d2l-ai/d2l-zh · error · AssertionError

test_acc <= 1 and test_acc > 0.7

Error message

test_acc <= 1 and test_acc > 0.7

What it means

An AssertionError in d2l.tensorflow.train_ch3 that final test accuracy must be in (0.7, 1] for the Fashion-MNIST softmax-regression benchmark. It guards the evaluation half of the pipeline: evaluate_accuracy must run against the real test_iter, and the net must generalize past 70%. A value > 1 indicates the metric denominator is wrong; <= 0.7 indicates undertraining, divergence, or evaluating an untrained/different model.

Source

Thrown at d2l/tensorflow.py:321

            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:
        break

View on GitHub (pinned to e6b18ccea7)

Solutions

1. Make sure evaluate_accuracy(net, test_iter) is called on the same net object that was trained, after the training loop.
2. Confirm loss matches output form: use from_logits=True only if the net returns logits (no softmax layer).
3. Rebuild train_iter/test_iter with identical preprocessing (resize to 28x28, scale to [0,1]) via d2l.load_data_fashion_mnist.
4. If accuracy is near 0.1 (chance), apply the fixes for the loss assert first (lr, updater scaling) — accuracy follows the loss.

Example fix

# before: net re-created before eval
net = tf.keras.Model(...)  # wipes trained weights
train_ch3(net, ...)  # test_acc <= 0.7 -> AssertionError
# after: train once, evaluate the same net
train_ch3(net, train_iter, test_iter, loss, 10, updater)

test_acc = evaluate_accuracy(net, test_iter)
if not (0.7 < test_acc <= 1.0):
    raise ValueError(f'test_acc out of expected (0.7, 1]: {test_acc:.3f}; '
                     f'verify same net is evaluated and preprocessing matches train')

try:
    train_ch3(net, train_iter, test_iter, loss, num_epochs, updater)
except AssertionError as e:
    raise RuntimeError(f'train_ch3 test-accuracy check failed ({e}); '
                       f'check test_iter construction and from_logits setting') from e

When it happens

Trigger: Passing an empty or wrong test_iter (accuracy 0/0 or NaN); evaluating a freshly re-initialized net instead of the trained one; label/shape mismatch making every prediction wrong; divergence from too-high lr dropping test accuracy to chance level (~0.1).

Common situations: Notebook cell reordering that re-creates net after training; using a test_iter built with a different preprocessing than train_iter; TF version differences in SparseCategoricalCrossentropy from_logits handling producing near-zero accuracy.

Related errors

• train_loss < 0.5
• train_acc <= 1 and train_acc > 0.7
• test_acc <= 1 and test_acc > 0.7
• train_loss < 0.5
• 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/98853a7fe8050007. Report an issue: GitHub.