def plot_state(self, state_idx, hl_rows=(), hl_cols=()):
""" Visualize a cross-categorization state.
.. note :: Work in progress
This function is currently only suited to small data tables. There
are problems with the row labels overlapping. There are problems
with singleton views and categories having negative size or
appearing as lines. Lots to fix.
"""
if self._subsampled:
raise NotImplementedError("Not implemented for sub-sampled states")
if hl_rows != ():
if not isinstance(hl_rows, (list, np.ndarray,)):
hl_rows = [hl_rows]
hl_rows = [self._converters['row2idx'][row] for row in hl_rows]
if hl_cols != ():
if not isinstance(hl_cols, (list, np.ndarray,)):
hl_cols = [hl_cols]
hl_cols = [self._converters['col2idx'][col] for col in hl_cols]
model = self._models[state_idx]
init_kwargs = {'dtypes': self._dtypes,
'distargs': self._distargs,
'Zv': model['col_assignment'],
'Zrcv': model['row_assignments'],
'col_hypers': model['col_hypers'],
'state_alpha': model['state_alpha'],
'view_alphas': model['view_alphas']}
state = BCState(self._data.T, **init_kwargs)
model_logps = state.get_logps()
pu.plot_cc_model(self._data, model, model_logps, self._df.index,
self._df.columns, hl_rows=hl_rows, hl_cols=hl_cols)
def _initialize(args):
data = args[0]
kwargs = args[1]
t_start = time.time()
state = BCState(data.T, **kwargs) # transpose data to col-major
metadata = state.get_metadata()
diagnostics = {
'log_score': state.log_score(),
'iters': 0,
'time': time.time() - t_start}
return metadata, [diagnostics]
def _initialize(args):
data = args[0]
kwargs = args[1]
t_start = time.time()
state = BCState(data.T, **kwargs) # transpose data to col-major
metadata = state.get_metadata()
diagnostics = {
'log_score': state.log_score(),
'iters': 0,
'time': time.time() - t_start
}
return metadata, [diagnostics]
def blank():
n_rows = 10
n_cols = 4
# column major
X = np.random.rand(n_cols, n_rows)
bcstate = BCState(X)
return bcstate
def plot_state(self, state_idx, hl_rows=(), hl_cols=()):
""" Visualize a cross-categorization state.
.. note :: Work in progress
This function is currently only suited to small data tables. There
are problems with the row labels overlapping. There are problems
with singleton views and categories having negative size or
appearing as lines. Lots to fix.
"""
if self._subsampled:
raise NotImplementedError("Not implemented for sub-sampled states")
if hl_rows != ():
if not isinstance(hl_rows, (
list,
np.ndarray,
)):
hl_rows = [hl_rows]
hl_rows = [self._converters['row2idx'][row] for row in hl_rows]
if hl_cols != ():
if not isinstance(hl_cols, (
list,
np.ndarray,
)):
hl_cols = [hl_cols]
hl_cols = [self._converters['col2idx'][col] for col in hl_cols]
model = self._models[state_idx]
init_kwargs = {
'dtypes': self._dtypes,
'distargs': self._distargs,
'Zv': model['col_assignment'],
'Zrcv': model['row_assignments'],
'col_hypers': model['col_hypers'],
'state_alpha': model['state_alpha'],
'view_alphas': model['view_alphas']
}
state = BCState(self._data.T, **init_kwargs)
model_logps = state.get_logps()
pu.plot_cc_model(self._data,
model,
model_logps,
self._df.index,
self._df.columns,
hl_rows=hl_rows,
hl_cols=hl_cols)
def _run(args):
data = args[0]
checkpoint = args[1]
t_id = args[2]
verbose = args[3]
init_kwargs = args[4]
trans_kwargs = args[5]
# create copy of trans_kwargs so we don't mutate
trans_kwargs = dict(trans_kwargs)
n_iter = trans_kwargs['N']
if checkpoint is None:
checkpoint = n_iter
n_sweeps = 1
else:
trans_kwargs['N'] = checkpoint
n_sweeps = int(n_iter/checkpoint)
diagnostics = []
state = BCState(data.T, **init_kwargs) # transpose dat to col-major
for i in range(n_sweeps):
t_start = time.time()
state.transition(**trans_kwargs)
t_iter = time.time() - t_start
n_views = state.n_views
log_score = state.log_score()
diagnostic = {
'log_score': log_score,
'n_views': n_views,
'iters': checkpoint,
'time': t_iter}
diagnostics.append(diagnostic)
if verbose:
msg = "Model {}:\n\t+ sweep {} of {} in {} sec."
