def gen_hierarchy(self, data):
"""
Create a new feature hierarchy:
(i) from input hierarchy if available, and
(ii) from feature set if not
"""
num_features = data.shape[1]
if self.feature_hierarchy is None:
# Create hierarchy if not available
if self.feature_names is None:
# Generate feature names if not available
self.feature_names = [f"{idx}" for idx in range(num_features)]
root = Feature(constants.DUMMY_ROOT, description=constants.DUMMY_ROOT, perturbable=False) # Dummy root node, shouldn't be perturbed
for idx, feature_name in enumerate(self.feature_names):
Feature(feature_name, parent=root, idx=[idx])
self.feature_hierarchy = root
else:
# TODO: Document real hierarchy with examples
# Input hierarchy needs a list of indices assigned to all base features
# Create hierarchy over features from input hierarchy
if isinstance(self.feature_hierarchy, str):
# JSON hierarchy - import to anytree
try:
importer = JsonImporter()
with open(self.feature_hierarchy, encoding="utf-8") as hierarchy_file:
self.feature_hierarchy = importer.read(hierarchy_file)
except JSONDecodeError as error:
raise ValueError(f"Feature hierarchy {self.feature_hierarchy} does not appear to be a valid JSON file:") from error
assert isinstance(self.feature_hierarchy, anytree.node.nodemixin.NodeMixin), "Feature hierarchy does not appear to be a valid JSON file or an anytree node"
feature_nodes = {}
all_idx = set()
# Parse and validate input hierarchy
for node in anytree.PostOrderIter(self.feature_hierarchy):
idx = []
if node.is_leaf:
valid = (hasattr(node, "idx") and
isinstance(node.idx, list) and
len(node.idx) >= 1 and
all(isinstance(node.idx[i], int) for i in range(len(node.idx))))
assert valid, f"Leaf node {node.name} must contain a non-empty list of integer indices under attribute 'idx'"
assert not all_idx.intersection(node.idx), f"Leaf node {node.name} has index overlap with other leaf nodes"
idx = node.idx
all_idx.update(idx)
else:
# Ensure internal nodes have empty initial indices
valid = not hasattr(node, "idx") or not node.idx
assert valid, f"Internal node {node.name} must have empty initial indices under attribute 'idx'"
description = getattr(node, "description", "")
feature_nodes[node.name] = Feature(node.name, description=description, idx=idx)
# Update feature group (internal node) indices and tree connections
assert min(all_idx) >= 0 and max(all_idx) < num_features, "Feature indices in hierarchy must be in range [0, num_features - 1]"
feature_node = None
for node in anytree.PostOrderIter(self.feature_hierarchy):
feature_node = feature_nodes[node.name]
parent = node.parent
if parent:
feature_node.parent = feature_nodes[parent.name]
for child in node.children:
feature_node.idx += feature_nodes[child.name].idx
self.feature_hierarchy = Feature(constants.DUMMY_ROOT, children=[feature_node], perturbable=False) # Dummy root node for consistency with flat hierarchy; last feature_node is original root
def simplify_tree(self):
"""
Goes through the tree and gets rid of nodes with only one child.
Replaced them with their only child.
