def pyquest_spin_bintrees(data,row_alpha=0.5,col_alpha=0.5,beta=1.0,bal_constant=1.0,n_iters=3,n_spin=10):
"""
runs what is momentarily the standard questionnaire algorithm:
initial affinity = mutual cosine similarity
initial tree based on median of successive eigenvectors
dual affinities based on earth mover distance.
dual trees based on eigen_cut method
"""
#Generate initial affinity
init_row_aff = affinity.mutual_cosine_similarity(data.T,False,0,threshold=0.1)
#Compute diffusion embedding of initial affinities
init_row_vecs,init_row_vals = markov.markov_eigs(init_row_aff, 12)
#Generate median trees
init_row_tree = bintree_construct.median_tree(init_row_vecs,init_row_vals,max_levels=12)
row_trees, col_trees = [],[]
for _ in xrange(n_spin):
dual_col_trees = []
dual_row_trees = [init_row_tree]
for _ in xrange(n_iters):
dual_col_trees.append(bintree_construct.eigen_tree(data,dual_row_trees[-1],alpha=col_alpha,beta=beta,bal_constant=bal_constant))
dual_row_trees.append(bintree_construct.eigen_tree(data.T,dual_col_trees[-1],alpha=row_alpha,beta=beta,bal_constant=bal_constant))
row_trees.append(dual_row_trees[-1])
col_trees.append(dual_col_trees[-1])
return row_trees,col_trees
def flex_tree_diffusion(affinity, penalty_constant, n_eigs=12):
"""
affinity is an nxn affinity matrix.
Creates a flexible tree by calculating the diffusion on the given affinity.
Then clusters at each level by the flexible tree algorithm. For each level
up, doubles the diffusion time.
penalty_constant is the multiplier of the median diffusion distance.
"""
#First, we calculate the first n eigenvectors and eigenvalues of the
#diffusion
cluster_list = []
vecs, vals = markov.markov_eigs(affinity, n_eigs)
diff_time = 1.0
q = np.eye(affinity.shape[0])
while 1:
#now we calculate the diffusion distances between points at the
#current diffusion time.
diff_vecs = vecs.dot(np.diag(vals**diff_time))
diff_dists = spsp.distance.squareform(spsp.distance.pdist(diff_vecs))
#we take the affinity between clusters to be the average diffusion
#distance between them.
avg_dists = q.dot(diff_dists).dot(q.T)
#now we cluster the points based on this distance
cluster_list.append(cluster_from_distance(avg_dists, penalty_constant))
#if there is only one node left, then we are done.
#otherwise, add another level to the tree, double the diffusion time
#and keep going.
if len(cluster_list[-1]) == 1:
break
temp_tree = clusterlist_to_tree(cluster_list)
cpart = ClusteringPartition(
[x.elements for x in temp_tree.dfs_level(2)])
q, _ = cluster_transform_matrices(cpart)
diff_time *= 2.0
return clusterlist_to_tree(cluster_list)
def flex_tree_diffusion(affinity, penalty_constant, n_eigs=12):
"""
affinity is an nxn affinity matrix.
Creates a flexible tree by calculating the diffusion on the given affinity.
Then clusters at each level by the flexible tree algorithm. For each level
up, doubles the diffusion time.
penalty_constant is the multiplier of the median diffusion distance.
"""
# First, we calculate the first n eigenvectors and eigenvalues of the
# diffusion
cluster_list = []
vecs, vals = markov.markov_eigs(affinity, n_eigs)
diff_time = 1.0
q = np.eye(affinity.shape[0])
while 1:
# now we calculate the diffusion distances between points at the
# current diffusion time.
diff_vecs = vecs.dot(np.diag(vals ** diff_time))
diff_dists = spsp.distance.squareform(spsp.distance.pdist(diff_vecs))
# we take the affinity between clusters to be the average diffusion
# distance between them.
avg_dists = q.dot(diff_dists).dot(q.T)
# now we cluster the points based on this distance
cluster_list.append(cluster_from_distance(avg_dists, penalty_constant))
