def emissionProb(self, node, is_partial_graph_index=False):
# Access the emission matrix with the full graph indices
node_full = self.partialGraphIndexToFullGraphIndex(
node) if is_partial_graph_index == True else node
prob = self.L[node_full].reshape((-1, ))
if (self.inFeedbackSet(node_full, is_partial_graph_index=False)):
return fbsData(prob, 0)
return fbsData(prob, -1)
def emissionProb(self, node, is_partial_graph_index=False):
# Access the emission matrix with the full graph indices
node_full = self.partialGraphIndexToFullGraphIndex(
node) if is_partial_graph_index == True else node
y = self.ys[int(node_full)]
if (not np.any(np.isnan(y))):
prob = self.recognizerFunc(y).reshape((-1, ))
else:
prob = np.zeros_like(self.pi0).reshape((-1, ))
if (self.inFeedbackSet(node_full, is_partial_graph_index=False)):
return fbsData(prob, 0)
return fbsData(prob, -1)
def vData( self, U, V, node, edges=None ):
# THESE MUST NOT MODIFY U OR V!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
# VERY IMPORTANT!!!!!! PROBABLY ENFORCE THIS SOMEHOW IN THE FURURE
V_row, V_col, _ = V
if( self.inFeedbackSet( node, is_partial_graph_index=True ) ):
# This is going to be empty because by construction,
# if a node is in the fbs, all the information from
# going down the down edges can be gained by going
# up an up edge.
if( edges is None ):
return []
elif( isinstance( edges, Iterable ) ):
return []
else:
return fbsData( np.array( [] ), -1 )
if( ~np.any( np.in1d( V_row, node ) ) ):
# If the node isn't in the V object (node is a leaf)
ans = [ fbsData( np.array( [] ), -1 ) ]
elif( edges is None ):
# Return the data over all down edges
ans = self.vDataFromMask( V, np.in1d( V_row, node ) )
elif( isinstance( edges, Iterable ) and len( edges ) > 0 ):
# Return over only certain edges
mask = np.in1d( V_row, node )
for e in edges:
mask &= np.in1d( V_col, e )
ans = self.vDataFromMask( V, mask )
elif( isinstance( edges, Iterable ) and len( edges ) == 0 ):
# This happens when we're not passed a down edge. In this
# case, we should return an empty v value, not a leaf v value
return [ fbsData( np.array( [] ), -1 ) ]
else:
# Only looking at one edge
ans = self.vDataFromMask( V, np.in1d( V_col, edges ) )
if( len( ans ) == 0 ):
# If we're looking at a leaf, return all 0s
nVals = 1 if edges is None else len( edges )
ans = []
for _ in range( nVals ):
ans.append( fbsData( np.array( [] ), -1 ) )
for v in ans:
data = v.data
assert np.any( np.isnan( data ) ) == False, data
return ans
def integrate(cls, integrand, axes):
# Check if we need to use the regular integrate
if (not isinstance(integrand, fbsData)):
return GraphHMM.integrate(integrand, axes)
# Need adjusted axes because the relative axes in integrand change as we reduce
# over each axis
assert isinstance(axes, Iterable)
if (len(axes) == 0):
return integrand
integrand, fbs_axis = (integrand.data, integrand.fbs_axis)
assert max(axes) < integrand.ndim
axes = np.array(axes)
axes[axes < 0] = integrand.ndim + axes[axes < 0]
adjusted_axes = np.array(sorted(axes)) - np.arange(len(axes))
for ax in adjusted_axes:
integrand = logsumexp(integrand, axis=ax)
if (fbs_axis > -1):
fbs_axis -= len(adjusted_axes)
# assert fbs_axis > -1, adjusted_axes
return fbsData(integrand, fbs_axis)
def genFilterProbs(self):
# Initialize U and V
U = []
for node in self.partial_graph.nodes:
U.append(fbsData(np.zeros(self.K), -1))
V_row = self.partial_graph.pmask.row
V_col = self.partial_graph.pmask.col
V_data = []
for node in self.partial_graph.pmask.row:
V_data.append(fbsData(np.zeros(self.K), -1))
# Invalidate all data elements
for node in self.partial_graph.nodes:
U[node].data[:] = np.nan
self.assignV((V_row, V_col, V_data), node, np.nan, keep_shape=True)
return U, (V_row, V_col, V_data)
def emissionProb(self, node, is_partial_graph_index=False):
# Access the emission matrix with the full graph indices
node_full = self.partialGraphIndexToFullGraphIndex(
