def test_incorporate_session():
rng = gu.gen_rng(4)
state = State(X,
cctypes=['normal'] * 5,
Zv={
0: 0,
1: 0,
2: 1,
3: 1,
4: 2
},
rng=rng)
# Incorporate row into a singleton cluster for all views.
previous = [len(state.views[v].Nk()) for v in [0, 1, 2]]
data = {i: rng.normal() for i in xrange(5)}
clusters = {
state.views[0].outputs[0]: previous[0],
state.views[1].outputs[0]: previous[1],
state.views[2].outputs[0]: previous[2],
}
state.incorporate(state.n_rows(), gu.merged(data, clusters))
assert [len(state.views[v].Nk()) for v in [0,1,2]] == \
[p+1 for p in previous]
# Incorporate row without specifying clusters, and some missing values
data = {i: rng.normal() for i in xrange(2)}
state.incorporate(state.n_rows(), data)
state.transition(N=3)
# Remove the incorporated rowid.
state.unincorporate(state.n_rows() - 1)
state.transition(N=3)
def test_categorical_forest_manual_inputs_errors():
state = State(
T, cctypes=CCTYPES, distargs=DISTARGS, rng=gu.gen_rng(1))
state.transition(N=1, progress=False)
cat_id = CCTYPES.index('categorical')
# Put 1201 into the first view.
view_idx = min(state.views)
state.incorporate_dim(
T[:,CCTYPES.index('categorical')], outputs=[1201],
cctype='categorical', distargs=DISTARGS[cat_id], v=view_idx)
# Updating cctype with completely invalid input should raise.
with pytest.raises(Exception):
distargs = DISTARGS[cat_id].copy()
distargs['inputs'] = [10000]
state.update_cctype(1201, 'random_forest', distargs=distargs)
# Updating cctype with input dimensions outside the view should raise.
cols_in_view = state.views[view_idx].dims.keys()
cols_out_view = [c for c in state.outputs if c not in cols_in_view]
assert len(cols_in_view) > 0 and len(cols_out_view) > 0
with pytest.raises(Exception):
distargs = DISTARGS[cat_id].copy()
distargs['inputs'] = cols_out_view
state.update_cctype(1201, 'random_forest', distargs=distargs)
# Updating cctype with no input dimensions should raise.
with pytest.raises(Exception):
distargs = DISTARGS[cat_id].copy()
distargs['inputs'] = []
state.update_cctype(1201, 'random_forest', distargs=distargs)
def test_cmi_different_views__ci_():
rng = gen_rng(0)
T = np.zeros((50,3))
T[:,0] = rng.normal(loc=-5, scale=1, size=50)
T[:,1] = rng.normal(loc=2, scale=2, size=50)
T[:,2] = rng.normal(loc=12, scale=3, size=50)
state = State(
T,
outputs=[0, 1, 2],
cctypes=['normal','normal','normal'],
Zv={0:0, 1:1, 2:2},
rng=rng
)
state.transition(N=30,
kernels=['alpha','view_alphas','column_params','column_hypers','rows'])
mi01 = state.mutual_information([0], [1])
mi02 = state.mutual_information([0], [2])
mi12 = state.mutual_information([1], [2])
# Marginal MI all zero.
assert np.allclose(mi01, 0)
assert np.allclose(mi02, 0)
assert np.allclose(mi12, 0)
