Skip to content

uqregressors.conformal.k_fold_cqr

This class implements conformal quantile regression in a K-fold manner to obtain uncertainty estimates which are often conservative, but use the entire dataset available. This can result in large improvements over split conformal quantile regression, particularly in cases where the dataset is sparse.

Tip

The quantiles of the underlying quantile regressor can be tuned with the parameters tau_lo and tau_hi as in the paper. This can often result in more efficient intervals.

Note

K-fold Conformal Quantile Regression can be overly conservative in prediction intervals, particularly in sparse data settings or when the underlying estimator has high variance.

K-Fold-CQR

This module implements conformal quantile regression in a K-fold manner for regression of a one dimensional output.

Key features are
  • Customizable neural network architecture
  • Tunable quantiles of the underyling regressors
  • Prediction intervals without distributional assumptions
  • Parallel training of ensemble models with Joblib
  • Customizable optimizer and loss function
  • Optional Input/Output Normalization

KFoldCQR

Bases: BaseEstimator, RegressorMixin

K-Fold Conformalized Quantile Regressor for uncertainty estimation in regression tasks.

This class trains an ensemble of quantile neural networks using K-Fold cross-validation, and applies conformal prediction to calibrate prediction intervals.

Parameters:

Name Type Description Default
name str

Name of the model.

'K_Fold_CQR_Regressor'
n_estimators int

Number of K-Fold models to train.

5
hidden_sizes list

Sizes of the hidden layers for each quantile regressor.

[64, 64]
dropout float or None

Dropout rate for the neural network layers.

None
alpha float

Miscoverage rate (1 - confidence level).

0.1
requires_grad bool

Whether inputs should require gradient.

False
tau_lo float

Lower quantile, defaults to alpha/2.

None
tau_hi float

Upper quantile, defaults to 1 - alpha/2.

None
n_jobs int

Number of parallel jobs for training.

1
activation_str str

String identifier of the activation function.

'ReLU'
learning_rate float

Learning rate for training.

0.001
epochs int

Number of training epochs.

200
batch_size int

Batch size for training.

32
optimizer_cls type

Optimizer class.

Adam
optimizer_kwargs dict

Keyword arguments for optimizer.

None
scheduler_cls type or None

Learning rate scheduler class.

None
scheduler_kwargs dict

Keyword arguments for scheduler.

None
loss_fn callable or None

Loss function, defaults to quantile loss.

None
device str

Device to use for training and inference.

'cpu'
use_wandb bool

Whether to log training with Weights & Biases.

False
wandb_project str or None

wandb project name.

None
wandb_run_name str or None

wandb run name.

None
scale_data bool

Whether to normalize input/output data.

True
input_scaler TorchStandardScaler

Scaler for input features.

None
output_scaler TorchStandardScaler

Scaler for target outputs.

None
random_seed int or None

Random seed for reproducibility.

None
tuning_loggers list

Optional list of loggers for tuning.

[]

Attributes:

Name Type Description
quantiles Tensor

The lower and upper quantiles for prediction.

models list[QuantNN]

A list of the models in the ensemble.

residuals Tensor

The combined residuals on the calibration sets.

conformal_width Tensor

The width needed to conformalize the quantile regressor, q.

_loggers list[Logger]

Training loggers for each ensemble member.

Source code in uqregressors\conformal\k_fold_cqr.py
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
class KFoldCQR(BaseEstimator, RegressorMixin): 
    """
    K-Fold Conformalized Quantile Regressor for uncertainty estimation in regression tasks.

    This class trains an ensemble of quantile neural networks using K-Fold cross-validation,
    and applies conformal prediction to calibrate prediction intervals.