Purges not needed Nodes from the node_dictionary
See Algorithm 1 (line 11)
"""
nodes = [node for node in at.PostOrderIter(self.root)]
for node in nodes:
children = node.children
if len(children) == 1:
children[0].parent = node.parent
if node.parent is None:
# The current root has only one child - let's replace it
# with the only child
self.root = children[0]
node.parent = None
try:
self.node_dictionary.pop(node.name)
except KeyError:
print(node.name)
try:
self.category_dictionary.pop(node.name)
except KeyError:
print(node.name)
# recompute distances for the entities
nodes = [node for node in at.PostOrderIter(self.root)]
for node in nodes:
if node.is_leaf:
node._compute_distances()
def update_hierarchy_relevance(hierarchy_root, relevant_feature_map, probs):
"""
Add feature relevance information to nodes of hierarchy:
their probabilty of being enabled,
their polynomial coefficient
"""
relevant_features = set()
for key in relevant_feature_map:
relevant_features.update(key)
for node in anytree.PostOrderIter(hierarchy_root):
node.description = constants.IRRELEVANT
if node.is_leaf:
idx = int(node.static_indices)
node.poly_coeff = 0.0
node.bin_prob = probs[idx]
coeff = relevant_feature_map.get(frozenset([idx]))
if coeff:
node.poly_coeff = coeff
node.description = (
"%s feature:\nPolynomial coefficient: %f\nBinomial probability: %f"
% (constants.RELEVANT, coeff, probs[idx]))
elif idx in relevant_features:
node.description = ("%s feature\n(Interaction-only)" %
constants.RELEVANT)
else:
for child in node.children:
if child.description != constants.IRRELEVANT:
node.description = constants.RELEVANT
def merkle_branch(hash_, tree):
tree_walker = anytree.PostOrderIter(tree)
node = next((node for node in tree_walker if node.name == hash_), None)
if not node:
return None
# TODO come up with a sane branch resprentation
return node
def get_input_wires(self):
if not self.input_wires:
leaves = [node.name for node
in anytree.PostOrderIter(self.tree) if node.is_leaf]
input_wires = [wire for leaf in leaves
for wire in [leaf.left_wire, leaf.right_wire]]
self.input_wires = input_wires
return self.input_wires
def prune_empty_groups(tree: LayerGroupNode, ) -> LayerGroupNode:
"""Remove any leaf nodes which are not LayerNodes."""
lt = copy.deepcopy(tree)
# NOTE: We MUST iterate using PostOrderIter(lt) instead of lt.leaves
# because removing a leaf group may leave its parent group newly enleafened.
for node in anytree.PostOrderIter(lt):
if node.is_leaf and type(node) is not LayerNode:
logger.warn(f'{node.group_name_path=}, {node.name=}')
_delete_node(node, msg='Removing empty group')
return lt
def evaluate_hierarchical(args, sfeatures, afeatures):
"""
Evaluate hierarchical analysis results - obtain power/FDR measures for all nodes/base features
"""
# pylint: disable = too-many-locals
# Map features in hierarchy to original features and identify ground-truth importances/scores
sfeatures_map = {sfeature.name: sfeature for sfeature in sfeatures}
importance_map = {}
score_map = {}
for node in anytree.PostOrderIter(afeatures[0].root):
if node.is_leaf:
sfeature_name = str(node.idx[0])
importance_map[node.name] = sfeatures_map[sfeature_name].important
score_map[node.name] = sfeatures_map[sfeature_name].effect_size
else:
importance_map[node.name] = any(importance_map[child.name] for child in node.children)
# Overall FDR/power
important = np.zeros(len(afeatures))
inferred_important = np.zeros(len(afeatures))
for idx, afeature in enumerate(afeatures):
important[idx] = importance_map[afeature.name]
inferred_important[idx] = afeature.important
imp_precision, imp_recall = get_precision_recall(important, inferred_important)
# Base features FDR/power
base_features = list(filter(lambda node: node.is_leaf, afeatures))
base_important = np.zeros(len(base_features))
inferred_base_important = np.zeros(len(base_features))
for idx, base_feature in enumerate(base_features):
base_important[idx] = importance_map[base_feature.name]
inferred_base_important[idx] = base_feature.important
base_imp_precision, base_imp_recall = get_precision_recall(base_important, inferred_base_important)
# Importance scores for base features
overall_scores_corr, overall_relevant_scores_corr = (1.0, 1.0)
if args.model_type == REGRESSOR:
scores = np.zeros(len(base_features))
inferred_scores = np.zeros(len(base_features))
for idx, base_feature in enumerate(base_features):
scores[idx] = score_map[base_feature.name]
inferred_scores[idx] = base_feature.importance_score
relevant_base_features = list(filter(lambda node: importance_map[node.name], base_features))
relevant_scores = np.zeros(len(relevant_base_features))
relevant_inferred_scores = np.zeros(len(relevant_base_features))
for idx, relevant_base_feature in enumerate(relevant_base_features):
relevant_scores[idx] = score_map[relevant_base_feature.name]
relevant_inferred_scores[idx] = relevant_base_feature.importance_score
overall_scores_corr = pearsonr(scores, inferred_scores)[0] if len(scores) >= 2 else 1
overall_relevant_scores_corr = pearsonr(relevant_scores, relevant_inferred_scores)[0] if len(relevant_scores) >= 2 else 1
vals = {FDR: 1 - imp_precision, POWER: imp_recall,
BASE_FEATURES_FDR: 1 - base_imp_precision, BASE_FEATURES_POWER: base_imp_recall,
OVERALL_SCORES_CORR: overall_scores_corr, OVERALL_RELEVANT_SCORES_CORR: overall_relevant_scores_corr}
return {key: value if isinstance(value, dict) else round(value, 10) for key, value in vals.items()} # Round values to avoid FP discrepancies
def delete_nodes(root_node, dry_run, delete_root, stop_node="root"):
"""
Cascade deletes OSDF nodes in a depth-first manner.
