git clone https://github.com/skojaku/node2vec.git
cd node2vec
pip install -e .Gensim implementation is by far the most common and fastest implementation of node2vec.
import node2vecs
dim = 32 # embedding dimension
model = node2vecs.GensimNode2Vec(vector_size = dim)
model.fit(A) # adjacency matrix in scipy csr_matrix format
emb = model.transform()Torch implementation of node2vec, with hyperparameters aligned to those of gensim as much as possible.
import node2vecs
dim = 32 # embedding dimension
model = node2vecs.TorchNode2Vec(vector_size = dim)
model.fit(A) # adjacency matrix in scipy csr_matrix format
emb = model.transform()Torch implementation gives more control over the calculations.
mport node2vecs
import networkx as nx
import numpy as np
G = nx.karate_club_graph()
A = nx.adjacency_matrix(G)
labels = np.unique([d[1]["club"] for d in G.nodes(data=True)], return_inverse=True)[1]
n_nodes = A.shape[0]
dim = 32
# Specify the sentense generator.
# The sentence generator has ``sampling'' method which
# generates a sequence of nodes for training word2vec.
sampler = node2vecs.RandomWalkSampler(A, walk_length=80)
# Negative node sampler used in negative sampling
# The default is ConfigModelNodeSampler which samples a node
# with probability proportional to it's degree.
# Changing noise sampler is useful to debias embedding.
# utils.node_sampler has different noise sampler.
# See residual2vec paper.
noise_sampler = node2vecs.utils.node_sampler.ConfigModelNodeSampler(ns_exponent=1.0)
noise_sampler.fit(A)
# Word2Vec model
model = node2vecs.Word2Vec(vocab_size=n_nodes, embedding_size=dim, padding_idx=n_nodes)
# Loss function
loss_func = node2vecs.Node2VecTripletLoss(n_neg=1)
# Set up negative sampler
dataset = node2vecs.TripletDataset(
adjmat=A,
num_walks=10,
window_length=10,
noise_sampler=noise_sampler,
padding_id=n_nodes,
buffer_size=1e4,
context_window_type="double", # we can limit the window to cover either side of center words. `context_window_type="double"` specifies a context window that extends both left and right of a focal node. context_window_type="left" or ="right" specifies that the window extends left or right, respectively.
epochs=5, # number of epochs
negative=1, # number of negative node per context
p=1, # (inverse) weight for the probability of backtracking walks
q=1, # (inverse) weight for the probability of depth-first walks
walk_length=80, # Length of walks
)
# Train
node2vecs.train(
model=model,
dataset=dataset,
loss_func=loss_func,
batch_size=256 * 4,
device="cpu", # gpu is also available
learning_rate=1e-3,
num_workers=4,
)
model.eval()
in_vec = model.ivectors.weight.data.cpu().numpy()[:n_nodes, :] # in_vector
out_vec = model.ovectors.weight.data.cpu().numpy()[:n_nodes, :] # out_vector
emb = model.embedding()import node2vecs
dim = 32 # embedding dimension
model = node2vecs.PYGNode2Vec(vector_size = dim)
model.fit(A) # adjacency matrix in scipy csr_matrix format
emb = model.transform()