"""
Probabilistic LVQ models: SLVQ and RSLVQ.
Soft LVQ (SLVQ) and Robust Soft LVQ (RSLVQ) use Gaussian mixture
models to define class-conditional probabilities and optimize
likelihood-based objectives.
References
----------
.. [1] Seo, S., & Obermayer, K. (2003). Soft Learning Vector
Quantization. Neural Computation.
.. [2] Seo, S., & Obermayer, K. (2007). Soft Nearest Prototype
Classification. IEEE Trans. Neural Networks.
"""
import jax
import jax.numpy as jnp
import numpy as np
from prosemble.models.prototype_base import SupervisedPrototypeModel
from prosemble.core.losses import nllr_loss, rslvq_loss
[docs]
class SLVQ(SupervisedPrototypeModel):
"""Soft Learning Vector Quantization.
Uses Gaussian mixture probabilities:
.. math::
p(k|x) = \\frac{\\exp(-d^2 / 2\\sigma^2)}{\\sum_j \\exp(-d_j^2 / 2\\sigma^2)}
.. math::
P(\\text{class}|x) = \\sum_{k \\in \\text{class}} p(k|x)
Loss: :math:`-\\log(P(\\text{correct}) / P(\\text{wrong}))`
Parameters
----------
sigma : float
Bandwidth of Gaussian mixture.
rejection_confidence : float, optional
Minimum class probability for a confident prediction (0 to 1).
Samples below this threshold are rejected (labeled -1) when using
``predict_with_rejection()``. Default is None (no rejection).
n_prototypes_per_class : int
Number of prototypes per class.
max_iter : int
Maximum training iterations.
lr : float
Learning rate.
epsilon : float
Convergence threshold on loss change.
random_seed : int
Random seed for reproducibility.
distance_fn : callable, optional
Distance function (default: squared Euclidean).
optimizer : str or optax optimizer, optional
Optimizer name ('adam', 'sgd') or optax GradientTransformation.
Default: 'adam'.
transfer_fn : callable, optional
Transfer function for loss shaping (default: identity).
margin : float
Margin for loss computation.
callbacks : list, optional
List of Callback objects.
use_scan : bool
If True (default), use jax.lax.scan for training (faster, JIT-compiled,
but runs all max_iter iterations even after convergence).
If False, use a Python for-loop with true early stopping (no wasted
compute after convergence, but slower per iteration).
batch_size : int, optional
Mini-batch size. If None (default), use full-batch training.
When set, each epoch iterates over shuffled mini-batches of this size.
lr_scheduler : str or optax.Schedule, optional
Learning rate schedule. Supported strings: 'exponential_decay',
'cosine_decay', 'warmup_cosine_decay', 'warmup_exponential_decay',
'warmup_constant', 'polynomial', 'linear', 'piecewise_constant',
'sgdr'. Or pass a custom optax.Schedule. Default: None.
lr_scheduler_kwargs : dict, optional
Keyword arguments passed to the learning rate scheduler
(e.g. ``decay_rate``, ``transition_steps``). Default: None.
prototypes_initializer : str or callable, optional
How to initialize prototypes. Supported strings: 'stratified_random'
(default), 'class_mean', 'class_conditional_mean', 'stratified_noise',
'random_normal', 'uniform', 'zeros', 'ones', 'fill_value'.
Or pass a callable ``(X, y, n_per_class, key) -> (protos, labels)``.
patience : int, optional
Number of consecutive epochs with no improvement before stopping.
If None (default), stops after a single non-improving step (epsilon
check). Requires use_scan=False for true early stopping.
restore_best : bool
If True, restore the parameters that achieved the lowest loss
(or validation loss if validation data is provided). Default: False.
class_weight : dict or 'balanced', optional
Weights for each class. Dict maps class label to weight, e.g.
{0: 1.0, 1: 2.0, 2: 1.5}. 'balanced' auto-computes weights
inversely proportional to class frequencies. Default: None (uniform).
gradient_accumulation_steps : int, optional
Accumulate gradients over this many steps before applying an update.
Effective batch size = batch_size * gradient_accumulation_steps.
Default: None (no accumulation).
ema_decay : float, optional
Exponential moving average decay for parameters (0 < ema_decay < 1).
After training, model parameters are replaced with EMA-smoothed values.
Typical values: 0.999, 0.9999. Default: None (no EMA).
freeze_params : list of str, optional
List of parameter group names to freeze (zero gradients).
E.g. ['backbone'] to freeze the backbone and only train prototypes.
Default: None (all parameters trainable).
lookahead : dict, optional
Enable lookahead optimizer wrapper. Dict with keys:
- 'sync_period': int (default 6) — sync every k steps
- 'slow_step_size': float (default 0.5) — interpolation factor
Default: None (no lookahead).
mixed_precision : str or None, optional
Compute dtype for mixed precision training. 'float16' or 'bfloat16'.
Master weights stay in float32; forward/backward pass runs in lower
precision for ~2x speed and ~half memory on GPU. Float16 uses static
loss scaling to prevent gradient underflow. Default: None (disabled).
"""
def __init__(self, sigma=1.0, rejection_confidence=None,
n_prototypes_per_class=1, max_iter=100, lr=0.01,
epsilon=1e-6, random_seed=42, distance_fn=None,
optimizer='adam', transfer_fn=None, margin=0.0,
callbacks=None, use_scan=True, batch_size=None,
lr_scheduler=None, lr_scheduler_kwargs=None,
prototypes_initializer=None, patience=None,
restore_best=False, class_weight=None,
gradient_accumulation_steps=None, ema_decay=None,
freeze_params=None, lookahead=None, mixed_precision=None,
**kwargs):
super().__init__(
n_prototypes_per_class=n_prototypes_per_class,
max_iter=max_iter, lr=lr, epsilon=epsilon,
random_seed=random_seed, distance_fn=distance_fn,
optimizer=optimizer, transfer_fn=transfer_fn, margin=margin,
callbacks=callbacks, use_scan=use_scan, batch_size=batch_size,
lr_scheduler=lr_scheduler,
lr_scheduler_kwargs=lr_scheduler_kwargs,
prototypes_initializer=prototypes_initializer,
patience=patience, restore_best=restore_best,
class_weight=class_weight,
gradient_accumulation_steps=gradient_accumulation_steps,
ema_decay=ema_decay, freeze_params=freeze_params,
lookahead=lookahead, mixed_precision=mixed_precision,
**kwargs,
)
self.sigma = sigma
self.rejection_confidence = rejection_confidence
def _compute_loss(self, params, X, y, proto_labels):
distances = self.distance_fn(X, params['prototypes'])
return nllr_loss(distances, y, proto_labels, sigma=self.sigma)
def predict_with_rejection(self, X, confidence=None):
"""Predict with rejection option.
Samples whose maximum class probability is below the confidence
threshold are assigned label -1 (rejected / "I don't know").
Parameters
----------
X : array-like of shape (n_samples, n_features)
confidence : float, optional
Override the model's rejection_confidence for this call.
Returns
-------
labels : array of shape (n_samples,)
Predicted labels, or -1 for rejected samples.
"""
self._check_fitted()
threshold = confidence if confidence is not None else self.rejection_confidence
if threshold is None:
return self.predict(X)
X = jnp.asarray(X, dtype=jnp.float32)
proba = self.predict_proba(X)
max_proba = jnp.max(proba, axis=1)
preds = jnp.argmax(proba, axis=1)
return jnp.where(max_proba >= threshold, preds, -1)
def _get_hyperparams(self):
hp = super()._get_hyperparams()
hp['sigma'] = self.sigma
hp['rejection_confidence'] = self.rejection_confidence
return hp
[docs]
class RSLVQ(SupervisedPrototypeModel):
"""Robust Soft Learning Vector Quantization.
Like SLVQ but with a more robust denominator:
.. math::
\\text{loss} = -\\log\\frac{P(\\text{correct}|x)}{P(\\text{all}|x)}
Parameters
----------
sigma : float
Bandwidth of Gaussian mixture.
rejection_confidence : float, optional
Minimum class probability for a confident prediction (0 to 1).
Samples below this threshold are rejected (labeled -1) when using
``predict_with_rejection()``. Default is None (no rejection).
n_prototypes_per_class : int
Number of prototypes per class.
max_iter : int
Maximum training iterations.
lr : float
Learning rate.
epsilon : float
Convergence threshold on loss change.
random_seed : int
Random seed for reproducibility.
distance_fn : callable, optional
Distance function (default: squared Euclidean).
optimizer : str or optax optimizer, optional
Optimizer name ('adam', 'sgd') or optax GradientTransformation.
Default: 'adam'.
transfer_fn : callable, optional
Transfer function for loss shaping (default: identity).
margin : float
Margin for loss computation.
callbacks : list, optional
List of Callback objects.
use_scan : bool
If True (default), use jax.lax.scan for training (faster, JIT-compiled,
but runs all max_iter iterations even after convergence).
If False, use a Python for-loop with true early stopping (no wasted
compute after convergence, but slower per iteration).
batch_size : int, optional
Mini-batch size. If None (default), use full-batch training.
When set, each epoch iterates over shuffled mini-batches of this size.
lr_scheduler : str or optax.Schedule, optional
Learning rate schedule. Supported strings: 'exponential_decay',
'cosine_decay', 'warmup_cosine_decay', 'warmup_exponential_decay',
'warmup_constant', 'polynomial', 'linear', 'piecewise_constant',
'sgdr'. Or pass a custom optax.Schedule. Default: None.
lr_scheduler_kwargs : dict, optional
Keyword arguments passed to the learning rate scheduler
(e.g. ``decay_rate``, ``transition_steps``). Default: None.
prototypes_initializer : str or callable, optional
How to initialize prototypes. Supported strings: 'stratified_random'
(default), 'class_mean', 'class_conditional_mean', 'stratified_noise',
'random_normal', 'uniform', 'zeros', 'ones', 'fill_value'.