# if there is only one node left, then we are done.
# otherwise, add another level to the tree, double the diffusion time
# and keep going.
if len(cluster_list[-1]) == 1:
break
temp_tree = clusterlist_to_tree(cluster_list)
cpart = ClusteringPartition([x.elements for x in temp_tree.dfs_level(2)])
q, _ = cluster_transform_matrices(cpart)
diff_time *= 2.0
return clusterlist_to_tree(cluster_list)
def pyquest_newtree(data,tree_constant=0.25,row_alpha=0.5,col_alpha=0.5,beta=1.0,n_iters=3):
init_row_aff = affinity.mutual_cosine_similarity(data.T,False,0,threshold=0.1)
#Compute diffusion embedding of initial affinities
init_row_vecs,init_row_vals = markov.markov_eigs(init_row_aff, 12)
init_row_vals[np.isnan(init_row_vals)] = 0.0
row_embedding = init_row_vecs.dot(np.diag(init_row_vals))
row_distances = spsp.distance.squareform(spsp.distance.pdist(row_embedding))
row_affinity = np.max(row_distances) - row_distances
#Generate initial tree
#print "call1 tree_constant:{}".format(tree_constant)
init_row_tree = tree_building.make_tree_embedding(row_affinity,tree_constant)
dual_col_trees = []
dual_row_trees = [init_row_tree]
for _ in xrange(n_iters):
#print "Beginning iteration {}".format(i)
col_emd = dual_affinity.calc_emd(data,dual_row_trees[-1],alpha=col_alpha,beta=beta)
col_aff = dual_affinity.emd_dual_aff(col_emd)
#print "call2 tree_constant:{}".format(tree_constant)
dual_col_trees.append(tree_building.make_tree_embedding(col_aff,tree_constant))
row_emd = dual_affinity.calc_emd(data.T,dual_col_trees[-1],alpha=row_alpha,beta=beta)
row_aff = dual_affinity.emd_dual_aff(row_emd)
#print "call3 tree_constant:{}".format(tree_constant)
dual_row_trees.append(tree_building.make_tree_embedding(row_aff,tree_constant))
col_tree = dual_col_trees[-1]
row_tree = dual_row_trees[-1]
col_emd = dual_affinity.calc_emd(data,row_tree,alpha=col_alpha,beta=beta)
row_emd = dual_affinity.calc_emd(data.T,col_tree,alpha=row_alpha,beta=beta)
row_aff = dual_affinity.emd_dual_aff(row_emd)
col_aff = dual_affinity.emd_dual_aff(col_emd)
row_vecs,row_vals = markov.markov_eigs(row_aff, 12)
col_vecs,col_vals = markov.markov_eigs(col_aff, 12)
return row_tree,col_tree,row_vecs,col_vecs,row_vals,col_vals
def zero_eigen_cut(node,affinity):
"""
Returns the cut of the affinity matrix (cutting at zero)
corresponding to the elements in node, under the condition of bal_constant.
"""
new_data = affinity[node.elements,:][:,node.elements]
vecs,_ = markov.markov_eigs(new_data, 2)
labels = vecs[:,1] < 0.0
return labels
def zero_eigen_cut(node, affinity):
"""
Returns the cut of the affinity matrix (cutting at zero)
corresponding to the elements in node, under the condition of bal_constant.
"""
new_data = affinity[node.elements, :][:, node.elements]
vecs, _ = markov.markov_eigs(new_data, 2)
labels = vecs[:, 1] < 0.0
return labels
def random_dyadic_cut(node,affinity,left,right):
"""
Returns a randomized cut of the affinity matrix (cutting at zero)
corresponding to the elements in node, under the condition of bal_constant.
"""
new_data = affinity[node.elements,:][:,node.elements]
vecs,_ = markov.markov_eigs(new_data, 2)
eig = vecs[:,1]
eig_sorted = eig.argsort().argsort()
cut_loc = np.random.randint(left,right+1)
labels = eig_sorted < cut_loc
return labels
def random_dyadic_cut(node, affinity, left, right):
"""
Returns a randomized cut of the affinity matrix (cutting at zero)
corresponding to the elements in node, under the condition of bal_constant.
"""
new_data = affinity[node.elements, :][:, node.elements]
vecs, _ = markov.markov_eigs(new_data, 2)
eig = vecs[:, 1]
eig_sorted = eig.argsort().argsort()
cut_loc = np.random.randint(left, right + 1)
labels = eig_sorted < cut_loc
return labels
def break_node(train_data,
col_tree_node,
row_tree,
regressors=None,
k=5,
alpha=0.0,
beta=1.0,
col_emd=None):
"""
First calculates the EMD on the columns of train_data
in col_tree_node.elements using row_tree. Converts that to an affinity.
Calculates the second eigenvector of the markov matrix based on that
affinity.
Then fits a linear model using the rows in regressors (all if it's None)
and uses the LASSO path to identify the best k rows.
