morphoclass.model_utils module

Utilities for GNN models. Currently, only the hierarchical labels.

class morphoclass.model_utils.HierarchicalLabels(labels, adj, flat_to_hierarchical_oh)

Bases: object

Class representing a set of hierarchical labels.

It holds the structural information of a set of hierarchical labels and is meant to be used conjointly with an appropriate hierarchical model.

The constructor constructs class instance from a given label adjacency matrix.

It takes labels for all hierarchical nodes and their hierarchical relationship in terms of the adjacency matrix. Additionally a dictionary mapping flat labels (=leaf nodes) to the hierarchical one-hot labels must be provided.

Parameters

  • labels – List of labels for all hierarchical nodes.

  • adj – Adjacency matrix representing the label hierarchy.

  • flat_to_hierarchical_oh – dictionary mapping flat labels to hierarchical one-hot labels.

Returns

An instance of the HierarchicalLabels class.

Return type

instance

property adj

Get the adjacency matrix for hierarchical labels.

Returns

Return type

The adjacency matrix representing hierarchical labels.

flat_labels()

Flat labels in numerical form.

Returns

The numerical values of the flat labels.

Return type

list

flat_to_hierarchical_oh(n)

Map flat numerical labels to hierarchical one-hot labels.

Parameters

n – Flat numerical label.

Returns

One-hot hierarchical label.

Return type

oh_label

classmethod from_adjacency_matrix(labels, adj, flat_to_hierarchical_oh)

Construct class instance from a given label adjacency matrix.

This factory takes labels for all hierarchical nodes and their hierarchical relationship in terms of the adjacency matrix. Additionally a dictionary mapping flat labels (=leaf nodes) to the hierarchical one-hot labels must be provided.

Parameters

  • labels – List of labels for all hierarchical nodes.

  • adj – Adjacency matrix representing the label hierarchy.

  • flat_to_hierarchical_oh – dictionary mapping flat labels to hierarchical one-hot labels.

Returns

An instance of the HierarchicalLabels class.

Return type

instance

classmethod from_flat_labels(label_dict)

Construct a class instance from a label dictionary.

This factory expects a dictionary mapping numerical flat labels, i.e. leaf nodes of the hierarchical tree, to the list of hierarchical nodes representing the path from the root of the hierarchy to the given leaf node.

For example, a valid representation for the following hierarchy

  A      B
 / |   /  |
a  b  c   d
     /|                x y z

would be given by

label_dict = {
    0: ['A', 'a'],
    1: ['A', 'b'],
    2: ['B, 'c', 'x'],
    3: ['B, 'c', 'y'],
    4: ['B, 'c', 'z'],
    5: ['B', 'd'],
}
Parameters

label_dict (dict) – Dictionary mapping dense numerical flat labels to a list of hierarchical string labels.

gen_class_masks(parent_mask=None)

Generate masks for labels probabilities of which should sum to 1.

Parameters

parent_mask – The labels to start with, defaults to root nodes.

Yields

mask – All possible masks for nodes with sums of probabilities equal to 1.

get_class_segmentation_mask()

Create a segmentation mask for labels.

The mast is such that each softmax block is marked by a different integer.

Returns

The class segmentation mask.

Return type

segmentation_mask

property labels

Get the labels.

Returns

Return type

The labels.

n_leafs()

Get the number of leaf nodes.

This is the same as the number of flat labels.

Returns

Return type

The number of leaf nodes.

property roots

Get the label roots.

Returns

Return type

The label roots.

class morphoclass.model_utils.HierarchicalLabelsDeprecated(label_dict)

Bases: object

Class representing a set of hierarchical labels.

The attributes are initialized from given label_dict

The argument label_dict is a dictionary mapping numerical labels to a list of strings representing the hierarchical labels. For example [‘animal’, ‘mammal’, ‘dog’, ‘Corgi’] would represent the hierarchical label for ‘Corgi’, which is a subclass of ‘dog’, which is a subclass of ‘mammal’, which is a subclass of ‘animal’.

Parameters

label_dict (dict) – Dictionary mapping dense numerical labels to a list of hierarchical string labels.

get_mask(order=0)

Get a mask for the nodes corresponding to a given hierarchy order.

The order 0 in the tree hierarchy is the root.

Actually all of the non-trivial masks are already computed in is_tree so if performance starts to matter one could extend is_tree to cache the child_mask and return.

static is_tree(tree_idx)

Check if the given parents tensor represents a valid tree.

Start with the parents and iteratively visit the children, and mark the visited nodes. Continue until all nodes are visited. If at one iteration step the number of visited nodes does not increase (so there’s a disconnected loop) or a child is found to have been seen before, then parents does not represent a tree.

Parameters

tree_idx – The tree index.

Returns

Whether or not the given tree index represents a valid tree.

Return type

bool

to_one_hot(dense_label)

Convert dense labels to one-hot labels.

Parameters

dense_label – Given dense labels.

Returns

Return type

One-hot labels.