d2l-ai/d2l-zh · error · AssertionError

train_acc <= 1 and train_acc > 0.7

Error message

train_acc <= 1 and train_acc > 0.7

What it means

An AssertionError in d2l.tensorflow.train_ch3 requiring final training accuracy to lie in (0.7, 1]. It is a post-training sanity check that the softmax-regression model actually learned Fashion-MNIST to a reasonable degree. Values <= 0.7 mean the model undertrained or diverged; values > 1 mean the metric computation itself is broken.

Source

Thrown at d2l/tensorflow.py:320

        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

class Updater():
    """用小批量随机梯度下降法更新参数

    Defined in :numref:`sec_softmax_scratch`"""
    def __init__(self, params, lr):
        self.params = params
        self.lr = lr

    def __call__(self, batch_size, grads):
        d2l.sgd(self.params, grads, self.lr, batch_size)

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:

View on GitHub (pinned to e6b18ccea7)

Solutions

1. Restore the book's hyperparameters (lr=0.1, batch_size=256, num_epochs=10) and rerun.
2. Check train_epoch_ch3's accuracy accumulator: metric.add(y == argmax(y_hat), y.numel()) style bookkeeping must sum counts and divide by totals exactly once.
3. Confirm updater.apply_gradients is actually called each batch; a broken update loop yields ~0.1 accuracy (chance level).
4. Read the Animator curves: if accuracy is still climbing at the final epoch, increase num_epochs.

Example fix

# before
train_ch3(net, train_iter, test_iter, loss, 1, updater)  # undertrained -> acc <= 0.7
# after
train_ch3(net, train_iter, test_iter, loss, 10, updater)

final_loss, final_acc = train_epoch_ch3(net, train_iter, loss, updater)
if not (0.7 < final_acc <= 1.0):
    raise ValueError(f'train_acc out of expected (0.7, 1]: {final_acc:.3f}; '
                     f'check metric accumulator and optimizer wiring')

try:
    train_ch3(net, train_iter, test_iter, loss, num_epochs, updater)
except AssertionError as e:
    raise RuntimeError(f'train_ch3 accuracy check failed ({e}); '
                       f'likely undertrained or broken updater') from e

When it happens

Trigger: Same divergence/undertraining causes as the loss assert: too-high lr, wrong updater scaling, too few epochs, untrained net; additionally a custom train_epoch_ch3 whose accuracy accumulation divides by the wrong denominator can push train_acc above 1 and trip the upper bound.

Common situations: Notebook users changing num_epochs from 10 to 1-2; using a custom metric accumulator with a mismatched count; training on a heavily subsampled train_iter; TF2 gradient-tape code that forgets to apply updates.

Related errors

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

AI-assisted analysis of d2l-ai/d2l-zh@e6b18ccea7 (2026-08-14). Data as JSON: /api/errors/53ba4edc3be185f3. Report an issue: GitHub.