Args:
root_node (anytree.Node): The root node of the tree to delete from.
dry_run (boolean): True/False to delete or print out the nodes to
be deleted.
stop_node (string): Name of the node to stop deletion on. Defaults to
the root node but can be any OSDF ID to stop on.
delete_root (boolean): If the stop_node parameter is set to 'root'
this parameter can be passed to indicate we want to delete the
root node as well.
Requires:
None
Returns:
None
"""
deleted = []
failed_delete = []
for node in anytree.PostOrderIter(root_node):
if dry_run:
if not node.name == "root":
print "DELETING NODE:", node
else:
osdf_obj = node.osdf
if not node.name == "root":
print "DELETING NODE:", node
res = osdf_obj.delete()
if not res:
print "FAILED TO DELETE NODE:", node
failed_delete.append(osdf_obj)
if failed_delete:
print "WARNING: The following OSDF nodes were not deleted:" + "\n".join(failed_delete)
if delete_root and stop_node == "root":
print "DELETING ROOT NODE:", root_node
if not dry_run:
root_node.osdf.delete()
def gen_hierarchy(args, clustering_data):
"""
Generate hierarchy over features
Args:
args: Command-line arguments
clustering_data: Data potentially used to cluster features
(depending on hierarchy generation method)
Returns:
hierarchy_root: root fo resulting hierarchy over features
"""
# TODO: Get rid of possibly redundant hierarchy attributes e.g. vidx
# Generate hierarchy
hierarchy_root = None
if args.hierarchy_type == constants.FLAT:
args.contiguous_node_names = False # Flat hierarchy should be automatically created; do not re-index hierarchy
elif args.hierarchy_type == constants.CLUSTER_FROM_DATA:
clusters = cluster_data(clustering_data)
hierarchy_root = gen_hierarchy_from_clusters(args, clusters)
elif args.hierarchy_type == constants.RANDOM:
hierarchy_root = gen_random_hierarchy(args)
else:
raise NotImplementedError("Need valid hierarchy type")
# Improve visualization - contiguous feature names
feature_id_map = {} # mapping from visual feature ids to original ids
if args.contiguous_node_names:
for idx, node in enumerate(anytree.PostOrderIter(hierarchy_root)):
node.vidx = idx
if node.is_leaf:
node.min_child_vidx = idx
node.max_child_vidx = idx
node.num_base_features = 1
node.name = str(idx)
feature_id_map[idx] = node.idx[0]
else:
node.min_child_vidx = min(
[child.min_child_vidx for child in node.children])
node.max_child_vidx = max(
[child.vidx for child in node.children])
node.num_base_features = sum(
[child.num_base_features for child in node.children])
node.name = "[%d-%d] (size: %d)" % (node.min_child_vidx,
node.max_child_vidx,
node.num_base_features)
return hierarchy_root, feature_id_map
def compare_with_ground_truth(args, hierarchy_root):
"""Compare results from mihifepe with ground truth results"""
# Generate ground truth results
# Write hierarchical FDR input file for ground truth values
args.logger.info("Compare mihifepe results to ground truth")
input_filename = "%s/ground_truth_pvalues.csv" % args.output_dir
with open(input_filename, "w", newline="") as input_file:
writer = csv.writer(input_file)
writer.writerow([
constants.NODE_NAME, constants.PARENT_NAME,
constants.PVALUE_LOSSES, constants.DESCRIPTION
])
for node in anytree.PostOrderIter(hierarchy_root):
parent_name = node.parent.name if node.parent else ""
# Decide p-values based on rough heuristic for relevance
node.pvalue = 1.0
if node.description != constants.IRRELEVANT:
if node.is_leaf:
node.pvalue = 0.001
if node.poly_coeff:
node.pvalue = min(
node.pvalue,
1e-10 / (node.poly_coeff * node.bin_prob)**3)
else:
node.pvalue = 0.999 * min(
[child.pvalue for child in node.children])
writer.writerow(
[node.name, parent_name, node.pvalue, node.description])
# Generate hierarchical FDR results for ground truth values
ground_truth_dir = "%s/ground_truth_fdr" % args.output_dir