    Args:
        name (str): Name of the model.
        n_estimators (int): Number of K-Fold models to train.
        hidden_sizes (list): Sizes of the hidden layers for each quantile regressor.
        dropout (float or None): Dropout rate for the neural network layers.
        alpha (float): Miscoverage rate (1 - confidence level).
        requires_grad (bool): Whether inputs should require gradient.
        tau_lo (float): Lower quantile, defaults to alpha/2.
        tau_hi (float): Upper quantile, defaults to 1 - alpha/2.
        n_jobs (int): Number of parallel jobs for training.
        activation_str (str): String identifier of the activation function.
        learning_rate (float): Learning rate for training.
        epochs (int): Number of training epochs.
        batch_size (int): Batch size for training.
        optimizer_cls (type): Optimizer class.
        optimizer_kwargs (dict): Keyword arguments for optimizer.
        scheduler_cls (type or None): Learning rate scheduler class.
        scheduler_kwargs (dict): Keyword arguments for scheduler.
        loss_fn (callable or None): Loss function, defaults to quantile loss.
        device (str): Device to use for training and inference.
        use_wandb (bool): Whether to log training with Weights & Biases.
        wandb_project (str or None): wandb project name.
        wandb_run_name (str or None): wandb run name.
        scale_data (bool): Whether to normalize input/output data.
        input_scaler (TorchStandardScaler): Scaler for input features.
        output_scaler (TorchStandardScaler): Scaler for target outputs.
        random_seed (int or None): Random seed for reproducibility.
        tuning_loggers (list): Optional list of loggers for tuning.

    Attributes: 
        quantiles (Tensor): The lower and upper quantiles for prediction.
        models (list[QuantNN]): A list of the models in the ensemble.
        residuals (Tensor): The combined residuals on the calibration sets. 
        conformal_width (Tensor): The width needed to conformalize the quantile regressor, q. 
        _loggers (list[Logger]): Training loggers for each ensemble member. 
    """
    def __init__(
            self, 
            name="K_Fold_CQR_Regressor",
            n_estimators=5,
            hidden_sizes=[64, 64], 
            dropout = None,
            alpha=0.1, 
            requires_grad=False,
            tau_lo = None, 
            tau_hi = None, 
            n_jobs=1, 
            activation_str="ReLU",
            learning_rate=1e-3,
            epochs=200,
            batch_size=32,
            optimizer_cls=torch.optim.Adam,
            optimizer_kwargs=None,
            scheduler_cls=None,
            scheduler_kwargs=None,
            loss_fn=None,
            device="cpu",
            use_wandb=False,
            wandb_project=None,
            wandb_run_name=None,
            scale_data = True, 
            input_scaler = None, 
            output_scaler = None,
            random_seed=None, 
            tuning_loggers = []
    ):
        self.name = name
        self.n_estimators = n_estimators
        self.hidden_sizes = hidden_sizes
        self.dropout = dropout
        self.alpha = alpha
        self.requires_grad = requires_grad
        self.tau_lo = tau_lo or alpha / 2 
        self.tau_hi = tau_hi or 1 - alpha / 2
        self.activation_str = activation_str
        self.learning_rate = learning_rate
        self.epochs = epochs
        self.batch_size = batch_size
        self.optimizer_cls = optimizer_cls
        self.optimizer_kwargs = optimizer_kwargs or {}
        self.scheduler_cls = scheduler_cls
        self.scheduler_kwargs = scheduler_kwargs or {}
        self.loss_fn = loss_fn or self.quantile_loss
        self.device = device

        self.use_wandb = use_wandb
        self.wandb_project = wandb_project
        self.wandb_run_name = wandb_run_name

        self.n_jobs = n_jobs
        self.random_seed = random_seed
        self.quantiles = torch.tensor([self.tau_lo, self.tau_hi], device=self.device)
        self.models = []
        self.residuals = []
        self.conformal_width = None
        self.input_dim = None
        if self.n_estimators == 1: 
            raise ValueError("n_estimators set to 1. To use a single Quantile Regressor, use a non-ensembled Quantile Regressor class")
        self.scale_data = scale_data 
        self.input_scaler = input_scaler or TorchStandardScaler() 
        self.output_scaler = output_scaler or TorchStandardScaler()

        self._loggers = []
        self.training_logs = None
        self.tuning_loggers = tuning_loggers
        self.tuning_logs = None


    def quantile_loss(self, preds, y): 
        """
        Quantile loss used for training the quantile regressors.

        Args:
            preds (Tensor): Predicted quantiles, shape (batch_size, 2).
            y (Tensor): True target values, shape (batch_size,).

