bayesflow.helper_classes module

class bayesflow.helper_classes.SimulationDataset(forward_dict, batch_size, buffer_size=1024)

Bases: object

Helper class to create a tensorflow.data.Dataset which parses simulation dictionaries and returns simulation dictionaries as expected by BayesFlow amortizers.

__init__(forward_dict, batch_size, buffer_size=1024)

Creates a wrapper holding a tf.data.Dataset instance for offline training in an amortized estimation context.

Parameters:

forward_dictdict

The outputs from a GenerativeModel or a custom function, stored in a dictionary with at least the following keys: sim_data - an array representing the batched output of the model prior_draws - an array with prior generated from the model’s prior

batch_sizeint

The total number of simulations from all models in a given batch. The batch size per model will be calculated as batch_size // num_models

buffer_sizeint, optional, default: 1024

The buffer size for shuffling elements in a tf.data.Dataset

__call__(batch_in)

Convert output of tensorflow.data.Dataset to dict.

class bayesflow.helper_classes.MultiSimulationDataset(forward_dict, batch_size, buffer_size=1024)

Bases: object

Helper class for model comparison training with multiple generative models.

Will create multiple SimulationDataset instances, each parsing their own simulation dictionaries and returning these as expected by BayesFlow amortizers.

__init__(forward_dict, batch_size, buffer_size=1024)

Creates a wrapper holding multiple tf.data.Dataset instances for offline training in an amortized model comparison context.

Parameters:

forward_dictdict

The outputs from a MultiGenerativeModel or a custom function, stored in a dictionary with at least the following keys: model_outputs - a list with length equal to the number of models, each element representing a batched output of a single model model_indices - a list with integer model indices, which will later be one-hot-encoded for the model comparison learning problem.

batch_sizeint

The total number of simulations from all models in a given batch. The batch size per model will be calculated as batch_size // num_models

buffer_sizeint, optional, default: 1024

The buffer size for shuffling elements in a tf.data.Dataset

class bayesflow.helper_classes.EarlyStopper(patience=5, tolerance=0.05)

Bases: object

This class will track the total validation loss and trigger an early stopping recommendation based on its hyperparameters.

__init__(patience=5, tolerance=0.05)

patienceint, optional, default: 5

How many successive times the tolerance value is reached before triggering an early stopping recommendation.

tolerancefloat, optional, default: 0.05

The minimum reduction of validation loss to be considered significant.

update_and_recommend(current_val_loss)

Adds loss to history and check difference between sequential losses.

class bayesflow.helper_classes.RegressionLRAdjuster(optimizer, period=1000, wait_between_fits=10, patience=10, tolerance=-0.05, reduction_factor=0.25, cooldown_factor=2, num_resets=3, **kwargs)

Bases: object

This class will compute the slope of the loss trajectory and inform learning rate decay.

file_name = 'lr_adjuster'

__init__(optimizer, period=1000, wait_between_fits=10, patience=10, tolerance=-0.05, reduction_factor=0.25, cooldown_factor=2, num_resets=3, **kwargs)

Creates an instance with given hyperparameters which will track the slope of the loss trajectory according to specified hyperparameters and then issue an optional stopping suggestion.

Parameters:

optimizertf.keras.optimizers.Optimizer instance

An optimizer implementing a lr() method

periodint, optional, default: 1000

How much loss values to consider from the past

wait_between_fitsint, optional, default: 10

How many backpropagation updates to wait between two successive fits

patienceint, optional, default: 10

How many successive times the tolerance value is reached before lr update.

tolerancefloat, optional, default: -0.05

The minimum slope to be considered substantial for training.

reduction_factorfloat in [0, 1], optional, default: 0.25

The factor by which the learning rate is reduced upon hitting the tolerance threshold for patience number of times

cooldown_factorfloat, optional, default: 2

The factor by which the period is multiplied to arrive at a cooldown period.

num_resetsint, optional, default: 3

How many times to reduce the learning rate before issuing an optional stopping

**kwargsdict, optional, default {}

Additional keyword arguments passed to the HuberRegression class.

get_slope(losses)

Fits a Huber regression on the provided loss trajectory or returns None if not enough data points present.

reset()

Resets all stateful variables in preparation for a new start.

save_to_file(file_path)

Saves the state parameters of a RegressionLRAdjuster object to a pickled dictionary in file_path.

load_from_file(file_path)

Loads the saved LRAdjuster object from file_path.

class bayesflow.helper_classes.LossHistory

Bases: object

Helper class to keep track of losses during training.

file_name = 'history'

__init__()

property total_loss

property total_val_loss

last_total_loss()

last_total_val_loss()

start_new_run()

add_val_entry(epoch, val_loss)

Add validation entry to loss structure. Assume loss_names already exists as an attribute, so no attempt will be made to create names.

add_entry(epoch, current_loss)

Adds loss entry for current epoch into internal memory data structure.

get_running_losses(epoch)

Compute and return running means of the losses for current epoch.

get_plottable()

Returns the losses as a nicely formatted pandas DataFrame, in case only train losses were collected, otherwise a dict of data frames.

flush()

Returns current history and removes all existing loss history, but keeps loss names.

save_to_file(file_path, max_to_keep)

Saves a LossHistory object to a pickled dictionary in file_path. If max_to_keep saved loss history files are found in file_path, the oldest is deleted before a new one is saved.

load_from_file(file_path)

Loads the most recent saved LossHistory object from file_path.

class bayesflow.helper_classes.SimulationMemory(stores_raw=True, capacity_in_batches=50)

Bases: object

Helper class to keep track of a pre-determined number of simulations during training.

file_name = 'memory'

__init__(stores_raw=True, capacity_in_batches=50)

store(forward_dict)

Stores simulation outputs in forward_dict, if internal buffer is not full.

Parameters:

forward_dictdict

The configured outputs of the forward model.

get_memory()

is_full()

Returns True if the buffer is full, otherwise False.

save_to_file(file_path)

Saves a SimulationMemory object to a pickled dictionary in file_path.

load_from_file(file_path)

Loads the saved SimulationMemory object from file_path.

class bayesflow.helper_classes.MemoryReplayBuffer(capacity_in_batches=500)

Bases: object

Implements a memory replay buffer for simulation-based inference.

__init__(capacity_in_batches=500)

Creates a circular buffer following the logic of experience replay.

Parameters:

capacity_in_batchesint, optional, default: 500

The capacity of the buffer in batches of simulations. Could potentially grow very large, so make sure you pick a reasonable number!

store(forward_dict)

Stores simulation outputs, if internal buffer is not full.

Parameters:

forward_dictdict

The confogired outputs of the forward model.

sample()

Samples batch_size number of parameter vectors and simulations from buffer.

Returns:

forward_dictdict

The (raw or configured) outputs of the forward model.