Splits the node using the predicted eigenvector values.
"""
import sklearn.linear_model as sklm
col_indices = col_tree_node.elements
node_data = train_data[:, col_indices].astype(np.float64)
if col_emd is None:
col_emd = dual_affinity.calc_emd(node_data, row_tree, alpha, beta)
col_aff = dual_affinity.emd_dual_aff(col_emd)
else:
col_aff = dual_affinity.emd_dual_aff(
col_emd[:, col_indices][col_tree_node.elements, :])
vecs, _ = markov.markov_eigs(col_aff, 2)
eig = vecs[:, 1]
if regressors is None:
regressors = range(row_tree.size)
_, active, _ = sklm.lars_path(node_data[regressors, :].T, eig, max_iter=50)
regr_indices = active[0:k]
lm = sklm.LinearRegression()
lm.fit(node_data[regr_indices, :].T, eig)
pred_eigs = lm.predict(node_data[regr_indices, :].T)
labels = pred_eigs > 0.0
partition = labels * np.ones(labels.shape[0])
col_tree_node.create_subclusters(partition)
return np.array([regressors[x] for x in regr_indices]), lm
def eigen_cut_zero(node,emd,eps=1.0):
affinity = dual_affinity.emd_dual_aff(emd[node.elements,:][:,node.elements]
,eps)
try:
vecs,_ = markov.markov_eigs(affinity,2)
except:
print affinity
print emd
print node.elements
raise
eig = vecs[:,1]
n = len(eig)
labels = np.ones(n)
labels *= (eig > 0.0)
return labels
def bal_eigen_cut(node,emd,bal_constant=1.0,eps=1.0):
affinity = dual_affinity.emd_dual_aff(emd[node.elements,:][:,node.elements]
,eps)
try:
vecs,_ = markov.markov_eigs(affinity,2)
except:
print affinity
print emd
print node.elements
raise
eig = vecs[:,1]
eig_sorted = np.argsort(eig)
n = len(eig)
l,r = bal_cut(n,bal_constant)
cut_loc = np.random.randint(l,r+1)
labels = np.zeros(n,np.int)
labels[eig_sorted[0:cut_loc]] = 1
return labels
def break_node(train_data,col_tree_node,row_tree,regressors=None,
k=5,alpha=0.0,beta=1.0,col_emd=None):
"""
First calculates the EMD on the columns of train_data
in col_tree_node.elements using row_tree. Converts that to an affinity.
Calculates the second eigenvector of the markov matrix based on that
affinity.
Then fits a linear model using the rows in regressors (all if it's None)
and uses the LASSO path to identify the best k rows.
Splits the node using the predicted eigenvector values.
"""
import sklearn.linear_model as sklm
col_indices = col_tree_node.elements
node_data = train_data[:,col_indices].astype(np.float64)
if col_emd is None:
col_emd = dual_affinity.calc_emd(node_data,row_tree,alpha,beta)
col_aff = dual_affinity.emd_dual_aff(col_emd)
else:
col_aff = dual_affinity.emd_dual_aff(col_emd[:,col_indices][col_tree_node.elements,:])
vecs,_ = markov.markov_eigs(col_aff,2)
eig = vecs[:,1]
if regressors is None:
regressors = range(row_tree.size)
_,active,_ = sklm.lars_path(node_data[regressors,:].T,eig,max_iter=50)
regr_indices = active[0:k]
lm = sklm.LinearRegression()
lm.fit(node_data[regr_indices,:].T,eig)
pred_eigs = lm.predict(node_data[regr_indices,:].T)
labels = pred_eigs > 0.0
partition = labels*np.ones(labels.shape[0])
col_tree_node.create_subclusters(partition)
return np.array([regressors[x] for x in regr_indices]),lm
def eigen_cut(node,emd,noise,eps=1.0):
affinity = dual_affinity.emd_dual_aff(emd[node.elements,:][:,node.elements]
,eps)
try:
vecs,_ = markov.markov_eigs(affinity,2)
except:
print affinity
print emd
print node.elements
raise
eig = vecs[:,1]
eig_sorted = np.sort(eig)
n = len(eig_sorted)
rnoise = np.random.uniform(-noise,noise)
if noise < 1e-8:
labels = np.zeros(n)
labels[np.argsort(eig)[0:int(n/2)]] = 1
else:
cut_loc = eig_sorted[int((n/2)+(rnoise*n))]
labels = np.ones(n)*(eig > cut_loc)
return labels
def calc_col_embedding(self):
self.col_vecs, self.col_vals = markov.markov_eigs(self.col_affinity, 8)
Publisher.sendMessage("embed.col.calc")
def calc_row_embedding(self):
self.row_vecs, self.row_vals = markov.markov_eigs(self.row_affinity, 8)
Publisher.sendMessage("embed.row.calc")
def pyquest(data,params):
#params should be a PyQuestParams object
Publisher.sendMessage("status.bar", "Calculating initial affinity...")
