cannot find save model and fine-tune interface about graphvite

[KiddoZhu]

I am just curious about in what case you need fine-tuning? Maybe checkpoint is needed sometimes.

There isn't a perfect support for that, but you can do it in the Python code. Suppose you have created two GraphSolver with the same dimension for embeddings.

solver1.train()
with open(file_name, "wb") as fout:
    pickle.dump(solver1.vertex_embeddings, fout)
    pickle.dump(solver1.context_embeddings, fout)

For the same graph, you can fine tune like this.

with open(file_name, "rb") as fin:
    # Note LRS must use index. You should never directly assign value to vertex_embeddings.
    solver2.vertex_embeddings[:] = pickle.load(fin)
    solver2.context_embeddings[:] = pickle.load(fin)
solver2.train(resume=True)

Because embeddings correspond to nodes, this requires two solvers to use the same node set. If you use two graphs with different edges, the orders of nodes are usually different. You can rearrange them using the following code.

embeddings = solver1.vertex_embeddings
new_embeddings = np.empty_like(embeddings)
for id, embedding in enumerate(embeddings):
    new_embeddings[solver2.name2id[solver1.id2name[id]]] = embedding
solver2.vertex_embeddings[:] = new_embeddings

[KiddoZhu]

Added in v0.2.1.