        Returns:
            (Tensor): Scalar loss.
        """
        error = y.view(-1, 1) - preds
        return torch.mean(torch.max(self.quantiles * error, (self.quantiles - 1) * error))

    def _train_single_model(self, X_tensor, y_tensor, input_dim, train_idx, cal_idx, model_idx): 
        if self.random_seed is not None: 
            torch.manual_seed(self.random_seed + model_idx)
            np.random.seed(self.random_seed + model_idx)

        activation = get_activation(self.activation_str)
        model = QuantNN(input_dim, self.hidden_sizes, self.dropout, activation).to(self.device)

        optimizer = self.optimizer_cls(
            model.parameters(), lr=self.learning_rate, **self.optimizer_kwargs
        )
        scheduler = None 
        if self.scheduler_cls: 
            scheduler = self.scheduler_cls(optimizer, **self.scheduler_kwargs)

        X_train = X_tensor.detach()[train_idx]
        y_train = y_tensor.detach()[train_idx]
        dataset = TensorDataset(X_train, y_train)
        dataloader = DataLoader(dataset, batch_size=self.batch_size, shuffle=True)

        logger = Logger(
            use_wandb=self.use_wandb,
            project_name=self.wandb_project,
            run_name=self.wandb_run_name + str(model_idx) if self.wandb_run_name is not None else None,
            config={"n_estimators": self.n_estimators, "learning_rate": self.learning_rate, "epochs": self.epochs},
            name=f"Estimator-{model_idx}"
        )

        model.train()
        for epoch in range(self.epochs): 
            model.train()
            epoch_loss = 0.0 
            for xb, yb in dataloader: 
                optimizer.zero_grad() 
                preds = model(xb)
                loss = self.loss_fn(preds, yb)
                loss.backward() 
                optimizer.step() 
                epoch_loss += loss 

            if epoch % (self.epochs / 20) == 0:
                logger.log({"epoch": epoch, "train_loss": epoch_loss})

            if scheduler: 
                scheduler.step()


        test_X = X_tensor[cal_idx]
        test_y = y_tensor[cal_idx]
        oof_preds = model(test_X)
        loss_matrix =(oof_preds - test_y) * torch.tensor([1.0, -1.0], device=self.device)
        residuals = torch.max(loss_matrix, dim=1).values
        logger.finish()
        return model, residuals, logger

    def fit(self, X, y): 
        """
        Fit the ensemble on training data.

        Args:
            X (array-like or torch.Tensor): Training inputs.
            y (array-like or torch.Tensor): Training targets.

        Returns:
            (KFoldCQR): Fitted estimator.
        """
        X_tensor, y_tensor = validate_and_prepare_inputs(X, y, device=self.device, requires_grad=self.requires_grad)
        input_dim = X_tensor.shape[1]
        self.input_dim = input_dim


        if self.scale_data:
            X_tensor = self.input_scaler.fit_transform(X_tensor)
            y_tensor = self.output_scaler.fit_transform(y_tensor)

        kf = TorchKFold(n_splits=self.n_estimators, shuffle=True)

        results = Parallel(n_jobs=self.n_jobs)(
            delayed(self._train_single_model)(X_tensor, y_tensor, input_dim, train_idx, cal_idx, i)
            for i, (train_idx, cal_idx) in enumerate(kf.split(X_tensor))
        )

        self.models = [result[0] for result in results]
        self.residuals = torch.cat([result[1] for result in results], dim=0).ravel()
        self._loggers = [result[2] for result in results]

        return self

    def predict(self, X): 
        """
        Predicts the target values with uncertainty estimates.

        Args:
            X (np.ndarray): Feature matrix of shape (n_samples, n_features).

        Returns:
            (Union[Tuple[np.ndarray, np.ndarray, np.ndarray], Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]): Tuple containing:
                mean predictions,
                lower bound of the prediction interval,
                upper bound of the prediction interval.

        !!! note
            If `requires_grad` is False, all returned arrays are NumPy arrays.
            Otherwise, they are PyTorch tensors with gradients.
        """
        X_tensor = validate_X_input(X, input_dim=self.input_dim, device=self.device, requires_grad=self.requires_grad)
        n = len(self.residuals)
        q = int((1 - self.alpha) * (n + 1))
        q = min(q, n-1)

        res_quantile = n-q

        self.conformal_width = torch.topk(self.residuals, res_quantile).values[-1]

        if self.scale_data: 
            X_tensor = self.input_scaler.transform(X_tensor)

        preds = [] 

        with torch.no_grad(): 
            for model in self.models: 
                model.eval()
                pred = model(X_tensor)
                preds.append(pred)

        preds = torch.stack(preds)