# CMI on variable in other view equal to MI.
assert np.allclose(state.mutual_information([0], [1], {2:10}), mi01)
assert np.allclose(state.mutual_information([0], [2], {1:0}), mi02)
assert np.allclose(state.mutual_information([1], [2], {0:-2}), mi12)
assert np.allclose(state.mutual_information([1], [2], {0:None}, T=5), mi12)
def test_poisson_categorical():
state = State(
T, cctypes=CCTYPES, distargs=DISTARGS, rng=gu.gen_rng(0))
state.transition(N=1, progress=False)
state.update_cctype(CCTYPES.index('categorical'), 'poisson')
state.transition(N=1, progress=False)
state.update_cctype(CCTYPES.index('categorical'), 'categorical',
distargs={'k':2})
def test_naive_bayes_independence():
rng = gu.gen_rng(1)
D = rng.normal(size=(10, 1))
T = np.repeat(D, 10, axis=1)
Ci = list(itertools.combinations(range(10), 2))
state = State(T, cctypes=['normal'] * 10, Ci=Ci, rng=rng)
state.transition(N=10, progress=0)
vu.validate_crp_constrained_partition(state.Zv(), [], Ci, {}, {})
def test_complex_independent_relationships():
rng = gu.gen_rng(1)
D = rng.normal(size=(10, 1))
T = np.repeat(D, 10, axis=1)
Ci = [(2, 8), (0, 3)]
state = State(T, cctypes=['normal'] * 10, Ci=Ci, rng=rng)
state.transition(N=10, progress=0)
vu.validate_crp_constrained_partition(state.Zv(), [], Ci, {}, {})
def test_geometric_exponential():
state = State(
T, cctypes=CCTYPES, distargs=DISTARGS, rng=gu.gen_rng(0))
state.transition(N=1, progress=False)
state.update_cctype(CCTYPES.index('geometric'), 'exponential')
state.transition(N=1, progress=False)
# Incompatible numeric conversion.
with pytest.raises(Exception):
state.update_cctype(CCTYPES.index('exponential'), 'geometric')
def test_vonmises_normal():
state = State(
T, cctypes=CCTYPES, distargs=DISTARGS, rng=gu.gen_rng(0))
state.transition(N=1, progress=False)
state.update_cctype(CCTYPES.index('vonmises'), 'normal')
state.transition(N=1, progress=False)
state.update_cctype(CCTYPES.index('vonmises'), 'vonmises')
# Incompatible numeric conversion.
with pytest.raises(Exception):
state.update_cctype(CCTYPES.index('normal'), 'vonmises')
def generate_gaussian_samples():
state = State(D,
cctypes=['normal', 'normal'],
Zv={
0: 0,
1: 0
},
rng=gu.gen_rng(0))
view = state.view_for(1)
state.transition(S=15, kernels=['rows', 'column_params', 'column_hypers'])
samples = view.simulate(-1, [0, 1, view.outputs[0]], N=100)
return [replace_key(s, view.outputs[0], -1) for s in samples]
def init_view_state(data, iters, cctypes):
if isinstance(data, list):
data = np.array(data)
D = len(data[0])
outputs = range(D)
X = {c: data[:, i].tolist() for i, c in enumerate(outputs)}
view = View(X, cctypes=cctypes, outputs=[1000] + outputs, rng=RNG)
state = State(data[:, 0:D], outputs=outputs, cctypes=cctypes, rng=RNG)
if iters > 0:
view.transition(iters)
state.transition(iters)
return view, state
def state():
rng = gu.gen_rng(5)
rows = 120
cctypes = ['normal', 'bernoulli', 'normal']
G = generate_quadrants(rows, rng)
B, Zv, Zrv = tu.gen_data_table(rows, [1], [[.5, .5]], ['bernoulli'],
[None], [.95],
rng=rng)
T = np.column_stack((G, B.T))[:, [0, 2, 1]]
state = State(T, outputs=[0, 1, 2], cctypes=cctypes, rng=rng)
state.transition(N=20)
return state
def test_simple_dependence_constraint(Ci):
rng = gu.gen_rng(1)
D = rng.normal(size=(10, 1))
T = np.repeat(D, 10, axis=1)
Cd = [(2, 0), (8, 3)]
state = State(T, cctypes=['normal'] * 10, Ci=Ci, Cd=Cd, rng=rng)
with pytest.raises(ValueError):
# Cannot transition columns with dependencies.
state.transition(N=10, kernels=['columns'], progress=0)
state.transition(
N=10,
kernels=['rows', 'alpha', 'column_hypers', 'alpha', 'view_alphas'],
progress=False)
vu.validate_crp_constrained_partition(state.Zv(), Cd, Ci, {}, {})
def generate_regression_samples():
state = State(D,
cctypes=['normal', 'normal'],
Zv={
0: 0,
1: 0
},
rng=gu.gen_rng(4))
view = state.view_for(1)
assert not state._composite
state.update_cctype(1, 'linear_regression')
assert state._composite
state.transition(S=30, kernels=['rows', 'column_params', 'column_hypers'])
samples = view.simulate(-1, [0, 1, view.outputs[0]], N=100)
return [replace_key(s, view.outputs[0], -1) for s in samples]
def test_linreg_missing_data_ignore():
dataset = [
[1, 3, 1],
[2, 4, 1.5],
[float('nan'), 5, 1]]
state = State(dataset, cctypes=['normal']*3, Zv={0:0, 1:0, 2:0},
rng=gu.gen_rng(1))
# Make sure that missing covariates are handles as missing cell.
state.update_cctype(2, 'linear_regression', distargs={'inputs': [0,1]})
assert state.dim_for(2).inputs[1:] == [0,1]
state.transition(N=5, kernels=['rows', 'column_hypers', 'view_alphas'])
state.update_cctype(2, 'normal', distargs={'inputs': [0,1]})
# Make sure that specified inputs are set correctly.
state.update_cctype(2, 'linear_regression', distargs={'inputs': [1]})
assert state.dim_for(2).inputs[1:] == [1]
def state():
cctypes, distargs = cu.parse_distargs(
['categorical(k=5)', 'normal', 'poisson', 'bernoulli'])
T, Zv, Zc = tu.gen_data_table(50, [1], [[.33, .33, .34]],
cctypes,
distargs, [.95] * len(cctypes),
rng=gu.gen_rng(0))
s = State(T.T,
cctypes=cctypes,
distargs=distargs,
Zv={i: 0
for i in xrange(len(cctypes))},
rng=gu.gen_rng(0))
s.update_cctype(0, 'random_forest', distargs={'k': 5})
# XXX Uncomment me for a bug!
# state.update_cctype(1, 'linear_regression')
kernels = [
'rows', 'view_alphas', 'alpha', 'column_params', 'column_hypers'
]
s.transition(N=1, kernels=kernels)
return s
def test_categorical_forest():
state = State(
T, cctypes=CCTYPES, distargs=DISTARGS, rng=gu.gen_rng(1))
state.transition(N=1, progress=False)
cat_id = CCTYPES.index('categorical')
# If cat_id is singleton migrate first.
if len(state.view_for(cat_id).dims) == 1:
distargs = DISTARGS[cat_id].copy()
state.unincorporate_dim(cat_id)
state.incorporate_dim(
T[:,cat_id], outputs=[cat_id], cctype='categorical',
distargs=distargs, v=0)
state.update_cctype(cat_id, 'random_forest', distargs=distargs)
bernoulli_id = CCTYPES.index('bernoulli')
state.incorporate_dim(
T[:,bernoulli_id], outputs=[191], cctype='bernoulli',
v=state.Zv(cat_id))
state.update_cctype(191, 'random_forest', distargs={'k':2})
# Run valid transitions.
state.transition(
N=2, kernels=['rows','column_params','column_hypers'],
views=[state.Zv(cat_id)], progress=False)
# Running column transition should raise.
with pytest.raises(ValueError):
state.transition(N=1, kernels=['columns'], progress=False)
# Updating cctype in singleton View should raise.
distargs = DISTARGS[cat_id].copy()
state.incorporate_dim(
T[:,CCTYPES.index('categorical')], outputs=[98],
cctype='categorical', distargs=distargs, v=max(state.views)+1)
with pytest.raises(Exception):
state.update_cctype(98, 'random_forest', distargs=distargs)
def test_incorporate_state():
state = State(
T[:,:2], cctypes=CCTYPES[:2], distargs=DISTARGS[:2], rng=gu.gen_rng(0))
state.transition(N=5)
target = state.views.keys()[0]
# Incorporate a new dim into view[0].
state.incorporate_dim(
T[:,2], outputs=[2], cctype=CCTYPES[2], distargs=DISTARGS[2], v=target)
assert state.Zv(2) == target
state.transition(N=1)
# Incorporate a new dim into view[0] with a non-contiguous output.
state.incorporate_dim(
T[:,2], outputs=[10], cctype=CCTYPES[2], distargs=DISTARGS[2], v=target)
assert state.Zv(10) == target
state.transition(N=1)
