def get_embedding_complex(text, word_weights):
words = text.split(' ')
topics = np.zeros(num_topics)
if embedding == 'LDA':
for w in words:
w_topics = text2topics.LDA(w, model, num_topics)
topics = topics + w_topics
topics = topics / len(words)
elif embedding == 'word2vec_mean':
topics = text2topics.word2vec_mean(text, word_weights, model,
num_topics)
elif embedding == 'word2vec_tfidf':
topics = text2topics.word2vec_tfidf(text, model, num_topics,
tfidf_model, tfidf_dictionary)
elif embedding == 'doc2vec':
topics = text2topics.doc2vec(text, model, num_topics)
elif embedding == 'glove':
topics = text2topics.glove(text, word_weights, model, num_topics)
elif embedding == 'glove_tfidf':
topics = text2topics.glove_tfidf(text, model, num_topics)
return topics
def get_results_complex(database, text, word_weights, num_results, results_path):
words = text.split(' ')
topics = np.zeros(num_topics)
if embedding == 'LDA':
for w in words:
w_topics = text2topics.LDA(w, model, num_topics)
topics = topics + w_topics
topics = topics / len(words)
elif embedding == 'word2vec_mean':
topics = text2topics.word2vec_mean(text, word_weights, model, num_topics)
elif embedding == 'word2vec_tfidf':
topics = text2topics.word2vec_tfidf(text, model, num_topics, tfidf_model, tfidf_dictionary)
topics = topics / len(words)
elif embedding == 'doc2vec':
topics = text2topics.doc2vec(text, model, num_topics)
elif embedding == 'glove':
topics = text2topics.glove(text, word_weights, model, num_topics)
elif embedding == 'glove_tfidf':
topics = text2topics.glove_tfidf(text, model, num_topics)
# Create empty dict for distances
distances = {}
print "Max text query: " + str(max(topics))
print "Max total query: " + str(max(topics))
# Compute distances
for id in database:
distances[id] = np.dot(database[id], topics)
# Sort dictionary
distances = sorted(distances.items(), key=operator.itemgetter(1), reverse=True)
# Get elements with min distances
for idx, id in enumerate(distances):
# Copy image results
if test_dataset == 'webvision':
copyfile('../../../datasets/WebVision/test_images_256/' + id[0] , results_path + id[0].replace('/', '_'))
else:
copyfile('../../../datasets/SocialMedia/img_resized_1M/cities_instagram/' + id[0] + '.jpg', results_path + id[0].replace('/', '_') + '.jpg')
if idx == num_results - 1: break
used_words += 1
topics = topics / used_words
elif embedding == 'word2vec_mean':
word_weights = 0
topics = text2topics.word2vec_mean(text_query, word_weights, model, num_topics)
elif embedding == 'word2vec_tfidf':
topics = text2topics.word2vec_tfidf(text_query, model, num_topics, tfidf_model, tfidf_dictionary)
elif embedding == 'doc2vec':
topics = text2topics.doc2vec(text_query, model, num_topics)
elif embedding == 'glove':
topics = text2topics.glove(text_query, model, num_topics)
elif embedding == 'glove_tfidf':
topics = text2topics.glove_tfidf(text_query, model, num_topics)
else:
print("Select a correct embedding")
raise SystemExit(0)
distances = {}
for id in database:
distances[id] = np.linalg.norm(database[id]-topics)
correct = 0
precisions = []
def get_results_complex(database, text, word_weights, im_id, num_results):
words = text.split(' ')
topics = np.zeros(num_topics)
if embedding == 'LDA':
for w in words:
w_topics = text2topics.LDA(w, model, num_topics)
topics = topics + w_topics
topics = topics / len(words)
elif embedding == 'word2vec_mean':
topics = text2topics.word2vec_mean(text, word_weights, model,
num_topics)
elif embedding == 'word2vec_tfidf':
topics = text2topics.word2vec_tfidf(text, model, num_topics,
tfidf_model, tfidf_dictionary)
elif embedding == 'doc2vec':
topics = text2topics.doc2vec(text, model, num_topics)
elif embedding == 'glove':
topics = text2topics.glove(text, word_weights, model, num_topics)
elif embedding == 'glove_tfidf':
topics = text2topics.glove_tfidf(text, model, num_topics)
if FC:
topics = topics - min(topics)
if max(topics) > 0:
topics = topics / max(topics)
topics = get_NN_txt_embedding.get_NN_txt_embedding(text_NN, topics)