Or pass a callable ``(X, y, n_per_class, key) -> (protos, labels)``.
patience : int, optional
Number of consecutive epochs with no improvement before stopping.
If None (default), stops after a single non-improving step (epsilon
check). Requires use_scan=False for true early stopping.
restore_best : bool
If True, restore the parameters that achieved the lowest loss
(or validation loss if validation data is provided). Default: False.
class_weight : dict or 'balanced', optional
Weights for each class. Dict maps class label to weight, e.g.
{0: 1.0, 1: 2.0, 2: 1.5}. 'balanced' auto-computes weights
inversely proportional to class frequencies. Default: None (uniform).
gradient_accumulation_steps : int, optional
Accumulate gradients over this many steps before applying an update.
Effective batch size = batch_size * gradient_accumulation_steps.
Default: None (no accumulation).
ema_decay : float, optional
Exponential moving average decay for parameters (0 < ema_decay < 1).
After training, model parameters are replaced with EMA-smoothed values.
Typical values: 0.999, 0.9999. Default: None (no EMA).
freeze_params : list of str, optional
List of parameter group names to freeze (zero gradients).
E.g. ['backbone'] to freeze the backbone and only train prototypes.
Default: None (all parameters trainable).
lookahead : dict, optional
Enable lookahead optimizer wrapper. Dict with keys:
- 'sync_period': int (default 6) — sync every k steps
- 'slow_step_size': float (default 0.5) — interpolation factor
Default: None (no lookahead).
mixed_precision : str or None, optional
Compute dtype for mixed precision training. 'float16' or 'bfloat16'.
Master weights stay in float32; forward/backward pass runs in lower
precision for ~2x speed and ~half memory on GPU. Float16 uses static
loss scaling to prevent gradient underflow. Default: None (disabled).
"""
def __init__(self, sigma=1.0, rejection_confidence=None,
n_prototypes_per_class=1, max_iter=100, lr=0.01,
epsilon=1e-6, random_seed=42, distance_fn=None,
optimizer='adam', transfer_fn=None, margin=0.0,
callbacks=None, use_scan=True, batch_size=None,
lr_scheduler=None, lr_scheduler_kwargs=None,
prototypes_initializer=None, patience=None,
restore_best=False, class_weight=None,
gradient_accumulation_steps=None, ema_decay=None,
freeze_params=None, lookahead=None, mixed_precision=None,
**kwargs):
super().__init__(
n_prototypes_per_class=n_prototypes_per_class,
max_iter=max_iter, lr=lr, epsilon=epsilon,
random_seed=random_seed, distance_fn=distance_fn,
optimizer=optimizer, transfer_fn=transfer_fn, margin=margin,
callbacks=callbacks, use_scan=use_scan, batch_size=batch_size,
lr_scheduler=lr_scheduler,
lr_scheduler_kwargs=lr_scheduler_kwargs,
prototypes_initializer=prototypes_initializer,
patience=patience, restore_best=restore_best,
class_weight=class_weight,
gradient_accumulation_steps=gradient_accumulation_steps,
ema_decay=ema_decay, freeze_params=freeze_params,
lookahead=lookahead, mixed_precision=mixed_precision,
**kwargs,
)
self.sigma = sigma
self.rejection_confidence = rejection_confidence
def _compute_loss(self, params, X, y, proto_labels):
distances = self.distance_fn(X, params['prototypes'])
return rslvq_loss(distances, y, proto_labels, sigma=self.sigma)
def predict_with_rejection(self, X, confidence=None):
"""Predict with rejection option.
Samples whose maximum class probability is below the confidence
threshold are assigned label -1 (rejected / "I don't know").
Parameters
----------
X : array-like of shape (n_samples, n_features)
confidence : float, optional
Override the model's rejection_confidence for this call.
Returns
-------
labels : array of shape (n_samples,)
Predicted labels, or -1 for rejected samples.
"""
self._check_fitted()
threshold = confidence if confidence is not None else self.rejection_confidence
if threshold is None:
return self.predict(X)
X = jnp.asarray(X, dtype=jnp.float32)
proba = self.predict_proba(X)
max_proba = jnp.max(proba, axis=1)
preds = jnp.argmax(proba, axis=1)
return jnp.where(max_proba >= threshold, preds, -1)
def _get_hyperparams(self):
hp = super()._get_hyperparams()
hp['sigma'] = self.sigma
hp['rejection_confidence'] = self.rejection_confidence
return hp