# Some crash testing queries.
state.logpdf(-1, {10:1}, constraints={0:2, 1:1})
state.simulate(-1, [10], constraints={0:2})
# Incorporating with a duplicated output should raise.
with pytest.raises(ValueError):
state.incorporate_dim(
T[:,2], outputs=[10], cctype=CCTYPES[2], distargs=DISTARGS[2],
v=target)
# Multivariate incorporate should raise.
with pytest.raises(ValueError):
state.incorporate_dim(
T[:,2], outputs=[10, 2], cctype=CCTYPES[2],
distargs=DISTARGS[2], v=target)
# Missing output should raise.
with pytest.raises(ValueError):
state.incorporate_dim(
T[:,2], outputs=[], cctype=CCTYPES[2],
distargs=DISTARGS[2], v=target)
# Wrong number of rows should raise.
with pytest.raises(ValueError):
state.incorporate_dim(
T[:,2][:-1], outputs=[11], cctype=CCTYPES[2],
distargs=DISTARGS[2], v=target)
# Inputs should raise.
with pytest.raises(ValueError):
state.incorporate_dim(
T[:,2], outputs=[11], inputs=[2], cctype=CCTYPES[2],
distargs=DISTARGS[2], v=target)
# Incorporate dim into a newly created singleton view.
target = max(state.views)+1
state.incorporate_dim(
T[:,3], outputs=[3], cctype=CCTYPES[3],
distargs=DISTARGS[3], v=target)
assert state.Zv(3) == target
state.transition(N=1)
# Incorporate dim without specifying a view.
state.incorporate_dim(T[:,4], outputs=[4],
cctype=CCTYPES[4], distargs=DISTARGS[4])
state.transition(N=1)
# Unincorporate first dim.
previous = state.n_cols()
state.unincorporate_dim(0)
assert state.n_cols() == previous-1
state.transition(N=1)
# Reincorporate dim without specifying a view.
state.incorporate_dim(
T[:,0], outputs=[0], cctype=CCTYPES[0], distargs=DISTARGS[0])
state.transition(N=1)
# Incorporate dim into singleton view, remove it, assert destroyed.
target = max(state.views)+1
state.incorporate_dim(
T[:,5], outputs=[5], cctype=CCTYPES[5], distargs=DISTARGS[5],
v=target)
previous = len(state.views)
state.unincorporate_dim(5)
assert len(state.views) == previous-1
state.transition(N=1)
# Reincorporate dim into a singleton view.
target = max(state.views)+1
state.incorporate_dim(T[:,5], outputs=[5], cctype=CCTYPES[5],
distargs=DISTARGS[5], v=target)
state.transition(N=1)
# Incorporate the rest of the dims in the default way.
for i in xrange(6, len(CCTYPES)):
state.incorporate_dim(
T[:,i], outputs=[max(state.outputs)+1],
cctype=CCTYPES[i], distargs=DISTARGS[i])
state.transition(N=1)
# Unincorporating non-existent dim should raise.
with pytest.raises(ValueError):
state.unincorporate_dim(9999)
# Unincorporate all the dims, except the last one.
for o in state.outputs[:-1]:
state.unincorporate_dim(o)
assert state.n_cols() == 1
state.transition(N=1)
# Unincorporating last dim should raise.
with pytest.raises(ValueError):
state.unincorporate_dim(state.outputs[0])
from cgpm.utils import general as gu
from cgpm.utils import test as tu
# Set up the data generation
cctypes, distargs = cu.parse_distargs([
'normal', 'poisson', 'bernoulli', 'categorical(k=4)', 'lognormal',
'exponential', 'beta', 'geometric', 'vonmises'
])
T, Zv, Zc = tu.gen_data_table(200, [1], [[.25, .25, .5]],
cctypes,
distargs, [.95] * len(cctypes),
rng=gu.gen_rng(10))
state = State(T.T, cctypes=cctypes, distargs=distargs, rng=gu.gen_rng(312))
state.transition(N=10, progress=1)
def test_crash_simulate_joint(state):
state.simulate(-1, [0, 1, 2, 3, 4, 5, 6, 7, 8], N=10)
def test_crash_logpdf_joint(state):
state.logpdf(-1, {
0: 1,
1: 2,
2: 1,
3: 3,
4: 1,
5: 10,
6: .4,
def test_serialize_composite_cgpm():
rng = gu.gen_rng(2)
