def _gen_partition(weights, n):
if isinstance(weights, int):
weights = [1./weights]*weights
elif isinstance(weights, (list, np.ndarray,)):
if abs(sum(weights) - 1.) > 10E12:
raise ValueError('weight should sum to 1.')
else:
msg = "{} is not valid type for weights".format(type(weights))
raise ValueError(msg)
k = len(weights)
if n == 1:
return [0], weights
# there should be at least one instance of each of the components in
# weights
z = list(range(k))
if n-k == 1:
z += [pflip(weights, n=1)]
elif n-k > 1:
z += pflip(weights, n=n-k).tolist()
assert min(z) == 0
assert max(z) == k-1
return z, weights
def _sample_multi_col(model, col_idxs, given=None, n=1):
n_cols = len(col_idxs)
assert n_cols > 1
view_idxs = [model['col_assignment'][col_idx] for col_idx in col_idxs]
col2pos = dict((col_idx, i) for i, col_idx in enumerate(col_idxs))
view2col = dict()
for col_idx, view_idx in zip(col_idxs, view_idxs):
view2col[view_idx] = view2col.get(view_idx, []) + [col_idx]
samples = np.zeros((n, n_cols,))
for i in range(n):
for view, cols in view2col.items():
if given is None:
weights = _get_view_weights(model, cols[0])
else:
weights = _get_given_view_weights(model, cols[0], given)
component_idx = pflip(weights)
for col_idx in cols:
f = DRAWFUNC[model['dtypes'][col_idx]]
x = f(model, col_idx, component_idx)
samples[i, col2pos[col_idx]] = x
# import pdb; pdb.set_trace()
if n > 1:
return samples
else:
return samples[0, :]
def _sample_multi_col(model, col_idxs, given=None, n=1):
n_cols = len(col_idxs)
assert n_cols > 1
view_idxs = [model['col_assignment'][col_idx] for col_idx in col_idxs]
col2pos = dict((col_idx, i) for i, col_idx in enumerate(col_idxs))
view2col = dict()
for col_idx, view_idx in zip(col_idxs, view_idxs):
view2col[view_idx] = view2col.get(view_idx, []) + [col_idx]
samples = np.zeros((
n,
n_cols,
))
for i in range(n):
for view, cols in view2col.items():
if given is None:
weights = _get_view_weights(model, cols[0])
else:
weights = _get_given_view_weights(model, cols[0], given)
component_idx = pflip(weights)
for col_idx in cols:
f = DRAWFUNC[model['dtypes'][col_idx]]
x = f(model, col_idx, component_idx)
samples[i, col2pos[col_idx]] = x
# import pdb; pdb.set_trace()
if n > 1:
return samples
else:
return samples[0, :]
def _sample_from_bivariate_discrete(p, n):
n_rows, n_cols = p.shape
samples = pflip(p.flatten(), n=n)
x = np.zeros((n, 2))
for s, sample in enumerate(samples):
i = int(sample / n_cols)
j = int(sample % n_cols)
x[s, 0] = i
x[s, 1] = j
return x
def _sample_single_col(model, col_idx, given=None, n=1):
""" Samples data from the column at col_idx """
if given is None:
weights = _get_view_weights(model, col_idx)
else:
weights = _get_given_view_weights(model, col_idx, given)
component_idx = pflip(weights)
f = DRAWFUNC[model['dtypes'][col_idx]]
x = f(model, col_idx, component_idx)
return x
def _sample_single_col(model, col_idx, given=None, n=1):
""" Samples data from the column at col_idx """
if given is None:
weights = _get_view_weights(model, col_idx)
else:
weights = _get_given_view_weights(model, col_idx, given)
component_idx = pflip(weights)
f = DRAWFUNC[model['dtypes'][col_idx]]
x = f(model, col_idx, component_idx)
return x
def gen_mixture_data(n, mprop=.1):
""" Generate 2-feature mixture data """
x = np.zeros((n, 2,))
weights = [0.3, 0.7]
mu = [-1.0, 3.0]
for i in range(n):
k = pflip(weights)
m = mu[k]
x[i, :] = np.random.normal(m, size=2)
df = pd.DataFrame(x)
df.columns = ['x_1', 'x_2']
return df
def _sample_from_bivariate_discrete(p, n):
n_rows, n_cols = p.shape
samples = pflip(p.flatten(), n=n)
x = np.zeros((
n,
2,
))
for s, sample in enumerate(samples):
i = int(sample / n_cols)
j = int(sample % n_cols)
x[s, 0] = i
x[s, 1] = j
return x
def gen_mixture_data(n, mprop=.1):
""" Generate 2-feature mixture data """
x = np.zeros((
n,
2,
))
weights = [0.3, 0.7]
mu = [-1.0, 3.0]
for i in range(n):
k = pflip(weights)
m = mu[k]
x[i, :] = np.random.normal(m, size=2)
df = pd.DataFrame(x)
df.columns = ['x_1', 'x_2']
return df
def _categorical_draw(params):
return pflip(params['p'])
assert min(z) == 0
assert max(z) == k-1
return z, weights
PARAM_FUNCS = {
'continuous': _continuous_params,
'categorical': _categorical_params}
LOGPDFS = {
'categorical': lambda x, alpha: np.log([alpha[xi] for xi in x]),
'continuous': norm.logpdf}
DRAW = {
'categorical': lambda alpha: pflip(alpha),
'continuous': norm.rvs}
# ---
class DataGenerator(object):
""" Generate and store data and its generating distribution
Attributes
----------
dtypes : list(str)
The datatype of each column, either 'continuous' or 'categorical'.
df : pandas.DataFrame
The generated data
params : list(list(dict))
The distribution parameters the generated the data in each column