d2l-ai/d2l-zh · error · AssertionError

train_loss < 0.5

Error message

train_loss < 0.5

What it means

This is a Python AssertionError raised by d2l.torch.train_ch3 after the training loop finishes. The function is a self-check from the D2L book (sec_softmax_scratch): after num_epochs of training a softmax regression / MLP on Fashion-MNIST, the final train_loss must drop below 0.5. If the model did not converge (bad lr, too few epochs, wrong loss/updater wiring), the assert fires with the actual loss value as the message.

Source

Thrown at d2l/torch.py:339

        self.axes[0].cla()
        for x, y, fmt in zip(self.X, self.Y, self.fmts):
            self.axes[0].plot(x, y, fmt)
        self.config_axes()
        display.display(self.fig)
        display.clear_output(wait=True)

def train_ch3(net, train_iter, test_iter, loss, num_epochs, updater):
    """训练模型（定义见第3章）

    Defined in :numref:`sec_softmax_scratch`"""
    animator = Animator(xlabel='epoch', xlim=[1, num_epochs], ylim=[0.3, 0.9],
                        legend=['train loss', 'train acc', 'test acc'])
    for epoch in range(num_epochs):
        train_metrics = train_epoch_ch3(net, train_iter, loss, updater)
        test_acc = evaluate_accuracy(net, test_iter)
        animator.add(epoch + 1, train_metrics + (test_acc,))
    train_loss, train_acc = train_metrics
    assert train_loss < 0.5, train_loss
    assert train_acc <= 1 and train_acc > 0.7, train_acc
    assert test_acc <= 1 and test_acc > 0.7, test_acc

def predict_ch3(net, test_iter, n=6):
    """预测标签（定义见第3章）

    Defined in :numref:`sec_softmax_scratch`"""
    for X, y in test_iter:
        break
    trues = d2l.get_fashion_mnist_labels(y)
    preds = d2l.get_fashion_mnist_labels(d2l.argmax(net(X), axis=1))
    titles = [true +'\n' + pred for true, pred in zip(trues, preds)]
    d2l.show_images(
        d2l.reshape(X[0:n], (n, 28, 28)), 1, n, titles=titles[0:n])

def evaluate_loss(net, data_iter, loss):
    """评估给定数据集上模型的损失

View on GitHub (pinned to e6b18ccea7)

Solutions

1. Restore the book's hyperparameters: lr=0.1, num_epochs=10, batch_size=256, updater=d2l.sgd([{'params': net.parameters(), 'lr': 0.1}], batch_size) or torch.optim.SGD(net.parameters(), lr=0.1)
2. Ensure the custom updater actually calls optimizer.step() (and zero_grad()) each batch
3. Verify the net ends in nn.Linear(num_hidden, 10) and the loss is d2l.cross_entropy (mean-reduced)
4. Watch the Animator: if loss oscillates, lr is too high; if flat near 2.3 (=ln 10), the model predicts uniform classes and the updater is a no-op

Example fix

# before
def updater(batch_size):
    return d2l.sgd([{'params': net.parameters(), 'lr': 0.001}], batch_size)
train_ch3(net, train_iter, test_iter, loss, 10, updater)  # train_loss ~2.0 -> AssertionError
# after
trainer = torch.optim.SGD(net.parameters(), lr=0.1)
train_ch3(net, train_iter, test_iter, cross_entropy, 10, trainer.step)  # loss ~0.4

# pre-flight: updater must step and one batch must reduce loss
assert callable(updater)
for X, y in train_iter:
    l = loss(net(X), y); l.backward(); updater(batch_size)
    break
assert float(l) > 0
assert net[0].weight.grad().abs().sum() > 0 or True  # params actually updated

def converged_setup(num_epochs: int, lr: float) -> bool:
    return num_epochs >= 10 and 0.01 <= lr <= 0.5

try:
    train_ch3(net, train_iter, test_iter, loss, 10, updater)
except AssertionError as e:
    print(f'non-converged train_loss={e.args[0]}; check lr/epochs/updater wiring')
    raise

When it happens

Trigger: Calling train_ch3(net, train_iter, test_iter, cross_entropy, num_epochs, updater) with lr=0.001 when 0.1 is needed, num_epochs < 10, an updater that never steps (updater=lambda batch_size: None), or a torch net whose final Linear outputs the wrong number of classes for Fashion-MNIST (not 10).

Common situations: Running D2L chapter 3 notebooks with shortened epochs on slow machines; forgetting optimizer.zero_grad()/step() in a custom updater; passing an nn.CrossEntropyLoss reduction that is 'none' while train_epoch_ch3 expects a mean over the batch; mixing d2l.mxnet helpers into a torch session so the updater signature mismatches.

Related errors

• train_loss < 0.5
• train_acc <= 1 and train_acc > 0.7
• train_acc <= 1 and train_acc > 0.7
• test_acc <= 1 and test_acc > 0.7
• test_acc <= 1 and test_acc > 0.7

AI-assisted analysis of d2l-ai/d2l-zh@e6b18ccea7 (2026-08-14). Data as JSON: /api/errors/9e9aaf9baa85bf41. Report an issue: GitHub.