        means = torch.mean(preds, dim=2) 
        mean = torch.mean(means, dim=0)

        lower_cq = torch.mean(preds[:, :, 0], dim=0)
        upper_cq = torch.mean(preds[:, :, 1], dim=0)

        lower = lower_cq - self.conformal_width
        upper = upper_cq + self.conformal_width

        if self.scale_data: 
            mean = self.output_scaler.inverse_transform(mean.view(-1, 1)).squeeze()
            lower = self.output_scaler.inverse_transform(lower.view(-1, 1)).squeeze()
            upper = self.output_scaler.inverse_transform(upper.view(-1, 1)).squeeze()

        if not self.requires_grad: 
            return mean.detach().cpu().numpy(), lower.detach().cpu().numpy(), upper.detach().cpu().numpy()

        else: 
            return mean, lower, upper

    def save(self, path):
        """
        Save the trained model and associated configuration to disk.

        Args:
            path (str or Path): Directory to save model files.
        """
        path = Path(path)
        path.mkdir(parents=True, exist_ok=True)

        # Save config (exclude non-serializable or large objects)
        config = {
            k: v for k, v in self.__dict__.items()
            if k not in ["models", "quantiles", "residuals", "conformal_width", "optimizer_cls", "optimizer_kwargs", "scheduler_cls", "scheduler_kwargs", 
                         "input_scaler", "output_scaler", "_loggers", "training_logs", "tuning_loggers", "tuning_logs"]
            and not callable(v)
            and not isinstance(v, (torch.nn.Module,))
        }

        config["optimizer"] = self.optimizer_cls.__class__.__name__ if self.optimizer_cls is not None else None
        config["scheduler"] = self.scheduler_cls.__class__.__name__ if self.scheduler_cls is not None else None
        config["input_scaler"] = self.input_scaler.__class__.__name__ if self.input_scaler is not None else None 
        config["output_scaler"] = self.output_scaler.__class__.__name__ if self.output_scaler is not None else None

        with open(path / "config.json", "w") as f:
            json.dump(config, f, indent=4)

        # Save model weights
        for i, model in enumerate(self.models):
            torch.save(model.state_dict(), path / f"model_{i}.pt")

        # Save residuals and conformity score
        torch.save({
            "conformal_width": self.conformal_width, 
            "residuals": self.residuals,
            "quantiles": self.quantiles,
        }, path / "extras.pt")

        with open(path / "extras.pkl", 'wb') as f: 
            pickle.dump([self.optimizer_cls, 
                        self.optimizer_kwargs, self.scheduler_cls, self.scheduler_kwargs, self.input_scaler, self.output_scaler], f)

        for i, logger in enumerate(getattr(self, "_loggers", [])):
            logger.save_to_file(path, idx=i, name="estimator")

        for i, logger in enumerate(getattr(self, "tuning_loggers", [])): 
            logger.save_to_file(path, name="tuning", idx=i)

    @classmethod
    def load(cls, path, device="cpu", load_logs=False):
        """
        Load a saved KFoldCQR model from disk.

        Args:
            path (str or Path): Directory containing saved model files.
            device (str): Device to load the model on ("cpu" or "cuda").
            load_logs (bool): Whether to also load training logs.

        Returns:
            (KFoldCQR): The loaded model instance.
        """
        path = Path(path)

        # Load config
        with open(path / "config.json", "r") as f:
            config = json.load(f)
        config["device"] = device

        config.pop("optimizer", None)
        config.pop("scheduler", None)
        config.pop("input_scaler", None)
        config.pop("output_scaler", None)

        input_dim = config.pop("input_dim", None)
        model = cls(**config)

        # Recreate models
        model.input_dim = input_dim
        activation = get_activation(config["activation_str"])
        model.models = []
        for i in range(config["n_estimators"]):
            m = QuantNN(model.input_dim, config["hidden_sizes"], config["dropout"], activation).to(device)
            m.load_state_dict(torch.load(path / f"model_{i}.pt", map_location=device))
            model.models.append(m)

        # Load extras
        extras_path = path / "extras.pt"
        if extras_path.exists():
            extras = torch.load(extras_path, map_location=device, weights_only=False)
            model.conformal_width = extras.get("conformal_width", None)
            model.residuals = extras.get("residuals", None)
            model.quantiles = extras.get("quantiles", None)
        else:
            model.conformal_width = None
            model.residuals = None
            model.quantiles = None

        with open(path / "extras.pkl", 'rb') as f: 
            optimizer_cls, optimizer_kwargs, scheduler_cls, scheduler_kwargs, input_scaler, output_scaler = pickle.load(f)

        model.optimizer_cls = optimizer_cls 
        model.optimizer_kwargs = optimizer_kwargs 
        model.scheduler_cls = scheduler_cls 
        model.scheduler_kwargs = scheduler_kwargs
        model.input_scaler = input_scaler
        model.output_scaler = output_scaler

        if load_logs: 
            logs_path = path / "logs"
            training_logs = [] 
            tuning_logs = []
            if logs_path.exists() and logs_path.is_dir(): 
                estimator_log_files = sorted(logs_path.glob("estimator_*.log"))
                for log_file in estimator_log_files:
                    with open(log_file, "r", encoding="utf-8") as f:
                        training_logs.append(f.read())

                tuning_log_files = sorted(logs_path.glob("tuning_*.log"))
                for log_file in tuning_log_files: 
                    with open(log_file, "r", encoding="utf-8") as f: 
                        tuning_logs.append(f.read())

            model.training_logs = training_logs
            model.tuning_logs = tuning_logs

        return model

fit(X, y)

Fit the ensemble on training data.

Parameters:

Name Type Description Default
X array - like or Tensor

Training inputs.

required
y array - like or Tensor

Training targets.

required

Returns:

Type Description
KFoldCQR

Fitted estimator.

Source code in uqregressors\conformal\k_fold_cqr.py
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
def fit(self, X, y): 
    """
    Fit the ensemble on training data.

    Args:
        X (array-like or torch.Tensor): Training inputs.
        y (array-like or torch.Tensor): Training targets.

    Returns:
        (KFoldCQR): Fitted estimator.
    """
    X_tensor, y_tensor = validate_and_prepare_inputs(X, y, device=self.device, requires_grad=self.requires_grad)
    input_dim = X_tensor.shape[1]
    self.input_dim = input_dim


    if self.scale_data:
        X_tensor = self.input_scaler.fit_transform(X_tensor)
        y_tensor = self.output_scaler.fit_transform(y_tensor)

    kf = TorchKFold(n_splits=self.n_estimators, shuffle=True)

    results = Parallel(n_jobs=self.n_jobs)(
        delayed(self._train_single_model)(X_tensor, y_tensor, input_dim, train_idx, cal_idx, i)
        for i, (train_idx, cal_idx) in enumerate(kf.split(X_tensor))
    )

    self.models = [result[0] for result in results]
    self.residuals = torch.cat([result[1] for result in results], dim=0).ravel()
    self._loggers = [result[2] for result in results]

    return self

load(path, device='cpu', load_logs=False) classmethod

Load a saved KFoldCQR model from disk.

Parameters:

Name Type Description Default
path str or Path

Directory containing saved model files.

required
device str

Device to load the model on ("cpu" or "cuda").

'cpu'
load_logs bool

Whether to also load training logs.

False

Returns:

Type Description
KFoldCQR

The loaded model instance.

Source code in uqregressors\conformal\k_fold_cqr.py
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
@classmethod
def load(cls, path, device="cpu", load_logs=False):
    """
    Load a saved KFoldCQR model from disk.

    Args:
        path (str or Path): Directory containing saved model files.
        device (str): Device to load the model on ("cpu" or "cuda").
        load_logs (bool): Whether to also load training logs.

    Returns:
        (KFoldCQR): The loaded model instance.
    """
    path = Path(path)

    # Load config
    with open(path / "config.json", "r") as f:
        config = json.load(f)
    config["device"] = device

    config.pop("optimizer", None)
    config.pop("scheduler", None)
    config.pop("input_scaler", None)
    config.pop("output_scaler", None)

    input_dim = config.pop("input_dim", None)
    model = cls(**config)

    # Recreate models
    model.input_dim = input_dim
    activation = get_activation(config["activation_str"])
    model.models = []
    for i in range(config["n_estimators"]):
        m = QuantNN(model.input_dim, config["hidden_sizes"], config["dropout"], activation).to(device)
        m.load_state_dict(torch.load(path / f"model_{i}.pt", map_location=device))
        model.models.append(m)

