在validation阶段为什么不使用model.eval()

[Ponytai1j]

大大好看见您最近上线了想问下这个问题。实在没搞明白，既然模型使用了dropout，BN，为什么只在推理阶段使用model.eval(),在validation阶段为什么不开model.eval()

def val(net, data_loader, device='cpu', supervision='full'):

accuracy, total = 0., 0.
ignored_labels = data_loader.dataset.ignored_labels
for batch_idx, (data, target) in enumerate(data_loader):
    with torch.no_grad():
        # Load the data into the GPU if required
        data, target = data.to(device), target.to(device)
        if supervision == 'full':
            output = net(data)
        elif supervision == 'semi':
            outs = net(data)
            output, rec = outs
        _, output = torch.max(output, dim=1)
        #target = target - 1
        for pred, out in zip(output.view(-1), target.view(-1)):
            if out.item() in ignored_labels:
                continue
            else:
                accuracy += out.item() == pred.item()
                total += 1
return accuracy / total`

[eecn]

是需要加上的^_^.