representations = torch.Tensor(V)
#### unsupervised evaluation ####
# fitting GMM
algs = GaussianMixtureSKLearn(n_communities)
algs.fit(representations)
#get and transform prediction
prediction = algs.predict(representations).long()
prediction_mat = torch.LongTensor([[
1 if (y == prediction[i]) else 0 for y in range(n_communities)
] for i in range(len(X))])
# get the conductance
conductance = evaluation.conductance(prediction_mat, X)
# get the NMI
ground_truth = torch.LongTensor(
[[1 if (y in Y[i]) else 0 for y in range(n_communities)]
for i in range(len(X))])
nmi = evaluation.nmi(prediction_mat, ground_truth)
# get the accuracy
accuracy = evaluation.precision1_unsupervised_evaluation(
prediction, D.Y, n_communities)
## fill results
results_unsupervised = {
"accuracy": accuracy,
"conductance": conductance,
for i in range(len(X))])
if args.verbose:
print("\nEvaluate accuracy, conductence and normalised mutual information")
for i in range(args.n):
current_gmm = gmm_list[i]
accuracies.append(
evaluation.poincare_unsupervised_em(representations,
D.Y,
n_gaussian,
em=current_gmm,
verbose=False))
prediction = current_gmm.predict(representations)
prediction_mat = torch.LongTensor(
[[1 if (y == prediction[i]) else 0 for y in range(n_gaussian)]
for i in range(len(X))])
conductances.append(evaluation.conductance(prediction_mat, adjency_matrix))
nmis.append(evaluation.nmi(prediction_mat, ground_truth))
# log results
log_in.append({
log_prefix + "unsupervised_eval": {
"accuracy": accuracies,
"conductance": conductances,
"nmi": nmis
}
})
import numpy as np
# print results
print("Results of unsupervised evaluation ")
print("\t Mean accuracy : ", sum(accuracies, 0.) / args.n)
prediction,
D.Y,
n_gaussian,
))
prediction_mat = torch.zeros(len(X), n_gaussian).cuda()
gt_mat = [
torch.LongTensor(ground_truth[i]).cuda()
for i in tqdm.trange(len(D.Y))
]
print("create prediction matrix")
for i in tqdm.trange(len(D.Y)):
prediction_mat[i][prediction[i]] = 1
# prediction_mat = torch.LongTensor([[ 1 if(y == prediction[i]) else 0 for y in range(n_gaussian)] for i in range(len(X))])
print("Evaluate Conductance")
conductances.append(
evaluation.conductance(prediction_mat, adjency_matrix))
nmis.append(evaluation.nmi(prediction_mat, ground_truth))
# log results
log_in.append({
log_prefix + "unsupervised_eval_k-means": {
"accuracy": accuracies,
"conductance": conductances,
"nmi": nmis
}
})
import numpy as np
# print results
print("Results of unsupervised evaluation for ", directory)
print("\t Mean accuracy : ", sum(accuracies, 0.) / args.n)
import torch
from rcome.evaluation_tools.evaluation import nmi, conductance
from rcome.data_tools import data_loader
from rcome.clustering_tools import poincare_em
# loading data and learned embeddings
X, Y = data_loader.load_corpus("LFR1", directed=False)
n_gaussian = 13
representations = torch.load("LOG/all_loss/representation.pth")
ground_truth = torch.LongTensor([[1 if(y in Y[i]) else 0 for y in range(n_gaussian)] for i in range(len(X))])
# estimate the gmm
em_alg = poincare_em.PoincareEM(13)
em_alg.fit(representations)
# predict associated gaussian
prediction = em_alg.predict(representations)
prediction_mat = torch.LongTensor([[1 if(y in prediction[i]) else 0 for y in range(n_gaussian)] for i in range(len(X))])
conductance_scores = conductance(prediction_mat, X)
print("Conductance score is ", torch.Tensor(conductance_scores).mean().item(), "+-",
torch.Tensor(conductance_scores).std().item())
nmi_score = nmi(prediction_mat, ground_truth)
print("NMI score is ", nmi_score)