# Generate the data.
cctypes, distargs = cu.parse_distargs([
'categorical(k=3)', # RandomForest 0
'normal', # LinearRegression 1
'categorical(k=3)', # GPMCC 2
'poisson', # GPMCC 3
'normal', # GPMCC 4
'lognormal' # GPMCC 5
])
T, Zv, Zc = tu.gen_data_table(
35, [.4, .6], [[.33, .33, .34], [.5, .5]],
cctypes, distargs, [.2]*len(cctypes), rng=rng)
D = np.transpose(T)
# Create GPMCC.
state = State(
D[:,2:], outputs=[2,3,4,5], cctypes=cctypes[2:],
distargs=distargs[2:], rng=rng)
# Create a Forest.
forest = RandomForest(
outputs=[0],
inputs=[1,2,3,4],
distargs={
'inputs': {
'stattypes': [cctypes[i] for i in [1,2,3,4]],
'statargs': [distargs[i] for i in [1,2,3,4]]},
'k': distargs[0]['k']},
rng=rng)
# Create a Regression.
linreg = LinearRegression(
outputs=[1],
inputs=[3,4,5],
distargs={
'inputs': {
'stattypes': [cctypes[i] for i in [3,4,5]],
'statargs': [distargs[i] for i in [3,4,5]]}},
rng=rng)
# Incorporate the data.
def incorporate_data(cgpm, rowid, row):
cgpm.incorporate(
rowid,
{i: row[i] for i in cgpm.outputs},
{i: row[i] for i in cgpm.inputs},
)
for rowid, row in enumerate(D):
incorporate_data(forest, rowid, row)
incorporate_data(linreg, rowid, row)
# Compose the CGPMs.
# Run state transitions.
state.transition(N=10, progress=False)
# Compose CGPMs, instructing State to run the transitions.
token_forest = state.compose_cgpm(forest)
token_linreg = state.compose_cgpm(linreg)
state.transition_foreign(N=10, cols=[forest.outputs[0], linreg.outputs[0]])
# Now run the serialization.
metadata = state.to_metadata()
state2 = State.from_metadata(metadata)
# Check that the tokens are in state2.
assert token_forest in state2.hooked_cgpms
assert token_linreg in state2.hooked_cgpms
# The hooked cgpms must be unique objects after serialize/deserialize.
assert state.hooked_cgpms[token_forest] != state2.hooked_cgpms[token_forest]
assert state.hooked_cgpms[token_linreg] != state2.hooked_cgpms[token_linreg]
# Check that the log scores of the hooked cgpms agree.
assert np.allclose(
state.hooked_cgpms[token_forest].logpdf_score(),
state2.hooked_cgpms[token_forest].logpdf_score())
assert np.allclose(
state.hooked_cgpms[token_linreg].logpdf_score(),
state2.hooked_cgpms[token_linreg].logpdf_score())
# Now run some tests for the engine.
e = Engine(
D[:,2:], outputs=[2,3,4,5], cctypes=cctypes[2:],
distargs=distargs[2:], num_states=2, rng=rng)
e.compose_cgpm([forest, forest], multiprocess=1)
e.compose_cgpm([linreg, linreg], multiprocess=1)
e.transition_foreign(N=1, cols=[forest.outputs[0], linreg.outputs[0]])
e.dependence_probability(0,1)
e.simulate(-1, [0,1], {2:1}, multiprocess=0)
e.logpdf(-1, {1:1}, {2:1, 0:0}, multiprocess=0)
state3 = e.get_state(0)
# There is no guarantee that the logpdf score improves with inference, but
# it should reduce by more than a few nats.
def check_logpdf_delta(before, after):
return before < after or (after-before) < 5
check_logpdf_delta(
before=state.hooked_cgpms[token_forest].logpdf_score(),
after=state3.hooked_cgpms[token_forest].logpdf_score())
check_logpdf_delta(
before=state.hooked_cgpms[token_linreg].logpdf_score(),
after=state3.hooked_cgpms[token_linreg].logpdf_score())