    # Load extras
    extras_path = path / "extras.pt"
    if extras_path.exists():
        extras = torch.load(extras_path, map_location=device, weights_only=False)
        model.conformal_width = extras.get("conformal_width", None)
        model.residuals = extras.get("residuals", None)
        model.quantiles = extras.get("quantiles", None)
    else:
        model.conformal_width = None
        model.residuals = None
        model.quantiles = None

    with open(path / "extras.pkl", 'rb') as f: 
        optimizer_cls, optimizer_kwargs, scheduler_cls, scheduler_kwargs, input_scaler, output_scaler = pickle.load(f)

    model.optimizer_cls = optimizer_cls 
    model.optimizer_kwargs = optimizer_kwargs 
    model.scheduler_cls = scheduler_cls 
    model.scheduler_kwargs = scheduler_kwargs
    model.input_scaler = input_scaler
    model.output_scaler = output_scaler

    if load_logs: 
        logs_path = path / "logs"
        training_logs = [] 
        tuning_logs = []
        if logs_path.exists() and logs_path.is_dir(): 
            estimator_log_files = sorted(logs_path.glob("estimator_*.log"))
            for log_file in estimator_log_files:
                with open(log_file, "r", encoding="utf-8") as f:
                    training_logs.append(f.read())

            tuning_log_files = sorted(logs_path.glob("tuning_*.log"))
            for log_file in tuning_log_files: 
                with open(log_file, "r", encoding="utf-8") as f: 
                    tuning_logs.append(f.read())

        model.training_logs = training_logs
        model.tuning_logs = tuning_logs

    return model

predict(X)

Predicts the target values with uncertainty estimates.

Parameters:

Name Type Description Default
X ndarray

Feature matrix of shape (n_samples, n_features).

required

Returns:

Type Description
Union[Tuple[ndarray, ndarray, ndarray], Tuple[Tensor, Tensor, Tensor]]

Tuple containing: mean predictions, lower bound of the prediction interval, upper bound of the prediction interval.

Note

If requires_grad is False, all returned arrays are NumPy arrays. Otherwise, they are PyTorch tensors with gradients.

Source code in uqregressors\conformal\k_fold_cqr.py
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
def predict(self, X): 
    """
    Predicts the target values with uncertainty estimates.

    Args:
        X (np.ndarray): Feature matrix of shape (n_samples, n_features).

    Returns:
        (Union[Tuple[np.ndarray, np.ndarray, np.ndarray], Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]): Tuple containing:
            mean predictions,
            lower bound of the prediction interval,
            upper bound of the prediction interval.

    !!! note
        If `requires_grad` is False, all returned arrays are NumPy arrays.
        Otherwise, they are PyTorch tensors with gradients.
    """
    X_tensor = validate_X_input(X, input_dim=self.input_dim, device=self.device, requires_grad=self.requires_grad)
    n = len(self.residuals)
    q = int((1 - self.alpha) * (n + 1))
    q = min(q, n-1)

    res_quantile = n-q

    self.conformal_width = torch.topk(self.residuals, res_quantile).values[-1]

    if self.scale_data: 
        X_tensor = self.input_scaler.transform(X_tensor)

    preds = [] 

    with torch.no_grad(): 
        for model in self.models: 
            model.eval()
            pred = model(X_tensor)
            preds.append(pred)

    preds = torch.stack(preds)

    means = torch.mean(preds, dim=2) 
    mean = torch.mean(means, dim=0)

    lower_cq = torch.mean(preds[:, :, 0], dim=0)
    upper_cq = torch.mean(preds[:, :, 1], dim=0)

    lower = lower_cq - self.conformal_width
    upper = upper_cq + self.conformal_width

    if self.scale_data: 
        mean = self.output_scaler.inverse_transform(mean.view(-1, 1)).squeeze()
        lower = self.output_scaler.inverse_transform(lower.view(-1, 1)).squeeze()
        upper = self.output_scaler.inverse_transform(upper.view(-1, 1)).squeeze()

    if not self.requires_grad: 
        return mean.detach().cpu().numpy(), lower.detach().cpu().numpy(), upper.detach().cpu().numpy()

    else: 
        return mean, lower, upper

quantile_loss(preds, y)

Quantile loss used for training the quantile regressors.

