Function encode_data

lib/evaluation.py:72–126

(model, data_loader, log_step=10, logging=logger.info)

def encode_data(model, data_loader, log_step=10, logging=logger.info):

    batch_time = AverageMeter()
    val_logger = LogCollector()

    # switch to evaluate mode
    model.eval()

    end = time.time()

    # np array to keep all the embeddings
    img_embs = None
    cap_embs = None
    
    # compute the number of max word
    max_n_word = model.opt.max_word

    for i, data_i in enumerate(data_loader):
        
        # make sure val logger is used       
        images, captions, lengths, ids, img_ids = data_i

        model.logger = val_logger

        # compute the embeddings
        img_emb, cap_emb, lengths = model.forward_emb(images, captions, lengths)

        if img_embs is None:
            # for local visual features
            img_embs = torch.zeros((len(data_loader.dataset), img_emb.size(1), img_emb.size(2)))
            # for local textual features
            cap_embs = torch.zeros((len(data_loader.dataset), max_n_word, cap_emb.size(2)))
            
            cap_lens = torch.zeros(len(data_loader.dataset)).long()

        # cache embeddings
        img_embs[ids] = img_emb.cpu()

        n_word = min(max(lengths), max_n_word)
        
        cap_embs[ids, :n_word, :] = cap_emb[:, :n_word, :].cpu()
        cap_lens[ids] = lengths.cpu()

        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()

        if i % log_step == 0:
            logging('Test: [{0}/{1}]\t'
                    '{e_log}\t'
                    'Batch-Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
                .format(i, len(data_loader.dataset) // data_loader.batch_size + 1, batch_time=batch_time, e_log=str(model.logger)))
        del images, captions

    return img_embs, cap_embs, cap_lens


def evalrank(model_path, model=None, data_path=None, split='dev', fold5=False, save_path=None):