature_hps=[
{'alpha': 2., 'beta': 2.},
{'mu': mu, 'kappa': kappa, 'psi': psi, 'nu': nu}
],
r=R)
bound_s = bind(s, view)
for _ in xrange(10):
gibbs_assign(bound_s, R)
def test_mnist_supervised():
mnist_dataset = _get_mnist_dataset()
classes = range(10)
classmap = {c: i for i, c in enumerate(classes)}
train_data, test_data = [], []
for c in classes:
Y = mnist_dataset['data'][
np.where(mnist_dataset['target'] == float(c))[0]]
Y_train, Y_test = train_test_split(Y, test_size=0.01)
train_data.append(Y_train)
test_data.append(Y_test)
sample_size_max = 10000
def mk_class_data(c, Y):
n, D = Y.shape
print 'number of digit', c, 'in training is', n
dtype = [('', bool)] * D + [('', int)]
inds = np.random.permutation(Y.shape[0])[:sample_size_max]
Y = np.array([tuple(list(y) + [classmap[c]]) for y in Y[inds]],
dtype=dtype)
return Y
Y_train = np.hstack([mk_class_data(c, y)
for c, y in zip(classes, train_data)])
Y_train = Y_train[np.random.permutation(np.arange(Y_train.shape[0]))]
n, = Y_train.shape
D = len(Y_train.dtype)
print 'training data is', n, 'examples'
print 'image dimension is', (D - 1), 'pixels'
view = numpy_dataview(Y_train)
defn = model_definition(n, [bb] * (D - 1) + [dd(len(classes))])
r = rng()
s = initialize(defn,
view,
cluster_hp={'alpha': 0.2},
feature_hps=[{'alpha': 1., 'beta': 1.}] *
(D - 1) + [{'alphas': [1. for _ in classes]}],
r=r)
bound_s = bind(s, view)
indiv_prior_fn = log_exponential(1.2)
hparams = {
i: {
'alpha': (indiv_prior_fn, 1.5),
'beta': (indiv_prior_fn, 1.5),
} for i in xrange(D - 1)}
hparams[D - 1] = {
'alphas[{}]'.format(idx): (indiv_prior_fn, 1.5)
for idx in xrange(len(classes))
}
def print_prediction_results():
results = []
for c, Y_test in zip(classes, test_data):
for y in Y_test:
query = ma.masked_array(
np.array([tuple(y) + (0,)],
dtype=[('', bool)] * (D - 1) + [('', int)]),
mask=[(False,) * (D - 1) + (True,)])[0]
samples = [
s.sample_post_pred(query, r)[1][0][-1] for _ in xrange(30)]
samples = np.bincount(samples, minlength=len(classes))
prediction = np.argmax(samples)
results.append((classmap[c], prediction, samples))
print 'finished predictions for class', c
Y_actual = np.array([a for a, _, _ in results], dtype=np.int)
Y_pred = np.array([b for _, b, _ in results], dtype=np.int)
print 'accuracy:', accuracy_score(Y_actual, Y_pred)
print 'confusion matrix:'
print confusion_matrix(Y_actual, Y_pred)
# AUROC for one vs all (each class)
for i, clabel in enumerate(classes):
Y_true = np.copy(Y_actual)
# treat class c as the "positive" example
positive_examples = Y_actual == i
negative_examples = Y_actual != i
Y_true[positive_examples] = 1
Y_true[negative_examples] = 0
Y_prob = np.array([float(c[i]) / c.sum() for _, _, c in results])
cls_auc = roc_auc_score(Y_true, Y_prob)
print 'class', clabel, 'auc=', cls_auc
#import matplotlib.pylab as plt
#Y_prob = np.array([c for _, _, c in results])
#fpr, tpr, thresholds = roc_curve(Y_actual, Y_prob, pos_label=0)
#plt.plot(fpr, tpr)
#plt.show()
def kernel(rid):
start0 = time.time()
assign(bound_s, r)
sec0 = time.time() - start0
start1 = time.time()
hp(bound_s, r, hparams=hparams)
sec1 = time.time() - start1
print 'rid=', rid, 'nclusters=', s.ngroups(), \
'iter0=', sec0, 'sec', 'iter1=', sec1, 'sec'
sec_per_post_pred = sec0 / (float(view.size()) * (float(s.ngroups())))
print ' time_per_post_pred=', sec_per_post_pred, 'sec'
# print group size breakdown
sizes = [(gid, s.groupsize(gid)) for gid in s.groups()]
sizes = sorted(sizes, key=lambda x: x[1], reverse=True)
print ' group_sizes=', sizes
print_prediction_results()
# save state
mkdirp("mnist-states")
fname = os.path.join("mnist-states", "state-iter{}.ser".format(rid))
with open(fname, "w") as fp:
fp.write(s.serialize())
# training
iters = 30
for rid in xrange(iters):
kernel(rid)
def test_mnist_supervised(n):
mnist_dataset = _get_mnist_dataset()
classes = range(10)
classmap = {c: i for i, c in enumerate(classes)}
train_data, test_data = [], []
for c in classes:
Y = mnist_dataset['data'][np.where(
mnist_dataset['target'] == float(c))[0]]
Y_train, Y_test = train_test_split(Y, test_size=0.01)
train_data.append(Y_train)
test_data.append(Y_test)
sample_size_max = n
def mk_class_data(c, Y):
n, D = Y.shape
print 'number of digit', c, 'in training is', n
dtype = [('', bool)] * D + [('', int)]
inds = np.random.permutation(Y.shape[0])[:sample_size_max]
Y = np.array([tuple(list(y) + [classmap[c]]) for y in Y[inds]],
dtype=dtype)
return Y
Y_train = np.hstack(
[mk_class_data(c, y) for c, y in zip(classes, train_data)])
Y_train = Y_train[np.random.permutation(np.arange(Y_train.shape[0]))]
n, = Y_train.shape
D = len(Y_train.dtype)
print 'training data is', n, 'examples'
print 'image dimension is', (D - 1), 'pixels'
view = numpy_dataview(Y_train)
defn = model_definition(n, [bb] * (D - 1) + [dd(len(classes))])
r = rng()
s = initialize(defn,
view,
cluster_hp={'alpha': 0.2},
feature_hps=[{
'alpha': 1.,
'beta': 1.
}] * (D - 1) + [{
'alphas': [1. for _ in classes]
}],
r=r)
bound_s = bind(s, view)
indiv_prior_fn = log_exponential(1.2)
hparams = {
i: {
'alpha': (indiv_prior_fn, 1.5),
'beta': (indiv_prior_fn, 1.5),
}
for i in xrange(D - 1)
}
hparams[D - 1] = {
'alphas[{}]'.format(idx): (indiv_prior_fn, 1.5)
for idx in xrange(len(classes))
}
def print_prediction_results():
results = []
for c, Y_test in zip(classes, test_data):
for y in Y_test:
query = ma.masked_array(
np.array([tuple(y) + (0, )],
dtype=[('', bool)] * (D - 1) + [('', int)]),
mask=[(False, ) * (D - 1) + (True, )])[0]
samples = [
s.sample_post_pred(query, r)[1][0][-1] for _ in xrange(30)
]
samples = np.bincount(samples, minlength=len(classes))
prediction = np.argmax(samples)
results.append((classmap[c], prediction, samples))
print 'finished predictions for class', c
Y_actual = np.array([a for a, _, _ in results], dtype=np.int)
Y_pred = np.array([b for _, b, _ in results], dtype=np.int)
print 'accuracy:', accuracy_score(Y_actual, Y_pred)
print 'confusion matrix:'
print confusion_matrix(Y_actual, Y_pred)
# AUROC for one vs all (each class)
for i, clabel in enumerate(classes):
Y_true = np.copy(Y_actual)
# treat class c as the "positive" example
positive_examples = Y_actual == i
negative_examples = Y_actual != i
Y_true[positive_examples] = 1
Y_true[negative_examples] = 0
Y_prob = np.array([float(c[i]) / c.sum() for _, _, c in results])
cls_auc = roc_auc_score(Y_true, Y_prob)
print 'class', clabel, 'auc=', cls_auc
#import matplotlib.pylab as plt
#Y_prob = np.array([c for _, _, c in results])
#fpr, tpr, thresholds = roc_curve(Y_actual, Y_prob, pos_label=0)
#plt.plot(fpr, tpr)
#plt.show()
def kernel(rid):
start0 = time.time()
assign(bound_s, r)
sec0 = time.time() - start0
start1 = time.time()
hp(bound_s, r, hparams=hparams)
sec1 = time.time() - start1
print 'rid=', rid, 'nclusters=', s.ngroups(), \
'iter0=', sec0, 'sec', 'iter1=', sec1, 'sec'
sec_per_post_pred = sec0 / (float(view.size()) * (float(s.ngroups())))
print ' time_per_post_pred=', sec_per_post_pred, 'sec'
# training
iters = 30
for rid in xrange(iters):
kernel(rid)
# print group size breakdown
sizes = [(gid, s.groupsize(gid)) for gid in s.groups()]
sizes = sorted(sizes, key=lambda x: x[1], reverse=True)
print ' group_sizes=', sizes
#print_prediction_results()
# save state
mkdirp("mnist-states")
fname = os.path.join("mnist-states", "state-iter{}.ser".format(rid))
with open(fname, "w") as fp:
fp.write(s.serialize())
