def __call__(self, model_name):
predictions = {}
print('Evaluating "%s" on "%s".' % (model_name, self.args['test']),
flush=True)
# load k-means model
kmeans = pickle.load(open(os.path.join(".", 'kmeans.pickle'), 'rb'))
print('The number of clusters is %d.' % (kmeans.n_clusters),
flush=True)
# partition test data according to k-means model
clusters_test = kmeans.predict(self.Y_all_test - self.X_all_test)
try:
with np.load(os.path.join(".", '%s.test.npz') % model_name) as npz:
Y_hat_clusters = {
int(cluster): npz[cluster]
for cluster in npz.files
}
except FileNotFoundError:
Y_hat_clusters = {}
if kmeans.n_clusters != len(Y_hat_clusters):
print('Missing the output for the model "%s"!' % model_name,
file=sys.stderr,
flush=True)
return predictions
# get estimated hypernyms for each term in test/validation set
Y_all_hat = extract(clusters_test, Y_hat_clusters)
# ensure we have the same number of estimates as we have of test terms
assert len(self.subsumptions_test) == Y_all_hat.shape[0]
# compute unit-norm of hypernym estimates
Y_all_hat_norm = Y_all_hat / np.linalg.norm(Y_all_hat,
axis=1)[:, np.newaxis]
# find similar words
print('nn_vec...')
similar_indices = nn_vec(Y_all_hat_norm,
self.w2v.syn0norm,
topn=15,
sort=True,
return_sims=False,
nthreads=self.args['threads'],
verbose=False)
print('nn_vec results covert...')
similar_words = [[self.w2v.index2word[ind] for ind in row]
for row in similar_indices]
print('done')
for i, (hyponym, hypernym) in enumerate(self.subsumptions_test):
predictions[hyponym] = similar_words[i]
return predictions
if kmeans.n_clusters != len(Y_hat_clusters):
print('Missing the output for the model "%s"!' % model)
continue
Y_all_hat = extract(clusters_test, Y_hat_clusters)
assert len(subsumptions_test) == Y_all_hat.shape[0]
measures = [{} for _ in range(10)]
if not args['non_optimized']:
# normalize Y_all_hat to make dot product equeal to cosine and monotonically decreasing function of euclidean distance
Y_all_hat_norm = Y_all_hat / np.linalg.norm(Y_all_hat,axis=1)[:,np.newaxis]
print('nn_vec...')
similar_indices = nn_vec(Y_all_hat_norm, w2v.syn0norm, topn=10, sort=True, return_sims=False, nthreads=args['threads'], verbose=False)
print('nn_vec results covert...')
similar_words = [[w2v.index2word[ind] for ind in row] for row in similar_indices]
print('done')
file_ptr_ms = open(str(model)+"_test_candidates1",'w')
file_ptr_hypo = open("test_hypo1",'w')
file_ptr_gold = open("test_gold1",'w')
prev_hypo = ''
gold_list = ''
out_ms = ''
count = 0
for i, (hyponym, hypernym) in enumerate(subsumptions_test):
if args['non_optimized']:
Y_hat = Y_all_hat[i].reshape(X_all_test.shape[1],)
actual = [w for w,_ in w2v.most_similar(positive=[Y_hat], topn=10)]
else: