from __future__ import absolute_import, division, print_function
import logging
import numpy as np
from collections import OrderedDict
import torch
from ..utils.ml.models.ratio import DenseSingleParameterizedRatioModel
from ..utils.ml.eval import evaluate_ratio_model
from ..utils.ml.utils import get_optimizer, get_loss
from ..utils.various import load_and_check, shuffle, restrict_samplesize
from ..utils.ml.trainer import SingleParameterizedRatioTrainer
from .base import ConditionalEstimator, TheresAGoodReasonThisDoesntWork
try:
FileNotFoundError
except NameError:
FileNotFoundError = IOError
logger = logging.getLogger(__name__)
[docs]class ParameterizedRatioEstimator(ConditionalEstimator):
"""
A neural estimator of the likelihood ratio as a function of the observation x as well as
the numerator hypothesis theta. The reference (denominator) hypothesis is kept fixed at some
reference value and NOT modeled by the network.
Parameters
----------
features : list of int or None, optional
Indices of observables (features) that are used as input to the neural networks. If None, all observables
are used. Default value: None.
n_hidden : tuple of int, optional
Units in each hidden layer in the neural networks. If method is 'nde' or 'scandal', this refers to the
setup of each individual MADE layer. Default value: (100,).
activation : {'tanh', 'sigmoid', 'relu'}, optional
Activation function. Default value: 'tanh'.
"""
[docs] def train(
self,
method,
x,
y,
theta,
r_xz=None,
t_xz=None,
x_val=None,
y_val=None,
theta_val=None,
r_xz_val=None,
t_xz_val=None,
alpha=1.0,
optimizer="amsgrad",
n_epochs=50,
batch_size=128,
initial_lr=0.001,
final_lr=0.0001,
nesterov_momentum=None,
validation_split=0.25,
early_stopping=True,
scale_inputs=True,
shuffle_labels=False,
limit_samplesize=None,
memmap=False,
verbose="some",
scale_parameters=True,
n_workers=8,
clip_gradient=None,
early_stopping_patience=None,
):
"""
Trains the network.
Parameters
----------
method : str
The inference method used for training. Allowed values are 'alice', 'alices', 'carl', 'cascal', 'rascal',
and 'rolr'.
x : ndarray or str
Observations, or filename of a pickled numpy array.
y : ndarray or str
Class labels (0 = numeerator, 1 = denominator), or filename of a pickled numpy array.
theta : ndarray or str
Numerator parameter point, or filename of a pickled numpy array.
r_xz : ndarray or str or None, optional
Joint likelihood ratio, or filename of a pickled numpy array. Default value: None.
t_xz : ndarray or str or None, optional
Joint scores at theta, or filename of a pickled numpy array. Default value: None.
x_val : ndarray or str or None, optional
Validation observations, or filename of a pickled numpy array. If None
and validation_split > 0, validation data will be randomly selected from the training data.
Default value: None.
y_val : ndarray or str or None, optional
Validation labels (0 = numerator, 1 = denominator), or filename of a pickled numpy array. If None
and validation_split > 0, validation data will be randomly selected from the training data.
Default value: None.
theta_val : ndarray or str or None, optional
Validation numerator parameter points, or filename of a pickled numpy array. If None
and validation_split > 0, validation data will be randomly selected from the training data.
Default value: None.
r_xz_val : ndarray or str or None, optional
Validation joint likelihood ratio, or filename of a pickled numpy array. If None
and validation_split > 0, validation data will be randomly selected from the training data.
Default value: None.
t_xz_val : ndarray or str or None, optional
Validation joint scores at theta, or filename of a pickled numpy array. If None
and validation_split > 0, validation data will be randomly selected from the training data.
Default value: None.
alpha : float, optional
Hyperparameter weighting the score error in the loss function of the 'alices', 'rascal', and 'cascal'
methods. Default value: 1.
optimizer : {"adam", "amsgrad", "sgd"}, optional
Optimization algorithm. Default value: "amsgrad".
n_epochs : int, optional
Number of epochs. Default value: 50.
batch_size : int, optional
Batch size. Default value: 128.
initial_lr : float, optional
Learning rate during the first epoch, after which it exponentially decays to final_lr. Default value:
0.001.
final_lr : float, optional
Learning rate during the last epoch. Default value: 0.0001.
nesterov_momentum : float or None, optional
If trainer is "sgd", sets the Nesterov momentum. Default value: None.
validation_split : float or None, optional
Fraction of samples used for validation and early stopping (if early_stopping is True). If None, the entire
sample is used for training and early stopping is deactivated. Default value: 0.25.
early_stopping : bool, optional
Activates early stopping based on the validation loss (only if validation_split is not None). Default value:
True.
scale_inputs : bool, optional
Scale the observables to zero mean and unit variance. Default value: True.
shuffle_labels : bool, optional
If True, the labels (`y`, `r_xz`, `t_xz`) are shuffled, while the observations (`x`) remain in their
normal order. This serves as a closure test, in particular as cross-check against overfitting: an estimator
trained with shuffle_labels=True should predict to likelihood ratios around 1 and scores around 0.
limit_samplesize : int or None, optional
If not None, only this number of samples (events) is used to train the estimator. Default value: None.
memmap : bool, optional.
