def create_generate_model(opts):
    assert isinstance(opts, options)
    with tf.variable_scope(tf.get_variable_scope(), reuse=True):
        model = inference_wrapper(opts, reuse=True)
        vocab = vocabulary(opts)
        generator = CaptionGenerator(model, vocab, beam_size=4)

    return generator
    'the place of h5 file that contains image attributes information')
tf.flags.DEFINE_string('train_dir', './model',
                       'the directory for training the model')
tf.flags.DEFINE_string('caption_file', 'result/caption_',
                       'the file which save results')
tf.flags.DEFINE_string('id', 'test', 'the id to specified')

opts = options()
opts.batch_size = 1
opts.data_h5 = FLAGS.data_h5
opts.data_json = FLAGS.data_json
opts.attributes_h5 = FLAGS.attributes_h5
opts.train_dir = FLAGS.train_dir

dataloader = DataLoader(opts, 'test')
model = inference_wrapper(opts)

vocab = vocabulary(opts)
generator = CaptionGenerator(model, vocab, beam_size=4)
nImage = len(dataloader.test_ix)

sess = tf.Session()
checkpoint = tf.train.latest_checkpoint(opts.train_dir)
ema = tf.train.ExponentialMovingAverage(decay=0.999)
variables_to_restore = ema.variables_to_restore()
saver = tf.train.Saver(var_list=variables_to_restore, max_to_keep=5)
saver.restore(sess, checkpoint)
print('restoring from checkpoint {:s}'.format(checkpoint))
res = []
for i in xrange(nImage):
Ejemplo n.º 3
0
def create_val_fn():

    FLAGS = namedtuple(
        'FLAGS',
        ['data_h5', 'data_json', 'attributes_h5', 'train_dir', 'caption_file'])
    FLAGS.data_h5 = './data/data.h5'
    FLAGS.data_json = './data/data.json'
    FLAGS.attributes_h5 = './data/tag_feats.h5'
    FLAGS.train_dir = './model'
    FLAGS.caption_file = 'result/caption.json'

    opts = options()
    opts.batch_size = 1
    opts.data_h5 = FLAGS.data_h5
    opts.data_json = FLAGS.data_json
    opts.attributes_h5 = FLAGS.attributes_h5
    opts.train_dir = FLAGS.train_dir

    dataloader = DataLoader(opts, 'val')
    # reuse variables
    with tf.variable_scope(tf.get_variable_scope(), reuse=True):
        model = inference_wrapper(opts, reuse=True)
        vocab = vocabulary(opts)
        generator = CaptionGenerator(model, vocab, beam_size=4)

    nImage = len(dataloader.val_ix)

    def val(sess):
        res = []
        t1 = 0
        for i in xrange(nImage):

            attributes, features, image_feature, image_ids, img = dataloader.get_batch(
                1)
            caption = generator.beam_search(sess, image_feature, attributes,
                                            features)
            caption = caption[0].sentence
            caption = [vocab.to_word(w) for w in caption]
            caption = ' '.join(caption)
            res.append({'image_id': image_ids[0], 'caption': caption})
            print('processing image:{:d}/{:d},time {:f}s'.format(
                i, nImage,
                time.time() - t1))
            t1 = time.time()

        current_dir = os.getcwd()
        caption_file_full_path = os.path.join(current_dir, FLAGS.caption_file)
        with open(caption_file_full_path, 'w') as f:
            json.dump(res, f)

        #evaluation.myeval(caption_file_full_path)
        os.system('./eval.sh {:s}'.format(caption_file_full_path))

        result_file = caption_file_full_path + '.json_out.json'
        with open(result_file) as f:
            result = json.load(f)

        for metric, val in result.iteritems():
            print('{:s}:{:f}'.format(metric, val))

        # return the cider metric
        return result['CIDEr']

    return val