Finetune Node2Vec model

[paulthemagno]

Is there a way to update/finetune an already trained Node2Vec model with new examples? Or with an updated graph without re-computing all the Node2Vec model?

from torch_geometric.nn import Node2Vec
import os.path as osp
import torch
import matplotlib.pyplot as plt
from sklearn.manifold import TSNE
from torch_geometric.datasets import Planetoid
from tqdm.notebook import tqdm

dataset = 'Cora'
path = osp.join('.', 'data', dataset)
dataset = Planetoid(path, dataset)
data = dataset[0]

device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = Node2Vec(data.edge_index, embedding_dim=128, 
                 walk_length=20,                        # lenght of rw
                 context_size=10, walks_per_node=20,
                 num_negative_samples=1, 
                 p=1, q=1,                             # bias parameters
                 sparse=True).to(device)

[rusty1s]

You should be able to set the the embedding of the Node2Vec model to a pre-trained state:

model.embedding.weight = pretrained_model.embedding.weight

Keep in mind that the ordering of nodes need to stay the same. Node2Vec is a transductive learning model and cannot easily be applied to new graphs.