topics = topics - min(topics)
topics = topics / max(topics)
# Create empty dict for distances
distances = {}
# Compute distances
for id in database:
distances[id] = np.dot(database[id], topics)
# Sort dictionary
distances = sorted(distances.items(),
key=operator.itemgetter(1),
reverse=True)
# Get elements with min distances
results = []
for idx, id in enumerate(distances):
copyfile('../../../datasets/COCO/val2014/' + id[0],
'../../../datasets/COCO/rr/' + id[0].replace('/', '_'))
results.append(int(id[0][13:-4]))
if idx == num_results - 1: break
if im_id in results:
return True
else:
return False
for cur_q in q:
cur_w = '1 1'
if len(cur_q.split(' ')) == 1:
if embedding == 'LDA':
topics = text2topics.LDA(cur_q, model, num_topics)
elif embedding == 'word2vec_mean':
topics = text2topics.word2vec_mean(cur_q, cur_w, model, num_topics)
elif embedding == 'doc2vec':
topics = text2topics.doc2vec(cur_q, model, num_topics)
elif embedding == 'word2vec_tfidf':
topics = text2topics.word2vec_tfidf(cur_q, model, num_topics,
tfidf_model, tfidf_dictionary)
elif embedding == 'glove':
topics = text2topics.glove(cur_q, cur_w, model, num_topics)
elif embedding == 'glove_tfidf':
topics = text2topics.glove_tfidf(cur_q, model, num_topics)
else:
topics = get_embedding_complex(cur_q, cur_w)
# Normalize by max to compare with net
topics = topics - min(topics)
topics = topics / max(topics)
f.write('word_images/' + cur_q.replace(' ', '_'))
for t in topics:
f.write(',' + str(t))
f.write('\n')
q.append('snow ski')
q.append('rain umbrella')
q.append('icecream beach')
q.append('chocolate cake')
q.append('pizza wine')
q.append('woman bag')
q.append('man boat')
q.append('kid dog')
file = open(out_file_path,'w')
for e,cur_q in enumerate(q):
if len(cur_q.split(' ')) == 1:
topics = text2topics.glove(cur_q, '1', model, num_topics)
if len(cur_q.split(' ')) == 2:
topics = text2topics.glove(cur_q, '0.5 0.5', model, num_topics)
file.write(cur_q)
for t in topics:
file.write(',' + str(t))
word_weights = 0
topics = text2topics.word2vec_mean(text_query, word_weights, model,
num_topics)
elif embedding == 'word2vec_tfidf':
topics = text2topics.word2vec_tfidf(text_query, model, num_topics,
tfidf_model, tfidf_dictionary)
# topics = topics + w_topics
# topics = topics / len(words)
elif embedding == 'doc2vec':
topics = text2topics.doc2vec(text_query, model, num_topics)
elif embedding == 'glove':
word_weights = 0
topics = text2topics.glove(text_query, word_weights, model, num_topics)
elif embedding == 'glove_tfidf':
topics = text2topics.glove_tfidf(text_query, model, num_topics)
# Create empty dict for ditances
distances = {}
topics = topics - min(topics)
topics = topics / max(topics)
# Compute distances)
for id in database:
distances[id] = np.dot(database[id], topics)
# Get elements with min distances
print(cur_q)
cur_w = w[e]
if test_dataset == 'webvision': results_path = "../../../datasets/WebVision/rr/" + data + "/" + cur_q.replace(' ', '_') + '__' + cur_w.replace(' ', '_') + '/'
else: results_path = "../../../datasets/SocialMedia/retrieval_results/" + data + "/" + cur_q.replace(' ', '_') + '__' + cur_w.replace(' ', '_') + '/'
if not os.path.exists(results_path):
print("Creating dir: " + results_path)
os.makedirs(results_path)
if len(cur_q.split(' ')) == 1:
if embedding == 'LDA': topics = text2topics.LDA(cur_q, model, num_topics)
elif embedding == 'word2vec_mean': topics = text2topics.word2vec_mean(cur_q, cur_w, model, num_topics)
elif embedding == 'doc2vec': topics = text2topics.doc2vec(cur_q, model, num_topics)
elif embedding == 'word2vec_tfidf': topics = text2topics.word2vec_tfidf(cur_q, model, num_topics, tfidf_model, tfidf_dictionary)
elif embedding == 'glove': topics = text2topics.glove(cur_q, cur_w, model, num_topics)
elif embedding == 'glove_tfidf': topics = text2topics.glove_tfidf(cur_q, model, num_topics)
get_results(database, topics, num_results,results_path)
else:
get_results_complex(database, cur_q, cur_w, num_results, results_path)