fix read_batch behavior for visgeno testing

[qleroy]

Testing scripts in exp-visgeno-rel

The DataReader allows to read the same batch several times even at testing time, if a batch raises an IOError in run_prefetch from util/visgeno_rel_train/rel_data_reader.py, it does not increment self.n_batch causing the reader to eventually read the first batch once again without incrementing self.n_epoch.

For example in exp-visgeno-rel/exp_test_visgeno_attbilstm.py, the test runs over reader.num_batch (equals 5000 in case of imdb_tst) but 127 batches of imdb_tst does not contain any relationship causing prepare_batch to raise an IOError, caught by run_prefetch and it does not do anything in that case. Eventually the first 127 batches of imdb_tst are read and been computed scores once again. I think this is not an intended behavior, rather exp_test_visgeno_attbilstm.py (and other testing scripts) should compute scores against batches containing relationships only once.

To solve the problem, I suggest that run_prefetch adds None to the prefetch_queue in case of an "empty" batch.

# from run_prefetch in util/visgeno_rel_train/rel_data_reader.py
except IOError as err:
            # Print error and move on to next batch
            print('data reader: skipped an image.', err)
            prefetch_queue.put(None, block=True)

In testing scripts such as exp_test_visgeno_attbilstm.py if read_batch is None n_iter is incremented and we move to the next batch.

# from exp_test_visgeno_attbilstm.py in exp-visgeno-rel/exp_test_visgeno_attbilstm.py
for n_iter in range(reader.num_batch):
    batch = reader.read_batch()

    ###
    # continue if the batch does not contain any relationship
    # and increment n_iter so that the first few batches are not read once again in case of
    ###
    if batch is None:
        n_iter = n_iter + 1
        continue

    # compute score for batch