Function validate

train.py:241–310

(opt, val_loader, model)

def validate(opt, val_loader, model):

    logger = logging.getLogger(__name__)
    
    model.eval()

    with torch.no_grad():
       img_embs, cap_embs, cap_lens = encode_data(model, val_loader, opt.log_step, logging.info)

    # have repetitive image features
    img_embs = img_embs[::5]

    start_time = time.time()

    if opt.multi_gpu:         
        sims = torch.zeros((len(img_embs), len(cap_embs))).cuda()
        
        num_tasks = utils.get_world_size()
        rank = utils.get_rank() 

        step = img_embs.size(0) // num_tasks + 1
        start = rank * step
        end = min(img_embs.size(0), start + step)

        sims_part = shard_attn_scores(model, img_embs[start:end], cap_embs, cap_lens, opt, gpu=True)
        sims[start:end] = sims_part

        # wait for synchronization 
        torch.distributed.barrier()
        # Aggregating results on different GPUs
        torch.distributed.all_reduce(sims, op=torch.distributed.ReduceOp.SUM) 
        sims = sims.cpu().numpy()
    else:
        sims = shard_attn_scores(model, img_embs, cap_embs, cap_lens, opt)      
        sims = sims.numpy()

    # compute metric
    if utils.is_main_process():
        
        logging.info("calculate similarity time: %.3f" % float(time.time() - start_time))

        npts = img_embs.shape[0]
        # print(npts)
        
        # caption retrieval
        (r1, r5, r10, medr, meanr) = i2t(npts, sims)
        logging.info("Image to text (R@1, R@5, R@10): %.1f, %.1f, %.1f" % (r1, r5, r10))

        # image retrieval
        (r1i, r5i, r10i, medri, meanr) = t2i(npts, sims)
        logging.info("Text to image (R@1, R@5, R@10): %.1f, %.1f, %.1f" % (r1i, r5i, r10i))

        # sum of recalls to be used for early stopping
        currscore = r1 + r5 + r10 + r1i + r5i + r10i
        logger.info('Current rsum is {}'.format(round(currscore, 1)))
            
        # record metrics in tensorboard
        tb_logger.log_value('r1', r1, step=model.Eiters)