if params.init_aff_type == INIT_AFF_COS_SIM:
init_row_aff = affinity.mutual_cosine_similarity(
data.T,False,0,threshold=params.init_aff_threshold)
elif params.init_aff_type == INIT_AFF_GAUSSIAN:
#add KNN to the page
init_row_aff = affinity.gaussian_euclidean(
data.T, 5, params.init_aff_epsilon)
#Compute diffusion embedding of initial affinities
init_row_vecs,init_row_vals = markov.markov_eigs(init_row_aff, 12)
init_row_vals[np.isnan(init_row_vals)] = 0.0
row_embedding = init_row_vecs.dot(np.diag(init_row_vals))
row_distances = spsp.distance.squareform(spsp.distance.pdist(row_embedding))
row_affinity = np.max(row_distances) - row_distances
#Generate initial tree
#print "call1 tree_constant:{}".format(tree_constant)
Publisher.sendMessage("status.bar", "Calculating initial row tree...")
if params.tree_type == TREE_TYPE_BINARY:
init_row_tree = bintree_construct.median_tree(
init_row_vecs,init_row_vals,max_levels=12)
elif params.tree_type == TREE_TYPE_FLEXIBLE:
# init_row_tree = tree_building.make_tree_embedding(
# row_affinity,params.tree_constant)
init_row_tree = tree_building.make_tree_embedding(
row_affinity,params.tree_constant)
dual_col_trees = []
dual_row_trees = [init_row_tree]
row_tree_descs = ["Initial tree"]
col_tree_descs = []
for i in xrange(params.n_iters):
message = "Iteration {}: calculating column affinity...".format(i)
Publisher.sendMessage("status.bar", message)
#print "Beginning iteration {}".format(i)
if params.col_affinity_type == DUAL_EMD:
col_emd = dual_affinity.calc_emd(data,dual_row_trees[-1],
params.col_alpha,params.col_beta)
col_aff = dual_affinity.emd_dual_aff(col_emd)
elif params.col_affinity_type == DUAL_GAUSSIAN:
print "Gaussian dual affinity not supported at the moment."
return None
message = "Iteration {}: calculating column tree...".format(i)
Publisher.sendMessage("status.bar", message)
if params.tree_type == TREE_TYPE_BINARY:
col_tree = bintree_construct.eigen_tree(data,dual_row_trees[-1],
params.col_alpha,params.col_beta,params.tree_bal_constant)
elif params.tree_type == TREE_TYPE_FLEXIBLE:
col_tree = tree_building.make_tree_embedding(col_aff,
params.tree_constant)
dual_col_trees.append(col_tree)
col_tree_descs.append("Iteration {}".format(i))
message = "Iteration {}: calculating row affinity...".format(i)
Publisher.sendMessage("status.bar", message)
if params.row_affinity_type == DUAL_EMD:
row_emd = dual_affinity.calc_emd(data.T,dual_col_trees[-1],
params.row_alpha,params.row_beta)
row_aff = dual_affinity.emd_dual_aff(row_emd)
elif params.row_affinity_type == DUAL_GAUSSIAN:
print "Gaussian dual affinity not supported at the moment."
return None
message = "Iteration {}: calculating row tree...".format(i)
Publisher.sendMessage("status.bar", message)
if params.tree_type == TREE_TYPE_BINARY:
row_tree = bintree_construct.eigen_tree(data.T,dual_col_trees[-1],
params.row_alpha,params.row_beta,params.tree_bal_constant)
elif params.tree_type == TREE_TYPE_FLEXIBLE:
row_tree = tree_building.make_tree_embedding(row_aff,
params.tree_constant)
dual_row_trees.append(row_tree)
row_tree_descs.append("Iteration {}".format(i))
quest_run_desc = "{}".format(datetime.datetime.now())
return PyQuestRun(quest_run_desc,dual_row_trees,dual_col_trees,
row_tree_descs,col_tree_descs,params)