If True, training files larger than 1 GB will not be loaded into memory at once. Default value: False.
verbose : {"all", "many", "some", "few", "none}, optional
Determines verbosity of training. Default value: "some".
scale_parameters : bool, optional
Whether parameters are rescaled to mean zero and unit variance before going into the neural network.
Default value: True.
Returns
-------
results: ndarray
Results from SingleParameterizedRatioTrainer.train or DoubleParameterizedRatioTrainer.train for example
"""
logger.info("Starting training")
logger.info(" Method: %s", method)
if method in ["cascal", "rascal", "alices"]:
logger.info(" alpha: %s", alpha)
logger.info(" Batch size: %s", batch_size)
logger.info(" Optimizer: %s", optimizer)
logger.info(" Epochs: %s", n_epochs)
logger.info(" Learning rate: %s initially, decaying to %s", initial_lr, final_lr)
if optimizer == "sgd":
logger.info(" Nesterov momentum: %s", nesterov_momentum)
logger.info(" Validation split: %s", validation_split)
logger.info(" Early stopping: %s", early_stopping)
logger.info(" Scale inputs: %s", scale_inputs)
logger.info(" Scale parameters: %s", scale_parameters)
logger.info(" Shuffle labels %s", shuffle_labels)
if limit_samplesize is None:
logger.info(" Samples: all")
else:
logger.info(" Samples: %s", limit_samplesize)
# Load training data
logger.info("Loading training data")
memmap_threshold = 1.0 if memmap else None
theta = load_and_check(theta, memmap_files_larger_than_gb=memmap_threshold)
x = load_and_check(x, memmap_files_larger_than_gb=memmap_threshold)
y = load_and_check(y, memmap_files_larger_than_gb=memmap_threshold)
r_xz = load_and_check(r_xz, memmap_files_larger_than_gb=memmap_threshold)
t_xz = load_and_check(t_xz, memmap_files_larger_than_gb=memmap_threshold)
self._check_required_data(method, r_xz, t_xz)
# Infer dimensions of problem
n_samples = x.shape[0]
n_observables = x.shape[1]
n_parameters = theta.shape[1]
logger.info("Found %s samples with %s parameters and %s observables", n_samples, n_parameters, n_observables)
# Limit sample size
if limit_samplesize is not None and limit_samplesize < n_samples:
logger.info("Only using %s of %s training samples", limit_samplesize, n_samples)
x, theta, y, r_xz, t_xz = restrict_samplesize(limit_samplesize, x, theta, y, r_xz, t_xz)
# Validation data
external_validation = x_val is not None and y_val is not None and theta_val is not None
if external_validation:
theta_val = load_and_check(theta_val, memmap_files_larger_than_gb=memmap_threshold)
x_val = load_and_check(x_val, memmap_files_larger_than_gb=memmap_threshold)
y_val = load_and_check(y_val, memmap_files_larger_than_gb=memmap_threshold)
r_xz_val = load_and_check(r_xz_val, memmap_files_larger_than_gb=memmap_threshold)
t_xz_val = load_and_check(t_xz_val, memmap_files_larger_than_gb=memmap_threshold)
logger.info("Found %s separate validation samples", x_val.shape[0])
assert x_val.shape[1] == n_observables
assert theta_val.shape[1] == n_parameters
if r_xz is not None:
assert r_xz_val is not None, "When providing r_xz and sep. validation data, also provide r_xz_val"
if t_xz is not None:
assert t_xz_val is not None, "When providing t_xz and sep. validation data, also provide t_xz_val"
# Scale features
if scale_inputs:
self.initialize_input_transform(x, overwrite=False)
x = self._transform_inputs(x)
if external_validation:
x_val = self._transform_inputs(x_val)
else:
self.initialize_input_transform(x, False, overwrite=False)
# Scale parameters
if scale_parameters:
logger.info("Rescaling parameters")
self.initialize_parameter_transform(theta)
theta = self._transform_parameters(theta)
t_xz = self._transform_score(t_xz, inverse=False)
if external_validation:
t_xz_val = self._transform_score(t_xz_val, inverse=False)
else:
self.initialize_parameter_transform(theta, False)
# Shuffle labels
if shuffle_labels:
logger.info("Shuffling labels")
y, r_xz, t_xz = shuffle(y, r_xz, t_xz)
# Features
if self.features is not None:
x = x[:, self.features]
logger.info("Only using %s of %s observables", x.shape[1], n_observables)
n_observables = x.shape[1]
if external_validation:
x_val = x_val[:, self.features]
# Check consistency of input with model
if self.n_observables is None:
self.n_observables = n_observables
if self.n_parameters is None:
self.n_parameters = n_parameters
if n_parameters != self.n_parameters:
raise RuntimeError(
"Number of parameters does not match model: {} vs {}".format(n_parameters, self.n_parameters)
)
if n_observables != self.n_observables:
raise RuntimeError(
"Number of observables does not match model: {} vs {}".format(n_observables, self.n_observables)
)
# Data
data = self._package_training_data(method, x, theta, y, r_xz, t_xz)
if external_validation:
data_val = self._package_training_data(method, x_val, theta_val, y_val, r_xz_val, t_xz_val)
else:
data_val = None
# Create model
if self.model is None:
logger.info("Creating model")
self._create_model()
# Losses
loss_functions, loss_labels, loss_weights = get_loss(method, alpha)
# Optimizer
opt, opt_kwargs = get_optimizer(optimizer, nesterov_momentum)
# Train model
logger.info("Training model")
trainer = SingleParameterizedRatioTrainer(self.model, n_workers=n_workers)
result = trainer.train(
data=data,
data_val=data_val,
loss_functions=loss_functions,
loss_weights=loss_weights,
loss_labels=loss_labels,
epochs=n_epochs,
batch_size=batch_size,
optimizer=opt,
optimizer_kwargs=opt_kwargs,
initial_lr=initial_lr,
final_lr=final_lr,
validation_split=validation_split,
early_stopping=early_stopping,
verbose=verbose,
clip_gradient=clip_gradient,
early_stopping_patience=early_stopping_patience,
)
return result
[docs] def evaluate_log_likelihood_ratio(self, x, theta, test_all_combinations=True, evaluate_score=False):
"""
Evaluates the log likelihood ratio for given observations x betwen the given parameter point theta and the
reference hypothesis.
Parameters
----------
x : str or ndarray
Observations or filename of a pickled numpy array.
theta : ndarray or str
Parameter points or filename of a pickled numpy array.
test_all_combinations : bool, optional
If False, the number of samples in the observable and theta
files has to match, and the likelihood ratio is evaluated only for the combinations
`r(x_i | theta0_i, theta1_i)`. If True, `r(x_i | theta0_j, theta1_j)` for all pairwise combinations `i, j`
are evaluated. Default value: True.
evaluate_score : bool, optional
Sets whether in addition to the likelihood ratio the score is evaluated. Default value: False.
Returns
-------
log_likelihood_ratio : ndarray
The estimated log likelihood ratio. If test_all_combinations is True, the result has shape
`(n_thetas, n_x)`. Otherwise, it has shape `(n_samples,)`.
score : ndarray or None
None if evaluate_score is False. Otherwise the derived estimated score at `theta0`. If test_all_combinations
is True, the result has shape `(n_thetas, n_x, n_parameters)`. Otherwise, it has shape
`(n_samples, n_parameters)`.
"""
if self.model is None:
raise ValueError("No model -- train or load model before evaluating it!")
# Load training data
logger.debug("Loading evaluation data")
x = load_and_check(x)
theta = load_and_check(theta)
# Scale observables
x = self._transform_inputs(x)
theta = self._transform_parameters(theta)
# Restrict features
if self.features is not None:
x = x[:, self.features]
all_log_r_hat = []
all_t_hat = []
if test_all_combinations:
logger.debug("Starting ratio evaluation for %s x-theta combinations", len(theta) * len(x))
for i, this_theta in enumerate(theta):
logger.debug("Starting ratio evaluation for thetas %s / %s: %s", i + 1, len(theta), this_theta)
_, log_r_hat, t_hat, _ = evaluate_ratio_model(
model=self.model,
method_type="parameterized_ratio",
theta0s=[this_theta],
theta1s=None,
xs=x,
evaluate_score=evaluate_score,
)
t_hat = self._transform_score(t_hat, inverse=True)
all_log_r_hat.append(log_r_hat)
all_t_hat.append(t_hat)
all_log_r_hat = np.array(all_log_r_hat)
all_t_hat = np.array(all_t_hat)
else:
logger.debug("Starting ratio evaluation")
_, all_log_r_hat, all_t_hat, _ = evaluate_ratio_model(
model=self.model,
method_type="parameterized_ratio",
theta0s=theta,
theta1s=None,
xs=x,
evaluate_score=evaluate_score,
)
all_t_hat = self._transform_score(all_t_hat, inverse=True)
logger.debug("Evaluation done")
return all_log_r_hat, all_t_hat
[docs] def evaluate_log_likelihood_ratio_torch(self, x, theta, test_all_combinations=True):
"""
Evaluates the log likelihood ratio for given observations x betwen the given parameter point theta and the
reference hypothesis.
