API

Array

Array-like type

alias of Union[numpy.ndarray, tensorflow.python.ops.tensor_array_ops.TensorArray, float, List[float]]

class EmptyPrior

Bases: maxent.core.Prior

No prior deviation from target for restraint (exact agreement)

expected(l)

Expected disagreement

Parameters

l (float) – The lagrange multiplier

Returns

expected disagreement

class Laplace(sigma)

Bases: maxent.core.Prior

Laplace distribution prior expected deviation from target for restraint

Parameters

sigma (float) – Parameter for Laplace prior - higher means more allowable disagreement

expected(l)

Expected disagreement

Parameters

l (float) – The lagrange multiplier

Return type

float

Returns

expected disagreement

class MaxentModel(restraints, name='maxent-model', **kwargs)

Bases: keras.engine.training.Model

Keras Maximum entropy model

Parameters
call(inputs)

Compute reweighted restraint values

Parameters

inputs (Union[ndarray, TensorArray, float, List[float], List[Union[ndarray, TensorArray, float, List[float]]], Tuple[Union[ndarray, TensorArray, float, List[float]]]]) – Restraint values

Return type

TensorArray

Returns

Weighted restraint values

compile(optimizer=<keras.optimizers.optimizer_v2.adam.Adam object>, loss='mean_squared_error', metrics=None, loss_weights=None, weighted_metrics=None, run_eagerly=None, steps_per_execution=None, **kwargs)

See compile method of tf.keras.Model

fit(trajs, input_weights=None, batch_size=None, epochs=128, **kwargs)

Fit to given observations with restraints

Parameters

  • trajs (Union[ndarray, TensorArray, float, List[float]]) – Observations, which can be input to Restraint

  • input_weights (Union[ndarray, TensorArray, float, List[float], None]) – Array of weights which will be start

  • batch_size (Optional[int]) – Almost always should be equal to number of trajs, unless you want to mix your Lagrange multipliers across trajectories

  • kwargs – See :class:tf.keras.Model fit method for further optional arguments, like verbose=0 to hide output

Return type

History

Returns

The history of fit

reset_weights()

Zero out the weights of the model

class Prior

Bases: object

Prior distribution for expected deviation from target for restraint

expected(l)

Expected disagreement

Parameters

l (float) – The lagrange multiplier

Return type

float

Returns

expected disagreement

class Restraint(fxn, target, prior=<maxent.core.EmptyPrior object>)

Bases: object

Restraint - includes function, target, and prior belief in deviation from target

Parameters

  • fxn (Callable[[Union[ndarray, TensorArray, float, List[float]]], float]) – callable that returns scalar

  • target (float) – desired scalar value

  • prior (Prior) – prior is a Prior for expected deviation from that target

class HyperMaxentModel(restraints, prior_model, simulation, reweight=True, name='hyper-maxent-model', **kwargs)

Bases: maxent.core.MaxentModel

Keras Maximum entropy model

Parameters

  • restraints (List[Restraint]) – List of Restraint

  • prior_model (ParameterJoint) – ParameterJoint that specifies prior

  • simulation (Callable[[Union[ndarray, TensorArray, float, List[float]]], Union[ndarray, TensorArray, float, List[float]]]) – Callable that will generate observations given the output from prior_model

  • reweight (bool) – True means use to remove effect of prior training updates via reweighting, which keeps as close as possible to given untrained prior_model

  • name (str) – Name of model

fit(sample_batch_size=256, final_batch_multiplier=4, param_epochs=None, outer_epochs=10, **kwargs)

Fit to given outcomes from simulation

Parameters

  • sample_batch_size (int) – Number of observations to sample per outer_epochs

  • final_batch_multiplier (int) – Sets number of final MaxEnt fitting step after training prior_model. Final number of MaxEnt steps will be final_batch_multiplier * sample_batch_size

  • param_epochs (Optional[int]) – Number of times prior_model will be fit to sampled observations

  • outer_epochs (int) – Number of loops of sampling/prior_model fitting

  • kwargs – See :class:tf.keras.Model fit method for further optional arguments, like verbose=0 to hide output

Return type

History

Returns

The tf.keras.callbacks.History of fit

class ParameterJoint(reshapers=None, inputs=None, outputs=None, **kwargs)

Bases: keras.engine.training.Model

Prior parameter model joint distribution

This packages up how you want to sample prior paramters into one joint distribution. This has an important ability of reshaping output from distributions in case your simulation requires matrices, applying constraints, or projections.

Parameters

  • inputs (Union[Input, Tuple[Input], None]) – tf.keras.Input or tuple of them.

  • outputs (Optional[List[Distribution]]) – list of tfp.distributions.Distribution

  • reshapers (Optional[List[Callable[[Union[ndarray, TensorArray, float, List[float]]], Union[ndarray, TensorArray, float, List[float]]]]]) – optional list of callables that will be called on outputs from your distribution

compile(optimizer, **kwargs)

See compile method of tf.keras.Model

sample(N, return_joint=False)

Generate sample

Parameters
Return type

Union[Tuple[Union[ndarray, TensorArray, float, List[float]], Union[ndarray, TensorArray, float, List[float]], Any], ndarray, TensorArray, float, List[float]]

Returns

the reshaped output samples and (if return_joint) a value y which can be used to compute probabilities and tfp.distributions.Distribution joint

class TrainableInputLayer(initial_value, constraint=None, **kwargs)

Bases: keras.engine.base_layer.Layer

Create trainable input layer for tfp.distributions.Distribution

This will, given a fake input, return a trainable weight set by initial_value. Use to feed into distributions that can be trained.

Parameters

  • initial_value (Union[ndarray, TensorArray, float, List[float]]) – starting value, determines shape/dtype of output

  • constraint (Optional[Callable[[Union[ndarray, TensorArray, float, List[float]]], float]]) – Callable that returns scalar given output. See tf.keras.layers.Layer

  • kwargs – See tf.Keras.layers.Layer for additional arguments

call(inputs)

See call of tf.keras.layers.Layer

Return type

Union[ndarray, TensorArray, float, List[float]]

negloglik(y, rv_y)

negative log likelihood

Parameters

  • y (Union[ndarray, TensorArray, float, List[float]]) – observations

  • rv_y (Distribution) – distribution

Return type

Union[ndarray, TensorArray, float, List[float]]

Returns

negative log likelihood of y