class AttentionVQARewards(object):
def __init__(self, ckpt_file='model/v1_vqa_VQA/v1_vqa_VQA_best2/model.ckpt-135000',
use_dis_reward=False):
self.g = tf.Graph()
self.ckpt_file = ckpt_file
from models.vqa_soft_attention import AttentionModel
from vqa_config import ModelConfig
config = ModelConfig()
self.ans2id = AnswerTokenToTopAnswer()
self.use_dis_reward = use_dis_reward
with self.g.as_default():
self.sess = tf.Session()
self.model = AttentionModel(config, phase='test_broadcast')
self.model.build()
vars = tf.trainable_variables()
self.saver = tf.train.Saver(var_list=vars)
self.saver.restore(self.sess, ckpt_file)
def process_answers(self, ans, ans_len):
ans_pathes = _parse_gt_questions(ans, ans_len)
return self.ans2id.get_top_answer(ans_pathes)
def get_reward(self, sampled, inputs):
if len(inputs) == 4:
images, res5c, ans, ans_len = inputs
top_ans_ids = self.process_answers(ans, ans_len)
else:
assert (len(inputs) == 5)
images, res5c, ans, ans_len, top_ans_ids = inputs
images_aug = []
top_ans_ids_aug = []
answer_aug = []
answer_len_aug = []
pathes = []
for _idx, ps in enumerate(sampled):
for p in ps:
if p[-1] == END_TOKEN:
pathes.append(p[1:-1]) # remove start end token
else:
pathes.append(p[1:]) # remove start end token
images_aug.append(images[_idx][np.newaxis, :])
answer_aug.append(ans[_idx][np.newaxis, :])
answer_len_aug.append(ans_len[_idx])
top_ans_ids_aug.append(top_ans_ids[_idx])
# put to arrays
arr, arr_len = put_to_array(pathes)
images_aug = np.concatenate(images_aug)
answer_aug = np.concatenate(answer_aug).astype(np.int32)
top_ans_ids_aug = np.array(top_ans_ids_aug)
answer_len_aug = np.array(answer_len_aug, dtype=np.int32)
# run inference in VQA
scores = self.model.inference(self.sess, [res5c, arr, arr_len])
if self.use_dis_reward:
vqa_scores = np.require(scores.argmax(axis=1) == top_ans_ids_aug,
np.float32)
else:
_this_batch_size = scores.shape[0]
vqa_scores = scores[np.arange(_this_batch_size), top_ans_ids_aug]
is_valid = top_ans_ids_aug != 2000
return vqa_scores, [images_aug, answer_aug, answer_len_aug, is_valid]
class MLBWrapper(object):
def __init__(
self,
ckpt_file='model/v1_vqa_VQA/v1_vqa_VQA_best2/model.ckpt-135000'):
self.g = tf.Graph()
self.ckpt_file = ckpt_file
from models.vqa_soft_attention import AttentionModel
from vqa_config import ModelConfig
config = ModelConfig()
self.name = ' ------- MLB-attention ------- '
with self.g.as_default():
self.sess = tf.Session()
self.model = AttentionModel(config, phase='test_broadcast')
self.model.build()
vars = tf.trainable_variables()
self.saver = tf.train.Saver(var_list=vars)
self.saver.restore(self.sess, ckpt_file)
def _load_image(self, image_id):
FEAT_ROOT = '/usr/data/fl302/data/VQA/ResNet152/resnet_res5c'
filename = '%s2014/COCO_%s2014_%012d.jpg' % ('val', 'val', image_id)
f = np.load(os.path.join(FEAT_ROOT, filename + '.npz'))['x']
return f.transpose((1, 2, 0))[np.newaxis, ::]
def get_scores(self, sampled, image_id, top_ans_id):
# process questions
pathes = []
for p in sampled:
if p[-1] == END_TOKEN:
pathes.append(p[1:-1]) # remove start end token
else:
pathes.append(p[1:]) # remove start end token
num_unk = len(sampled)
arr, arr_len = put_to_array(pathes)
# load image
res5c = self._load_image(image_id)
images_aug = np.tile(res5c, [num_unk, 1, 1, 1])
# inference
scores = self.model.inference(self.sess, [images_aug, arr, arr_len])
vqa_scores = scores[:, top_ans_id].flatten()
return vqa_scores