토론: Reproducibility of a machine-learning experiment: seeds, environment, data and the limits of determinism

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

번역이 없어 원문을 표시합니다. 원문

The concrete settings behind step 2, all from the cited PyTorch note: `torch.use_deterministic_algorithms(True)` makes non-deterministic operations raise or switch implementations, but on CUDA 10.2 and later it additionally requires `CUBLAS_WORKSPACE_CONFIG=:4096:8` (or `:16:8`) in the environment before the process starts, and `torch.backends.cudnn.benchmark` must stay `False` because the autotuner may pick different kernels per run. `DataLoader` worker processes need `worker_init_fn` (or a `generator` passed to the loader) so that augmentation randomness is seeded per worker, otherwise workers duplicate or vary seeds between runs. Outside the frameworks, `PYTHONHASHSEED` fixes string hashing and therefore the iteration order of sets and of dictionaries built from them, which silently reorders features in code that collects column names in a set.

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

번역이 없어 원문을 표시합니다. 원문

Step 1 recommends declaring one random state and passing it to every estimator and splitter, but it does not say whether that state is an integer or a `RandomState` instance, and the cited pitfalls page makes the two behave differently in a way that matters for the article's goal. With an integer, every call to `fit` starts from the same generator state, so in cross-validation each fold's estimator draws the same random subsets; the page says this can make a score look good or bad because of one lucky seed and recommends passing an instance so that each fold uses a different RNG, which gives a more representative estimate of the estimator's own variance. With an instance, the sequence is consumed across calls, so results depend on the exact order of every randomised operation in the program: adding a `fit` earlier in the script changes every later result, which is the opposite of what a reproducibility protocol wants. Neither choice is wrong, but the protocol should choose: integers for the run-to-run repeatability it is about (accepting that the CV estimate reflects one seed, which step 7's seed interval already repairs), instances when the purpose is a robust model comparison. As written, two readers following step 1 build incomparable pipelines.