def evaluate()

in train_retriever.py [0:0]

• 35 lines of code
• 8 McCabe index (conditional complexity)

def evaluate(model, dataset, collator, opt):
    sampler = SequentialSampler(dataset)
    dataloader = DataLoader(
        dataset, 
        sampler=sampler, 
        batch_size=opt.per_gpu_batch_size,
        drop_last=False, 
        num_workers=10, 
        collate_fn=collator
    )
    model.eval()
    if hasattr(model, "module"):
        model = model.module
    total = 0
    eval_loss = []

    avg_topk = {k:[] for k in [1, 2, 5] if k <= opt.n_context}
    idx_topk = {k:[] for k in [1, 2, 5] if k <= opt.n_context}
    inversions = []
    with torch.no_grad():
        for i, batch in enumerate(dataloader):
            (idx, question_ids, question_mask, context_ids, context_mask, gold_score) = batch

            _, _, scores, loss = model(
                question_ids=question_ids.cuda(),
                question_mask=question_mask.cuda(),
                passage_ids=context_ids.cuda(),
                passage_mask=context_mask.cuda(),
                gold_score=gold_score.cuda(),
            )

            src.evaluation.eval_batch(scores, inversions, avg_topk, idx_topk)
            total += question_ids.size(0)

    inversions = src.util.weighted_average(np.mean(inversions), total, opt)[0]
    for k in avg_topk:
        avg_topk[k] = src.util.weighted_average(np.mean(avg_topk[k]), total, opt)[0]
        idx_topk[k] = src.util.weighted_average(np.mean(idx_topk[k]), total, opt)[0]

    return loss, inversions, avg_topk, idx_topk