Documents GMNet in docs/source/manuals/methods.md, mirroring the existing
HistNetQ entry: paper reference, torch/geotorch requirement, a basic usage
example, and a second example showing multi-branch + CKA regularization.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
quapy/classification/__init__.py imported a labelshift module that doesn't exist
anywhere in the repo (and nothing else references it), breaking `import quapy`
entirely; the import is removed.
quapy/method/aggregative.py imported a nonexistent `_liep` module instead of the
actual file, _liep_draft.py, which itself never got a LEIP class added (only
helper functions such as leip()). Guards the import so LEIP degrades to an
"not available" placeholder, like the other optional neural methods, instead of
crashing the whole package. test_leip/test_leip_fixed_tau still fail as a result;
finishing LEIP is left for a follow-up.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Ports GMNet (from https://github.com/pglez84/gmnet) into quapy/method/_gmnet.py,
mirroring how HistNetQ was ported: dropping that repo's quantificationlib-backed
bag generators in favor of QuaPy's own sampling protocols, and adding geotorch
(now a 'neural' extra dependency) to keep the Gaussian layers' covariance matrices
positive-definite during training.
- GMNet represents each bag instance by its likelihood under one or more learned
mixtures of Gaussians ("GM branches"), mean-pools these representations over the
bag, and predicts prevalence from the result. Supports multiple stacked GM
branches with an optional CKA-regularization term encouraging their latent
representations to be dissimilar.
- Fixes two aspects of the original architecture that assumed a fixed, training-time
bag_size baked into the network (a reshape step, and forward-hook-based activation
capture for CKA): both are now computed from the actual input shape/plain
attributes at forward time, so the model also works on predict()'s arbitrary-sized
test samples, not just same-size bags.
- Factors the bag-based training loop shared by HistNetQ and GMNet (bag generation,
fit/fit_from_samples, early stopping, LR scheduling, checkpointing, predict) out of
_histnet.py into a new BagTrainedQuantifier base class in
quapy/method/_neural_bags.py; HistNetQ's public API and behavior are unchanged.
- Aliased in meta.py (torch/geotorch-optional, mirroring HistNetQ/QuaNet) and
registered in META_METHODS.
- Adds test_gmnet covering single-branch and multi-branch+CKA (via
fit_from_samples/mix_bags) variants.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Ports the "hard" histogram variant of HistNetQ (from
https://github.com/pglez84/histnetq) into quapy/method/_histnet.py, dropping
that repo's quantificationlib-backed bag generators in favor of QuaPy's own
sampling protocols (UPP by default). Implemented as a BaseQuantifier,
alongside QuaNet, since it trains end-to-end on samples of known prevalence
rather than following the classify-then-aggregate pattern.
- HistNetQ.fit(X, y): resamples training/validation bags from a
LabelledCollection via a configurable protocol (UPP by default; fresh
random bags each training epoch, a fixed reproducible sequence for
validation).
- HistNetQ.fit_from_samples(protocol, val_protocol=None, mix_bags=False):
trains directly from a protocol that already yields bags (e.g. LeQua's
SamplesFromDir), with an optional mixer to synthesize extra
intermediate-prevalence bags from the given ones.
- Aliased in meta.py (torch-optional, mirroring the existing QuaNet guard)
and registered in META_METHODS.
- Adds test_histnetq covering both entry points on binary and multiclass
synthetic data.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Not having jax/numpyro/pystan installed is the expected, common case
for a plain `pip install quapy`; warning about it every import just
nags users who never asked for the bayes extra. Kept at debug level
so it's still available to diagnose a broken/partial bayes install.
fit_tr_val's train_test_split had no random_state, so val_split=float
picked a different split every run; occasionally the split produced a
posterior distribution that made abstention's temperature-scaling
L-BFGS optimizer diverge to NaN. Threaded random_state through the
base class and all four calibrator subclasses, and pinned the test to
a seed confirmed stable across repeated runs.
fit_tr_val's train_test_split had no random_state, so val_split=float
picked a different split every run; occasionally the split produced a
posterior distribution that made abstention's temperature-scaling
L-BFGS optimizer diverge to NaN. Threaded random_state through the
base class and all four calibrator subclasses, and pinned the test to
a seed confirmed stable across repeated runs.
Same fix as master (63075e4): conf.py uses pydata_sphinx_theme and
sphinx_design, but the docs extra never listed them, so pip install
.[docs] couldn't build the docs.
conf.py switched to the pydata_sphinx_theme and added the sphinx_design
extension, but the docs extra in setup.py was never updated to match,
so a fresh `pip install .[docs]` (as CI does) couldn't build the docs.
stan.plugins.get_plugins() calls pkg_resources.iter_entry_points()
from several call sites during model building, each re-emitting
setuptools' deprecation notice since Python's default warning dedup
is keyed per call site, not per message. Not actionable upstream
noise, so filter it by message instead.
setuptools 82 removed pkg_resources entirely, which httpstan still
imports at runtime; CI's prior `pip install --upgrade setuptools`
step grabs the latest, so an unpinned setuptools didn't help.
- Add smoke tests covering data/reader.py, method/_threshold_optim.py,
classification/calibration.py, method/confidence.py, and the pure-numpy
helpers in method/_bayesian.py (skipping the jax/stan-dependent model
code itself, consistent with how the aggregative-method registry already
treats it as optional)
- SVMperf: stop creating a tempfile.TemporaryDirectory() just to discard it
immediately for its .name; generate the path directly instead of doing a
pointless create/delete/recreate cycle (cleanup already happens via the
class's own __del__)
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Correctness:
- OneVsAllAggregative.aggregation_fit: fix undefined variable and call to
the nonexistent aggregate_fit (should be aggregation_fit)
- solve_adjustment: stop mutating the caller's fitted arrays in place for
method='invariant-ratio'
- LabelledCollection.join(): fix classes always being None due to
ndarray.sort() returning None; now unions each collection's own classes_
so a class absent from a particular join is kept at zero prevalence
- Rename the duplicate newSVMKLD (nkld variant) to newSVMNKLD so both loss
variants are reachable
- EMQ/DyS/DMy: resolve n_jobs via qp._get_njobs() like the other methods,
so qp.environ['N_JOBS'] is respected
- AggregativeMedianEstimator: drop backend='threading' (global np.random
state mutated via temp_seed is not thread-safe); use the safe process
based default instead
- NeuralClassifier: default device now 'cpu', matching its own docstring
- ConfidenceEllipseSimplex: narrow bare except to np.linalg.LinAlgError
- SVMperf: stop merging stderr into stdout so failures report the actual
subprocess error instead of crashing with AttributeError
- ConfidenceRegionABC: replace @lru_cache on bound methods (leaked every
instance for the process lifetime) with per-instance caching
Style/quality:
- Replace print() with warnings.warn()/logging across aggregative.py,
base.py, meta.py, model_selection.py, classification/neural.py,
method/_neural.py, classification/svmperf.py, data/reader.py,
data/datasets.py; also fixes a `raise RuntimeWarning(...)` in EMQ that
would have crashed instead of warning
- Remove dead duplicate class MedianEstimator2 in meta.py
- Rename misleading _compute_tpr(TP, FP) parameter to FN, matching what
callers actually pass
- Replace argparse.ArgumentError misuse with ValueError
- Remove commented-out dead code in protocol.py
- _lequa.py: fix CSV-parse failure raising an unrelated UnboundLocalError
instead of a clear ValueError
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Alejandro Moreo Fernandez
Alejandro Moreo
Andrea Esuli