cmd = (
"python -m mihifepe.fdr.hierarchical_fdr_control -output_dir %s -procedure yekutieli "
"-rectangle_leaves %s" % (ground_truth_dir, input_filename))
args.logger.info("Running cmd: %s" % cmd)
pass_args = cmd.split()[2:]
with patch.object(sys, 'argv', pass_args):
hierarchical_fdr_control.main()
# Compare results
ground_truth_outputs_filename = "%s/%s.png" % (ground_truth_dir,
constants.TREE)
args.logger.info("Ground truth results: %s" %
ground_truth_outputs_filename)
mihifepe_outputs_filename = "%s/%s/%s.png" % (
args.output_dir, constants.HIERARCHICAL_FDR_DIR, constants.TREE)
args.logger.info("mihifepe results: %s" % mihifepe_outputs_filename)
def prune(self, error_bound):
clf = self.copy()
def compute_error(clf, original, alternative):
# just compute the error if no switch is needed
if original == alternative:
return error_bound(clf)
# switch the original node with the alternative in clf
old_parents = (original.parent, alternative.parent)
Node.replace(original, alternative)
if old_parents[0] is None:
clf.tree = alternative
# compute the error with the resulting tree
output = error_bound(clf)
# return to the original tree
Node.replace(alternative, original, old_parents[-1])
if old_parents[0] is None:
clf.tree = original
return output
# go over all nodes in a bottom-up manner
for node in anytree.PostOrderIter(clf.tree):
# add all alternatives to a list
alternatives = []
for label in clf.label_values:
alternatives.append(Node(label))
for child in node.children:
alternatives.append(child)
alternatives.append(node)
# find the alternative which minimizes the bound on the error
best_alternative = argmin(lambda x: compute_error(clf, node, x),
alternatives)
if best_alternative != node:
Node.replace(node, best_alternative)
node = best_alternative
if best_alternative.is_root:
clf.tree = best_alternative
return clf
def update_hierarchy_descriptions(hierarchy_root, relevant_feature_map,
features):
"""
Add feature relevance information to nodes of hierarchy
"""
relevant_features = set()
for key in relevant_feature_map:
relevant_features.update(key)
for node in anytree.PostOrderIter(hierarchy_root):
node.description = constants.IRRELEVANT
if node.is_leaf:
idx = node.idx[0]
coeff = relevant_feature_map.get(frozenset([idx]))
if coeff:
node.description = f"{constants.RELEVANT} feature:\nPolynomial coefficient: {coeff}\nSummary: {features[idx].summary()}"
elif idx in relevant_features:
node.description = f"{constants.RELEVANT} feature\n(Interaction-only)\nSummary: {features[idx].summary()}"
else:
for child in node.children:
if child.description != constants.IRRELEVANT:
node.description = constants.RELEVANT
root = root.children[0]
# Identify the workspaces
for display in root.children:
for wk in display.children[0].children:
wk.workspace = True
# Get the current workspace
proc_out = subprocess.run(['swaymsg', '-t', 'get_workspaces'], stdout=subprocess.PIPE)
wkList = json.loads(proc_out.stdout.decode('utf-8'))
focWkName = nf.getFocusedWK(wkList)
# Change the tree such that the workspaces are children to the root
# while ignoring the current workspace
root.children = [node
for node in at.PostOrderIter(root, filter_=lambda x: x.workspace)
if node.id != focWkName]
# If workspace contains only one container, then remove that container
for node in at.PostOrderIter(root, filter_=lambda x: x.workspace):
if len(node.children) == 1:
node.children = node.children[0].children
# If containers have only one element, then remove such containers
for node in at.PreOrderIter(root, filter_=lambda x: x.container):
if len(node.children) == 1:
node.children[0].parent = node.parent
node.parent = None
# Create names for containers
for node in at.PreOrderIter(root, filter_=lambda x: x.container):