node) if is_partial_graph_index == True else node
group = self.node_groups[int(node_full)]
y = self.ys[int(node_full)]
if (not np.any(np.isnan(y))):
prob = self.emission_dists[group][:, y].sum(axis=-1)
else:
prob = np.zeros_like(self.emission_dists[group][:, 0])
if (self.inFeedbackSet(node_full, is_partial_graph_index=False)):
fbs_index = self.fbsIndex(node_full,
is_partial_graph_index=False,
within_graph=True)
for _ in range(fbs_index):
prob = prob[None]
return fbsData(prob, 0)
return fbsData(prob, -1)
def initialProb(self, node, is_partial_graph_index=False):
pi = np.copy(self.pi0)
node_full = self.partialGraphIndexToFullGraphIndex(
node) if is_partial_graph_index == True else node
if (int(node_full) in self.possible_latent_states):
states = self.possible_latent_states[int(node_full)]
impossible_states = np.setdiff1d(np.arange(pi.shape[-1]), states)
for state in impossible_states:
pi[impossible_states] = np.NINF
pi[states] -= logsumexp(pi)
return fbsData(pi, -1)
def emissionProb(self, node, is_partial_graph_index=False):
# Access the emission matrix with the full graph indices
node_full = self.partialGraphIndexToFullGraphIndex(
node) if is_partial_graph_index == True else node
group = self.node_groups[int(node_full)]
y = self.ys[int(node_full)]
cond = self.conds[int(node_full)]
if (not np.any(np.isnan(y))):
prob = self.recognizerFuncs[group](y, cond).reshape((-1, ))
else:
prob = np.zeros_like(self.pi0s[group][:, 0]).reshape((-1, ))
if (self.inFeedbackSet(node_full, is_partial_graph_index=False)):
fbs_index = self.fbsIndex(node_full,
is_partial_graph_index=False,
within_graph=True)
for _ in range(fbs_index):
prob = prob[None]
return fbsData(prob, 0)
return fbsData(prob, -1)
def extendAxes( self, node, term, target_axis, max_dim ):
# Push the first axis out to target_axis, but don't change
# the axes past max_dim
term, fbs_axis = ( term.data, term.fbs_axis )
# Add axes before the fbsAxes
for _ in range( max_dim - target_axis - 1 ):
term = np.expand_dims( term, 1 )
if( fbs_axis != -1 ):
fbs_axis += 1
# Prepend axes
for _ in range( target_axis ):
term = np.expand_dims( term, 0 )
if( fbs_axis != -1 ):
fbs_axis += 1
return fbsData( term, fbs_axis )
def transitionProb(self, child, is_partial_graph_index=False):
parents, parent_order = self.getFullParents(
child,
get_order=True,
is_partial_graph_index=is_partial_graph_index,
return_partial_graph_index=True)
child_full = self.partialGraphIndexToFullGraphIndex(
child) if is_partial_graph_index == True else child
ndim = len(parents) + 1
group = self.node_groups[int(child_full)]
shape = []
for p, _ in sorted(zip(parents, parent_order), key=lambda po: po[1]):
full_p = int(self.partialGraphIndexToFullGraphIndex(p))
shape.append(self.getNodeDim(full_p))
shape.append(self.getNodeDim(int(child_full)))
shape = tuple(shape)
pi = np.copy(self.pis[group][shape])
# Reshape pi's axes to match parent order
assert len(parents) + 1 == pi.ndim
assert parent_order.shape[0] == parents.shape[0]
# Sort the parent dimensions by parent order
pi = np.moveaxis(pi, np.arange(ndim),
np.hstack((parent_order, ndim - 1)))
# If we know the latent state for child, then ensure that we
# transition there. This is intervention!
modified = False
for parent, order in zip(parents, parent_order):
parent_full = self.partialGraphIndexToFullGraphIndex(parent)
if (int(parent_full) in self.possible_latent_states):
parent_states = self.possible_latent_states[int(parent_full)]
impossible_parent_axes = np.setdiff1d(
np.arange(pi.shape[order]), parent_states)
index = [slice(0, s) for s in pi.shape]
index[order] = impossible_parent_axes
pi[tuple(index)] = np.NINF
modified = True
if (int(child_full) in self.possible_latent_states):
states = self.possible_latent_states[int(child_full)]
impossible_axes = np.setdiff1d(np.arange(pi.shape[-1]), states)
