V_ft = vfeat_entry['vfeat'].value
pad_V_ft = np.zeros([max_box_num - num_box, V_ft.shape[1]],
dtype=V_ft.dtype)
V_ft = np.concatenate([V_ft, pad_V_ft], axis=0)
q_intseq = entry['question_intseq'].value
q_intseq_len = np.array(len(q_intseq), dtype=np.int32)
q_intseq_pad = np.zeros([max_q_len - q_intseq_len], dtype=q_intseq.dtype)
q_intseq = np.concatenate([q_intseq, q_intseq_pad], axis=0)
answer_id = np.array(entry['answer_id'].value, dtype=np.int32)
tf_example = tf.train.Example(features=tf.train.Features(
feature={
'qid': tf_util.int64_feature(int(id)),
'image_id': tf_util.bytes_feature(str(image_id)),
'image/encoded': tf_util.bytes_feature(image_str),
'image/format': tf_util.bytes_feature('png'),
'image/height': tf_util.int64_feature(IMAGE_HEIGHT),
'image/width': tf_util.int64_feature(IMAGE_WIDTH),
'box/list': tf_util.float_feature(list(box.reshape([-1]))),
'box/shape': tf_util.int64_feature(list(box.shape)),
'num_box': tf_util.int64_feature(num_box),
'V_ft/list': tf_util.float_feature(list(V_ft.reshape([-1]))),
'V_ft/shape': tf_util.int64_feature(list(V_ft.shape)),
'q_intseq/list': tf_util.int64_feature(list(q_intseq)),
'q_intseq/len': tf_util.int64_feature(q_intseq_len),
'answer_id': tf_util.int64_feature(answer_id),
}))
tf_record_writer.write(tf_example.SerializeToString())
V_ft = vfeat_entry['vfeat'].value
pad_V_ft = np.zeros([max_box_num - num_box, V_ft.shape[1]],
dtype=V_ft.dtype)
V_ft = np.concatenate([V_ft, pad_V_ft], axis=0)
q_intseq = entry['question_intseq'].value
q_intseq_len = np.array(len(q_intseq), dtype=np.int32)
q_intseq_pad = np.zeros([max_q_len - q_intseq_len], dtype=q_intseq.dtype)
q_intseq = np.concatenate([q_intseq, q_intseq_pad], axis=0)
answer_id = np.array(entry['answer_id'].value, dtype=np.int32)
tf_example = tf.train.Example(features=tf.train.Features(
feature={
'qid': tf_util.int64_feature(int(id)),
'image_id': tf_util.bytes_feature(str(image_id)),
'box/list': tf_util.float_feature(list(box.reshape([-1]))),
'box/shape': tf_util.int64_feature(list(box.shape)),
'num_box': tf_util.int64_feature(num_box),
'V_ft/list': tf_util.float_feature(list(V_ft.reshape([-1]))),
'V_ft/shape': tf_util.int64_feature(list(V_ft.shape)),
'q_intseq/list': tf_util.int64_feature(list(q_intseq)),
'q_intseq/len': tf_util.int64_feature(q_intseq_len),
'answer_id': tf_util.int64_feature(answer_id),
}))
tf_record_writer.write(tf_example.SerializeToString())
data.close()
vfeat.close()
tqdm(entries, desc='write tfrecord {}'.format(split))):
if i % config.num_record_per_shard == 0:
shard_id = int(i / config.num_record_per_shard)
shard_name = '{}-{:05d}-of-{:05d}'.format(split, shard_id,
num_shards)
shard_path = os.path.join(config.save_dir, split, shard_name)
if os.path.exists(shard_path):
raise ValueError('Existing shard path: {}'.format(shard_path))
tf_record_writer = tf.python_io.TFRecordWriter(shard_path)
max_q_len = max(max_q_len, len(entry['q_token']))
max_num_answer = max(max_num_answer, len(entry['answer']['labels']))
tf_example = tf.train.Example(features=tf.train.Features(
feature={
'qid': tf_util.int64_feature(int(entry['question_id'])),
'image_id': tf_util.bytes_feature(str(entry['image_id'])),
'image_idx': tf_util.int64_feature(int(entry['image'])),
'q_intseq/list': tf_util.int64_feature(entry['q_token']),
'q_intseq/len': tf_util.int64_feature(len(entry['q_token'])),
'answers/ids': tf_util.int64_feature(entry['answer']
['labels']),
'answers/scores': tf_util.float_feature(entry['answer']
['scores']),
}))
tf_record_writer.write(tf_example.SerializeToString())
# Construct image_info
used_image_paths = set()
image_id2idx = {}
image_path2idx = {}
image_num2path = {}