Quiz połączony z przewodnikiem · Średni Poziom

Testowanie jednostkowe Quiz dotyczący kodu uczenia maszynowego

Wybierz odpowiednie testy dla transformacji, strat, kontraktów tensorowych i niewielkich zachowań związanych z nadmiernym dopasowaniem, nie myląc kontroli kodu z uogólnianiem.

Powiązane ścieżki prowadząceKod uczenia maszynowego do testowania jednostkowego
Pytanie 1 z 8

Which assertions are appropriate for a preprocessing transformer?

Podgląd każdego pytania8 pytania

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

  1. Which assertions are appropriate for a preprocessing transformer?
    • Expected output shape, column order and known transformed values
    • Held-out accuracy on a deployment-like population
    • Recall across rare production subgroups
    • End-to-end latency under peak service load
  2. Why compare a loss implementation with a hand calculation?
    • It proves the model will generalize
    • It can catch formula or reduction errors on a known example
    • It replaces training data quality checks
    • It guarantees the optimizer converges
  3. What does a tiny-overfit test help detect?
    • Fairness across all subgroups
    • Reliable performance on unseen data
    • Potential training-path or gradient bugs
    • Production latency under load
  4. How should tests compare floating-point outputs?
    • Skip checks for all model outputs
    • Require exact bitwise equality on every device
    • Round every value to an integer
    • Use explicit tolerances appropriate to numerical precision
  5. Which check belongs in a model wrapper interface test?
    • A statistical calibration curve on a held-out set
    • One prediction per input row and clear rejection of malformed features
    • Subgroup performance estimates with confidence intervals
    • Training loss after a full retraining job
  6. Why separate unit tests from model-quality evaluation?
    • Quality metrics test only syntax
    • Unit tests cannot execute Python functions
    • Code correctness does not establish statistical usefulness
    • Model validation needs no data
  7. What can a finite-gradient assertion establish?
    • The model is calibrated
    • The gradients are mathematically optimal
    • The model has no bias
    • A basic gradient path exists for expected parameters
  8. Where should slow full-dataset training usually be placed?
    • In a separate integration or training validation stage
    • In every tiny unit test
    • Inside a feature-name assertion
    • In a package lockfile