def sentence_generator():
category_name, plural_name, categories = get_open_image_categories()
replace = dict(zip(plural_name, category_name))
category_set = set(category_name)
with open('data/sentences.pkl', 'r') as f:
captions = pkl.load(f)
if FLAGS.new_dict:
_create_vocab(captions)
with open('data/glove_vocab.pkl', 'r') as f:
glove = pkl.load(f)
glove.append('<S>')
glove.append('</S>')
glove = set(glove)
with open(FLAGS.word_counts_output_file, 'r') as f:
vocab = list(f)
vocab = [i.strip() for i in vocab]
vocab = [i.split() for i in vocab]
vocab = [(i, int(j)) for i, j in vocab if i in glove]
word_counts = [i for i in vocab if i[0] in category_set or i[1] >= 40]
words = set([i[0] for i in word_counts])
for i in category_name:
if i not in words:
word_counts.append((i, 0))
with open(FLAGS.word_counts_output_file, 'w') as f:
f.write('\n'.join(['%s %d' % (w, c) for w, c in word_counts]))
vocab = vocabulary.Vocabulary(FLAGS.word_counts_output_file)
all_ids = dict([(k, vocab.word_to_id(v)) for k, v in categories.items()])
with open('data/all_ids.pkl', 'w') as f:
pkl.dump(all_ids, f)
context = tf.train.Features()
random.shuffle(captions)
for c in captions:
for i, w in enumerate(c):
if w in replace:
c[i] = replace[w]
k = parse_key_words(c, category_set)
c = [vocab.word_to_id(word) for word in c]
if c.count(vocab.unk_id) > len(c) * 0.15:
continue
k = [vocab.word_to_id(i) for i in k]
feature_lists = tf.train.FeatureLists(
feature_list={
'key': _int64_feature_list(k),
'sentence': _int64_feature_list(c)
})
sequence_example = tf.train.SequenceExample(
context=context, feature_lists=feature_lists)
yield sequence_example.SerializeToString()
def image_generator(split):
with open('data/coco_%s.txt' % split, 'r') as f:
filename = list(f)
filename = [i.strip() for i in filename]
if split == 'train':
random.shuffle(filename)
with open('data/all_ids.pkl', 'r') as f:
all_ids = pkl.load(f)
with h5py.File('data/object.hdf5', 'r') as f:
for i in filename:
name = os.path.splitext(i)[0]
detection_classes = f[name + '/detection_classes'][:].astype(
np.int32)
detection_scores = f[name + '/detection_scores'][:]
detection_classes, ind = np.unique(detection_classes,
return_index=True)
detection_scores = detection_scores[ind]
detection_classes = [all_ids[j] for j in detection_classes]
image_path = FLAGS.image_path + '/' + i
with tf.gfile.FastGFile(image_path, 'r') as g:
image = g.read()
context = tf.train.Features(
feature={
'image/name': _bytes_feature(i),
'image/data': _bytes_feature(image),
})
feature_lists = tf.train.FeatureLists(
feature_list={
'classes': _int64_feature_list(detection_classes),
'scores': _float_feature_list(detection_scores)
})
sequence_example = tf.train.SequenceExample(
context=context, feature_lists=feature_lists)
yield sequence_example.SerializeToString()
def run(iname):
tf = infer.tf
if iname in self_caps:
sentence = self_caps[iname].split()
sentence = [infer.vocab.word_to_id(i) for i in sentence]
context = tf.train.Features()
feature_lists = tf.train.FeatureLists(
feature_list={'sentence': _int64_feature_list(sentence)})
sequence_example = tf.train.SequenceExample(
context=context, feature_lists=feature_lists)
return sequence_example.SerializeToString()
else:
sentence = [-1]
context = tf.train.Features()
feature_lists = tf.train.FeatureLists(
feature_list={'sentence': _int64_feature_list(sentence)})
sequence_example = tf.train.SequenceExample(
context=context, feature_lists=feature_lists)
return sequence_example.SerializeToString()
def run(classes):
tf = infer.tf
sentences = infer.infer(classes[::-1])
sentence = sentences[0][0].split()
sentence = [infer.vocab.word_to_id(i) for i in sentence]
context = tf.train.Features()
feature_lists = tf.train.FeatureLists(
feature_list={'sentence': _int64_feature_list(sentence)})
sequence_example = tf.train.SequenceExample(context=context,
feature_lists=feature_lists)
return sequence_example.SerializeToString()