pytorch-cpn的refine_loss

refine_output和 target7 的shape均为[batch,17,64,48]，经过计算每个元素对应位置的loss后，得到[batch,17,64,48]的refine_loss，然后经过refine_loss.mean(dim=3).mean(dim=2)，得到[batch,17]的loss。
MSELoss的说明如下

    target7 = torch.zeros((1,num_class, output_shape[0], output_shape[1]))
    refine_output = torch.ones((1,num_class, output_shape[0], output_shape[1]))
    criterion2 = torch.nn.MSELoss(reduce=False)
    refine_loss = criterion2(target7, refine_output)
    print(refine_loss.shape)
    refine_loss = refine_loss.mean(dim=3).mean(dim=2)
    print(refine_loss.shape)