Support for custom tree (tree_structure in python interface) about napkinxc HOT 2 CLOSED

Hi @xiaohan2012, thank you for your kind words.

It's true, that the Python interface doesn't have the tree_structure parameter right now (I would like to improve this functionality first before adding it), but constructors accept **kwargs that are passed to the underlying CPP module. This actually allows using all undocumented parameters that are implemented in src/args.cpp file also from Python (so for experimental purposes new options can be implemented just in CPP, without the need of updating the Python module).

So you can use the tree_structure out of the box, like in this example below that trains two PLTs: first trains one constructed its tree using hierarchical k-means clustering, the second one loads a tree created by for the first one.

from napkinxc.datasets import load_dataset
from napkinxc.models import PLT
from napkinxc.measures import precision_at_k

X_train, Y_train = load_dataset("eurlex-4k", "train")
X_test, Y_test = load_dataset("eurlex-4k", "test")
plt = PLT("eurlex-model")
plt.fit(X_train, Y_train)

Y_pred = plt.predict(X_test, top_k=5)
print("Precision at k:", precision_at_k(Y_test, Y_pred, k=5))

plt2 = PLT("eurlex-model2", tree_structure="eurlex-model/tree", verbose=True) # I added the verbose option here as a  proof, it will print the confirmation that the tree was loaded from a given file.
plt2.fit(X_train, Y_train)
Y_pred = plt2.predict(X_test, top_k=5)
print("Precision at k:", precision_at_k(Y_test, Y_pred, k=5)) # Since the tree was the same, this should give very similar results

When it comes to the tree format, it's pretty strict and limited right now:

• In the first line, it expects 2 numbers space separated: m - numbers of labels and t - numbers of tree nodes.
• Then t - 1 lines are expected, each specifies one tree node with two or three numbers space separated: p - id of the parent node, n - node id, l - label id (optional).
• 0 is always the id of the root node. p and n should be < t and l < m.

from napkinxc.

Thank you, it works!

from napkinxc.