pi[..., impossible_axes] = np.NINF
modified = True
# In case entire rows summed to -inf
pi[np.isnan(pi)] = np.NINF
# Check if there are nodes in [ child, *parents ] that are in the fbs.
# If there are, then move their axes
fbsOffset = lambda x: self.fbsIndex(
x, is_partial_graph_index=True, within_graph=True) + 1
fbs_indices = [
fbsOffset(parent) for parent in parents
if self.inFeedbackSet(parent, is_partial_graph_index=True)
]
if (self.inFeedbackSet(child,
is_partial_graph_index=is_partial_graph_index)):
fbs_indices.append(
self.fbsIndex(child,
is_partial_graph_index=is_partial_graph_index,
within_graph=True) + 1)
if (len(fbs_indices) > 0):
expand_by = max(fbs_indices)
for _ in range(expand_by):
pi = pi[..., None]
# If there are parents in the fbs, move them to the appropriate axes
for i, parent in enumerate(parents):
if (self.inFeedbackSet(parent, is_partial_graph_index=True)):
pi = np.swapaxes(pi, i, fbsOffset(parent) + ndim - 1)
if (self.inFeedbackSet(
child, is_partial_graph_index=is_partial_graph_index)):
# If the child is in the fbs, then move it to the appropriate axis
pi = np.swapaxes(pi, ndim - 1, fbsOffset(child) + ndim - 1)
return fbsData(pi, ndim)
return fbsData(pi, -1)
def multiplyTerms(cls, terms):
# Basically np.einsum but in log space
assert isinstance(terms, Iterable)
# Check if we should use the multiply for fbsData or for regular data
fbs_data_count, non_fbs_data_count = (0, 0)
for t in terms:
if (isinstance(t, fbsData)):
fbs_data_count += 1
else:
non_fbs_data_count += 1
# Can't mix types
if (not (fbs_data_count == 0 or non_fbs_data_count == 0)):
print('fbs_data_count', fbs_data_count)
print('non_fbs_data_count', non_fbs_data_count)
print(terms)
for t in terms:
if (isinstance(t, fbsData)):
print('this ones good', t, type(t))
else:
print('this ones bad', t, type(t))
assert 0
# Use the regular multiply if we don't have fbs data
if (fbs_data_count == 0):
return GraphHMM.multiplyTerms(terms)
# Remove the empty terms
terms = [t for t in terms if np.prod(t.shape) > 1]
if (len(terms) == 0):
return fbsData(np.array([]), 0)
# Separate out where the feedback set axes start and get the largest fbs_axis.
# Need to handle case where ndim of term > all fbs axes
# terms, fbs_axes_start = list( zip( *terms ) )
fbs_axes_start = [term.fbs_axis for term in terms]
terms = [term.data for term in terms]
if (max(fbs_axes_start) != -1):
max_fbs_axis = max([
ax if ax != -1 else term.ndim
for ax, term in zip(fbs_axes_start, terms)
])
if (max_fbs_axis > 0):
# Pad extra dims at each term so that the fbs axes start the same way for every term
for i, ax in enumerate(fbs_axes_start):
if (ax == -1):
for _ in range(max_fbs_axis - terms[i].ndim + 1):
terms[i] = terms[i][..., None]
else:
for _ in range(max_fbs_axis - ax):
terms[i] = np.expand_dims(terms[i], axis=ax)
else:
max_fbs_axis = -1
ndim = max([len(term.shape) for term in terms])
axes = [[i for i, s in enumerate(t.shape) if s != 1] for t in terms]
# Get the shape of the output
shape = np.ones(ndim, dtype=int)
for ax, term in zip(axes, terms):
shape[np.array(ax)] = term.squeeze().shape
total_elts = shape.prod()
if (total_elts > 1e8):
assert 0, 'Don\'t do this on a cpu! Too many terms: %d' % (
int(total_elts))
# Build a meshgrid out of each of the terms over the right axes
# and sum. Doing it this way because np.einsum doesn't work
# for matrix multiplication in log space - we can't do np.einsum
# but add instead of multiply over indices
ans = np.zeros(shape)
for ax, term in zip(axes, terms):
for _ in range(ndim - term.ndim):
term = term[..., None]
ans += np.broadcast_to(term, ans.shape)
return fbsData(ans, max_fbs_axis)
def vBaseCase( self, node ):
return fbsData( np.array( [] ), -1 )