Parameters:

Name Type Description Default
preds Tensor

Predicted quantiles, shape (batch_size, 2).

required
y Tensor

True target values, shape (batch_size,).

required

Returns:

Type Description
Tensor

Scalar loss.

Source code in uqregressors\conformal\k_fold_cqr.py
172
173
174
175
176
177
178
179
180
181
182
183
184
def quantile_loss(self, preds, y): 
    """
    Quantile loss used for training the quantile regressors.

    Args:
        preds (Tensor): Predicted quantiles, shape (batch_size, 2).
        y (Tensor): True target values, shape (batch_size,).

    Returns:
        (Tensor): Scalar loss.
    """
    error = y.view(-1, 1) - preds
    return torch.mean(torch.max(self.quantiles * error, (self.quantiles - 1) * error))

save(path)

Save the trained model and associated configuration to disk.

Parameters:

Name Type Description Default
path str or Path

Directory to save model files.

required
Source code in uqregressors\conformal\k_fold_cqr.py
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
def save(self, path):
    """
    Save the trained model and associated configuration to disk.

    Args:
        path (str or Path): Directory to save model files.
    """
    path = Path(path)
    path.mkdir(parents=True, exist_ok=True)

    # Save config (exclude non-serializable or large objects)
    config = {
        k: v for k, v in self.__dict__.items()
        if k not in ["models", "quantiles", "residuals", "conformal_width", "optimizer_cls", "optimizer_kwargs", "scheduler_cls", "scheduler_kwargs", 
                     "input_scaler", "output_scaler", "_loggers", "training_logs", "tuning_loggers", "tuning_logs"]
        and not callable(v)
        and not isinstance(v, (torch.nn.Module,))
    }

    config["optimizer"] = self.optimizer_cls.__class__.__name__ if self.optimizer_cls is not None else None
    config["scheduler"] = self.scheduler_cls.__class__.__name__ if self.scheduler_cls is not None else None
    config["input_scaler"] = self.input_scaler.__class__.__name__ if self.input_scaler is not None else None 
    config["output_scaler"] = self.output_scaler.__class__.__name__ if self.output_scaler is not None else None

    with open(path / "config.json", "w") as f:
        json.dump(config, f, indent=4)

    # Save model weights
    for i, model in enumerate(self.models):
        torch.save(model.state_dict(), path / f"model_{i}.pt")

    # Save residuals and conformity score
    torch.save({
        "conformal_width": self.conformal_width, 
        "residuals": self.residuals,
        "quantiles": self.quantiles,
    }, path / "extras.pt")

    with open(path / "extras.pkl", 'wb') as f: 
        pickle.dump([self.optimizer_cls, 
                    self.optimizer_kwargs, self.scheduler_cls, self.scheduler_kwargs, self.input_scaler, self.output_scaler], f)

    for i, logger in enumerate(getattr(self, "_loggers", [])):
        logger.save_to_file(path, idx=i, name="estimator")

    for i, logger in enumerate(getattr(self, "tuning_loggers", [])): 
        logger.save_to_file(path, name="tuning", idx=i)

QuantNN

Bases: Module

A simple quantile neural network that estimates the lower and upper quantile when trained with a pinball loss function.

Parameters:

Name Type Description Default
input_dim int

Number of input features

required
hidden_sizes list of int

List of hidden layer sizes

required
dropout None or float

The dropout probability - None if no dropout

required
activation Module

Activation function class (e.g., nn.ReLU).

required
Source code in uqregressors\conformal\k_fold_cqr.py
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
class QuantNN(nn.Module): 
    """
    A simple quantile neural network that estimates the lower and upper quantile when trained
    with a pinball loss function. 

    Args: 
        input_dim (int): Number of input features 
        hidden_sizes (list of int): List of hidden layer sizes
        dropout (None or float): The dropout probability - None if no dropout
        activation (torch.nn.Module): Activation function class (e.g., nn.ReLU).
    """
    def __init__(self, input_dim, hidden_sizes, dropout, activation): 
        super().__init__()
        layers = []
        for h in hidden_sizes:
            layers.append(nn.Linear(input_dim, h))
            layers.append(activation())
            if dropout is not None: 
                layers.append(nn.Dropout(dropout))
            input_dim = h
        output_layer = nn.Linear(hidden_sizes[-1], 2)
        layers.append(output_layer)
        self.model = nn.Sequential(*layers)

    def forward(self, x):
        return self.model(x)