Parameters
----------
x : torch.tensor
Observations.
theta : torch.tensor
Parameter points.
test_all_combinations : bool, optional
If False, the number of samples in the observable and theta
files has to match, and the likelihood ratio is evaluated only for the combinations
`r(x_i | theta0_i, theta1_i)`. If True, `r(x_i | theta0_j, theta1_j)` for all pairwise combinations `i, j`
are evaluated. Default value: True.
Returns
-------
log_likelihood_ratio : torch.tensor
The estimated log likelihood ratio. If test_all_combinations is True, the result has shape
`(n_thetas, n_x)`. Otherwise, it has shape `(n_samples,)`.
"""
if self.model is None:
raise ValueError("No model -- train or load model before evaluating it!")
# Scale observables and parameters
x = self._transform_inputs(x)
theta = self._transform_parameters(theta)
# Eval
if test_all_combinations:
logger.debug("Starting ratio evaluation for %s x-theta combinations", len(theta) * len(x))
all_log_r_hat = []
for theta_ in theta:
theta_ = torch.stack([theta_ for _ in x])
_, log_r_hat, _ = self.model(theta_, x)
all_log_r_hat.append(log_r_hat[:, 0].unsqueeze(0))
log_r_hat = torch.cat(all_log_r_hat, dim=0)
else:
logger.debug("Starting ratio evaluation")
_, log_r_hat = self.model(theta, x)
return log_r_hat
[docs] def evaluate_log_likelihood(self, *args, **kwargs):
raise TheresAGoodReasonThisDoesntWork(
"This estimator can only estimate likelihood ratios, not the likelihood " "itself!"
)
[docs] def evaluate_score(self, x, theta, nuisance_mode="keep"):
"""
Evaluates the scores for given observations x betwen at a given parameter point theta.
Parameters
----------
x : str or ndarray
Observations or filename of a pickled numpy array.
theta : ndarray or str
Parameter points or filename of a pickled numpy array.
nuisance_mode : {"auto", "keep", "profile", "project"}
Decides how nuisance parameters are treated. If nuisance_mode is "keep", the
returned score is always (n+k)-dimensional.
Returns
-------
score : ndarray or None
The estimated score at `theta`. If test_all_combinations is True, the
result has shape `(n_thetas, n_x, n_parameters)`. Otherwise, it has shape
`(n_samples, n_parameters)`.
"""
if nuisance_mode == "keep":
logger.debug("Keeping nuisance parameter in score")
else:
raise ValueError("Unknown nuisance_mode {}".format(nuisance_mode))
_, all_t_hat = self.evaluate_log_likelihood_ratio(x, theta, test_all_combinations=False, evaluate_score=True)
return all_t_hat
[docs] def evaluate(self, *args, **kwargs):
return self.evaluate_log_likelihood_ratio(*args, **kwargs)
def _create_model(self):
self.model = DenseSingleParameterizedRatioModel(
n_observables=self.n_observables,
n_parameters=self.n_parameters,
n_hidden=self.n_hidden,
activation=self.activation,
dropout_prob=self.dropout_prob,
)
@staticmethod
def _check_required_data(method, r_xz, t_xz):
if method in ["cascal", "alices", "rascal"] and t_xz is None:
raise RuntimeError("Method {} requires joint score information".format(method))
if method in ["rolr", "alice", "alices", "rascal"] and r_xz is None:
raise RuntimeError("Method {} requires joint likelihood ratio information".format(method))
@staticmethod
def _package_training_data(method, x, theta, y, r_xz, t_xz):
data = OrderedDict()
data["x"] = x
data["theta"] = theta
data["y"] = y
if method in ["rolr", "alice", "alices", "rascal"]:
data["r_xz"] = r_xz
if method in ["cascal", "alices", "rascal"]:
data["t_xz"] = t_xz
return data
def _wrap_settings(self):
settings = super(ParameterizedRatioEstimator, self)._wrap_settings()
settings["estimator_type"] = "parameterized_ratio"
return settings
def _unwrap_settings(self, settings):
super(ParameterizedRatioEstimator, self)._unwrap_settings(settings)
estimator_type = str(settings["estimator_type"])
if estimator_type != "parameterized_ratio":
raise RuntimeError("Saved model is an incompatible estimator type {}.".format(estimator_type))