Tree Vizualization

[nmanginas]

I noticed that there is no way to visualize the query trees (at least that I could find) and it has helped me quite a bit in debugging to be able to look at them. This involves writing a method in TabledAndOrTrees which I have done. Please let me know whether this is something you want to add to the code.

def to_dot(self, term: Term):
    if term not in self._and_or_tree:
        raise RuntimeError("term {} is not in the tree".format(term))
    from deepstochlog.logic import Or, And, NNLeaf, StaticProbability, TermLeaf

    nodes, edges = [], []
    num_and_nodes = 0
    num_or_nodes = 0

    def recurse_term_node(term: Term, parent_hash):
        node = self._and_or_tree[term]
        node_hash = abs(hash(node) + parent_hash)
        if isinstance(node, Or):
            nonlocal num_or_nodes
            num_or_nodes += 1
            nodes.append(
                "{} [label={}]".format(
                    node_hash,
                    '<OR<BR/> <FONT POINT-SIZE="7"> {} </FONT>>'.format(str(term)),
                )
            )
        elif isinstance(node, And):
            nonlocal num_and_nodes
            num_and_nodes += 1
            nodes.append(
                "{} [label={}]".format(
                    node_hash,
                    '<AND<BR/> <FONT POINT-SIZE="7"> {} </FONT>>'.format(str(term)),
                )
            )

        elif isinstance(node, NNLeaf):
            nodes.append('{} [label="{}"]'.format(node_hash, str(node)))
            return

        for child in node.children:
            child_hash = abs(hash(child) + node_hash)
            if not isinstance(child, TermLeaf):
                edges.append((node_hash, child_hash))
            resolve_other_node(child, node_hash)

    def resolve_other_node(node, parent_hash):
        node_hash = abs(hash(node) + parent_hash)
        if isinstance(node, TermLeaf):
            edges.append(
                (
                    parent_hash,
                    abs(hash(self._and_or_tree[node.term]) + node_hash),
                )
            )
            recurse_term_node(node.term, node_hash)
            return
        elif isinstance(node, Or):
            nonlocal num_or_nodes
            num_or_nodes += 1
            nodes.append("{} [label={}]".format(node_hash, '"OR"'))
        elif isinstance(node, And):
            nonlocal num_and_nodes
            num_and_nodes += 1
            nodes.append("{} [label={}]".format(node_hash, '"AND"'))
        elif isinstance(node, StaticProbability):
            nodes.append("{} [label={}]".format(node_hash, node.probability))
            return
        elif isinstance(node, NNLeaf):
            nodes.append('{} [label="{}"]'.format(node_hash, str(node)))
            return
        else:
            raise RuntimeError("unexpected node type")

        for child in node.children:
            child_hash = abs(hash(child) + node_hash)
            if not isinstance(child, TermLeaf):
                edges.append((node_hash, child_hash))
            resolve_other_node(child, node_hash)

    recurse_term_node(term, 0)
    dot_string = (
        "Digraph {\n"
        + "\n".join(nodes)
        + "\n"
        + "\n".join(
            [
                "{} -> {}".format(source, destination)
                for source, destination in edges
            ]
        )
        + "\n}"
    )
    print(
        "run circuit with {} nodes and {} edges ({} and nodes, {} or nodes)".format(
            len(nodes), len(edges), num_and_nodes, num_or_nodes
        )
    )
    return dot_string, {
        "num_nodes": len(nodes),
        "num_edges": len(edges),
        "num_or_nodes": num_or_nodes,
        "num_and_nodes": num_and_nodes,
    }

It is possible it can be done much more compactly by somehow utilizing the visitors similarly to how the tree is actually traversed when computing queries. This produces something of this sort (truncated here for ease).

If you are interested please let me know if there is anything else I can do to help.

[twinters]

Thanks a lot! That looks like something that we should have probably added somewhere too, e.g. in a demo notebook or something. I'll look into integrating this into the repo soon. Thanks again!