3.6节train_ch3函数loss计算和sgd问题

[jhj0411jhj]

bug描述 3.6节train_ch3函数，如果传入的loss已经是求过平均的，train_l_sum每次只累加一个batch的平均值，最后却除以总样本数，打印的loss结果就会很小，例如3.9节的调用。

如果每个batch不一样大（例如Fashion-MNIST设置batch_size=256时，最后一个batch是96），当optimizer=None时，默认的sgd传入batch_size应该会在最后一个batch造成误差，似乎应该使用y.shape[0]。

另外train_ch5是用batch_count，如果batch大小不一致，最后打印的loss也应该会有微小误差。

def train_ch3(net, train_iter, test_iter, loss, num_epochs, batch_size,
              params=None, lr=None, optimizer=None):
    for epoch in range(num_epochs):
        train_l_sum, train_acc_sum, n = 0.0, 0.0, 0
        for X, y in train_iter:
            y_hat = net(X)
            l = loss(y_hat, y).sum()

            # 梯度清零
            if optimizer is not None:
                optimizer.zero_grad()
            elif params is not None and params[0].grad is not None:
                for param in params:
                    param.grad.data.zero_()

            l.backward()
            if optimizer is None:
                sgd(params, lr, batch_size)
            else:
                optimizer.step()  # “softmax回归的简洁实现”一节将用到

            train_l_sum += l.item()
            train_acc_sum += (y_hat.argmax(dim=1) == y).sum().item()
            n += y.shape[0]
        test_acc = evaluate_accuracy(test_iter, net)
        print('epoch %d, loss %.4f, train acc %.3f, test acc %.3f'
              % (epoch + 1, train_l_sum / n, train_acc_sum / n, test_acc))

版本信息 pytorch: torchvision: torchtext: ...

[CKing111]

你好，我在这个部分出现了下面的error，请问应该怎么修正，有点无从下手

---

RuntimeError Traceback (most recent call last)

in 31 % (epoch + 1, train_l_sum / n, train_acc_sum / n, test_acc)) 32 ---> 33 train_ch3(net, train_iter, test_iter, cross_entropy, num_epochs, batch_size, [W, b], lr) in train_ch3(net, train_iter, test_iter, loss, num_epochs, batch_size, params, lr, optimizer) 7 train_l_sum, train_acc_sum, n = 0.0, 0.0, 0 8 for X, y in train_iter: ----> 9 y_hat = net(X) 10 l = loss(y_hat, y).sum() 11 in net(X) 1 def net(X): ----> 2 return softmax(torch.mm(X.view((-1, num_inputs)), W) + b) RuntimeError: The size of tensor a (10) must match the size of tensor b (3) at non-singleton dimension 1