msg += "\n\t+ log score: {}"
msg += "\n\t+ n_views: {}\n"
print(msg.format(t_id, i, n_sweeps, t_iter, log_score, n_views))
metadata = state.get_metadata()
return metadata, diagnostics
def _init_bcstates(self):
self._bcstates = []
for model in self._models:
sd = np.random.randint(2**31 - 1)
init_kwarg = {
'dtypes': self._dtypes,
'distargs': self._distargs,
'Zv': model['col_assignment'],
'Zrcv': model['row_assignments'],
'col_hypers': model['col_hypers'],
'state_alpha': model['state_alpha'],
'view_alphas': model['view_alphas'],
'seed': sd
}
self._bcstates.append(BCState(self._data.T, **init_kwarg))
def _run(args):
data = args[0]
checkpoint = args[1]
t_id = args[2]
verbose = args[3]
init_kwargs = args[4]
trans_kwargs = args[5]
# create copy of trans_kwargs so we don't mutate
trans_kwargs = dict(trans_kwargs)
n_iter = trans_kwargs['N']
if checkpoint is None:
checkpoint = n_iter
n_sweeps = 1
else:
trans_kwargs['N'] = checkpoint
n_sweeps = int(n_iter / checkpoint)
diagnostics = []
state = BCState(data.T, **init_kwargs) # transpose dat to col-major
for i in range(n_sweeps):
t_start = time.time()
state.transition(**trans_kwargs)
t_iter = time.time() - t_start
n_views = state.n_views
log_score = state.log_score()
diagnostic = {
'log_score': log_score,
'n_views': n_views,
'iters': checkpoint,
'time': t_iter
}
diagnostics.append(diagnostic)
if verbose:
msg = "Model {}:\n\t+ sweep {} of {} in {} sec."
msg += "\n\t+ log score: {}"
msg += "\n\t+ n_views: {}\n"
print(msg.format(t_id, i, n_sweeps, t_iter, log_score, n_views))
metadata = state.get_metadata()
return metadata, diagnostics
def test_transition(gdat):
X, _, _ = gdat
bcstate = BCState(X)
bcstate.transition()
def test_predictive_logp(gdat):
X, _, _ = gdat
bcstate = BCState(X)
res = bcstate.predictive_logp([[2, 2]], [1.2], [[1, 3]], [0.0])
res = bcstate.predictive_logp([[2, 2]], [1.2])
assert res is not None
def test_predictive_draw(gdat):
X, _, _ = gdat
bcstate = BCState(X)
res = bcstate.predictive_draw([[2, 2]], [[1, 3]], [4.1])
res = bcstate.predictive_draw([[2, 2]])
assert res is not None
def test_get_metadata(gdat):
X, _, _ = gdat
bcstate = BCState(X)
res = bcstate.get_metadata()
assert res is not None
def test_transition(gdat):
X, _, _ = gdat
bcstate = BCState(X)
bcstate.transition()
def test_initalizes_no_args(gdat):
X, _, _ = gdat
bcstate = BCState(X)
assert bcstate is not None
def test_initalizes_base(gdat):
X, datatypes_list, distargs = gdat
bcstate = BCState(X, datatypes_list, distargs)
assert bcstate is not None
def test_get_metadata(gdat):
X, _, _ = gdat
bcstate = BCState(X)
res = bcstate.get_metadata()
assert res is not None
def test_predictive_draw(gdat):
X, _, _ = gdat
bcstate = BCState(X)
res = bcstate.predictive_draw([[2, 2]], [[1, 3]], [4.1])
res = bcstate.predictive_draw([[2, 2]])
assert res is not None
def test_predictive_probability(gdat):
X, _, _ = gdat
bcstate = BCState(X)
res = bcstate.predictive_probability([[2, 2]], [1.2], [[1, 3]], [0.0])
res = bcstate.predictive_probability([[2, 2]], [1.2])
assert res is not None
def test_conditioned_row_resample(gdat):
X, _, _ = gdat
bcstate = BCState(X)
bcstate.conditioned_row_resample(2, lambda x: 0.0)
