DevNet: no class docstring, and five constructor parameters (including random_seed) are never used #714

DevNet has no class docstring, and five of its ten constructor parameters are stored but never read. One of them is random_seed, so runs are not reproducible.

Evidence

DevNet.__doc__ is None — the class goes straight from class DevNet(BaseDetector): at line 209 to def __init__ at line 210. It is the only detector in pyod/models/ with no class docstring, so it renders as a bare signature on the API page and help(DevNet) shows nothing.

Of the ten constructor parameters, these five are assigned to self and then never referenced again anywhere in the module:

data_format is worse than unused. A module-level data_format = 0 exists at line 23, and inside the training path line 190 re-assigns a local data_format = 0 immediately before the branch that reads it:

# Assuming data_format variable should be defined somewhere in the context or as a parameter
data_format = 0  # Assuming it's set correctly according to your use-case

if data_format == 0:

So self.data_format is shadowed by a hardcoded local, and passing data_format=1 has no effect. The two "Assuming ..." comments suggest this was left unfinished.

Why it matters

random_seed is the sharpest one: a user who sets it expects reproducible results and does not get them, with nothing to indicate why. known_outliers and cont_rate are core DevNet concepts from the paper (it is a weakly-supervised method that trains against a small set of labelled anomalies), so a user tuning them and seeing no change has reason to doubt the implementation rather than their own usage.

Suggested resolution

1. Wire up random_seed (seed torch, numpy, and Python random at the start of fit) — this one is worth doing regardless of what happens to the others.
2. Either implement nb_batch, known_outliers, cont_rate, and data_format, or remove them from the signature so the API stops promising behavior that is not there.
3. Delete the shadowing local at line 190 and the module-level data_format at line 23 once the parameter is either honoured or removed.
4. Add a class docstring in the numpydoc style used by the other detectors, documenting whatever parameter set survives.

A recent documentation pass deliberately skipped this class rather than writing prose describing parameters that do nothing, since every resolution here changes behavior. Found while auditing constructor signatures against their docstrings.