Datastore

Purpose of the datastore

The most important class here is the History class. The History class is the interface to the database in which pyABC stores and logs information during the ABC-SMC run, but also the interface which allows you to query that information later on.

Initializing the databse interface from a file

For querying, you initialize a History object with a valid SQLAlchmey database identifier. For example, if you ABC-SMC data is stored in a file “data.db”, you initialize the History with:

history = History("sqlite:///data.db")

Don’t mind the three slashes. This is SQLAlchemy syntax.

If more than one ABC-SMC run is stored in your database file, these runs will have IDs. The first run has ID=1, the second run ID=2 and so on. Per default, the first run found in the database is automatically selected. To select a specific run by manually, do

history.id = n

if n is the run number, e.g. n=3.

Querying the database

The History class has a number of methods which are relevant for querying the stored data. The most important ones are:

  • History.get_distribution to retrieve information on the parameter posteriors,
  • History.get_model_probabilities to retrieve information on the model probabilities in case you’re doing model selection,
  • History.get_all_populations, to retrieve information on the evolution of the acceptance threshold and the number of sample attempts per population,
  • History.get_nr_particles_per_population, to retrieve the number of particles per population (this number os not necessariliy constant),
  • History.get_weighted_distances, to retrieve the distances the parameter samples achieved,
  • History.n_populations to get the total number of populations, and
  • History.total_nr_simulations to get the total number of simulations, i.e. sample attempts.

Use get_distribution to retrieve your posterior particle population. For example,

df, w = history.get_distribution(m)

will return a DataFrame df of parameters and an array w of weights of the particles of model m in the last available population. If you’re interested in intermediate populations, add the optional t parameter, which indicates the population number (the first population is t=0)

df, w = history.get_distribution(m, t)

What can be stored as summary statistics

Currently, integers, floats, strings and in general everything that can be converted to a numpy array can be stored. In addition it is also possible to store pandas DataFrames.

Warning

Storage of pandas DataFrames is considered experimental at this point.

class pyabc.storage.History(db: str)

Bases: object

History for ABCSMC.

This class records the evolution of the populations and stores the ABCSMC results.

Parameters:db (str) – SQLAlchemy database identifier.
alive_models(t) → List

Get the models which are still alive at time t.

Parameters:t (int) – Population nr
Returns:alive – A list which contains the indices of those models which are still alive
Return type:List
all_runs()

Get all ABCSMC runs which are stored in the database.

append_population(t: int, current_epsilon: float, population: pyabc.storage.db_model.Population, nr_simulations: int, model_names)

Append population to database.

Parameters:
  • t (int) – Population number.
  • current_epsilon (float) – Current epsilon value.
  • population (Population) – List of sampled particles.
  • nr_simulations (int) – The number of model evaluations for this population.
  • model_names (list) – The model names.

Note. This function is called by the pyabc.ABCSMC class internally. You should most likely not find it necessary to call this method under normal circumstances.

db_size

Size of the database.

Returns:db_size – Size of the SQLite database in MB. Currently this only works for SQLite databases.

Returns an error string if the DB size cannot be calculated.

Return type:int, str
done()

Close database sessions and store end time of population.

Note. This function is called by the pyabc.ABCSMC class internally. You should most likely not find it necessary to call this method under normal circumstances.

get_all_populations()

Returns a pandas Dataframe with columns

  • t: Popultion number
  • population_end_time: The end time of the population
  • samples: The number of sample attempts performed
    for a population
  • epsilon: The acceptence threshold for the population.
Returns:all_populations – DataFrame with population info
Return type:pd.DataFrame
get_distribution(m: int, t: int = None) -> (<class 'pandas.core.frame.DataFrame'>, <class 'numpy.ndarray'>)

Returns the weighted population sample as pandas DataFrame.

Parameters:
  • m (int) – model index
  • t (int, optional) – Population number. If t is not specified, then the last population is returned.
Returns:

  • df, w (pandas.DataFrame, np.ndarray)
  • df – is a DataFrame of parameters
  • w – are the weights associated with each parameter

get_model_probabilities(t=None) → pandas.core.frame.DataFrame

Model probabilities.

Parameters:t (int or None) – Population. Defaults to None, i.e. the last population.
Returns:probabilities – Model probabilities
Return type:np.ndarray
get_nr_particles_per_population() → pandas.core.series.Series
Returns:nr_particles_per_population – A pandas DataFrame containing the number of particles for each population
Return type:pd.DataFrame
get_population_extended(*, m=None, t='last', tidy=True) → pandas.core.frame.DataFrame

Get extended population information, including parameters, distances, summary statistics, weights and more.

Parameters:
  • m (int, optional) – The model to query. If omitted, all models are returned
  • t (str, optional) – Can be “last” or “all” In case of “all”, all populations are returned. If “last”, only the last population is returned.
  • tidy (bool, optional) – If True, try to return a tidy DataFrame, where the individual parameters and summary statistics are pivoted. Setting tidy to true will only work for a single model and a single population.
Returns:

full_population

Return type:

DataFrame

get_population_strategy()
Returns:The population strategy.
Return type:population_strategy
get_sum_stats(t: int, m: int) -> (<class 'numpy.ndarray'>, typing.List)

Summary statistics

Parameters:
  • t (int) – Population number
  • m (int) – Model index
Returns:

w, sum_stats

  • w: the weights associated with the summary statistics
  • sum_stats: list of summary statistics

Return type:

np.ndarray, list

get_weighted_distances(t: Union[int, NoneType]) → pandas.core.frame.DataFrame

Population’s weighted distances to the measured sample. These weights do not necessarily sum up to 1. In case more than one simulation per parameter is performed and accepted the sum might be larger.

Parameters:t (int, None) – Population number. If t is None the last population is selected.
Returns:median – Weighted distances. The dataframe has column “w” for the weights and column distance for the distances
Return type:DataFrame
max_t

The population number of the last populations. This is equivalent to n_populations - 1.

n_populations

Number of populations stored in the database. This is equivalent to max_t + 1.

nr_of_models_alive(t=None) → int

Number of models still alive.

Parameters:t (int) – Population number
Returns:nr_alive – Number of models still alive. None is for the last population
Return type:int >= 0 or None
store_initial_data(ground_truth_model: int, options: dict, observed_summary_statistics: dict, ground_truth_parameter: dict, model_names: List[str], distance_function_json_str: str, eps_function_json_str: str, population_strategy_json_str: str)

Store the initial configuration data.

Parameters:
  • ground_truth_model (int) – Nr of the ground truth model.
  • options (dict) – Of ABC metadata
  • observed_summary_statistics (dict) – the measured summary statistics
  • ground_truth_parameter (dict) – the ground truth parameters
  • model_names (List) – A list of model names
  • distance_function_json_str (str) – The distance function represented as json string
  • eps_function_json_str (str) – The epsilon represented as json string
  • population_strategy_json_str (str) – The population strategy represented as json string

Note. This function is called by the pyabc.ABCSMC class internally. You should most likely not find it necessary to call this method under normal circumstances.

total_nr_simulations

Number of sample attempts for the ABC run.

Returns:Total nr of sample attempts for the ABC run.
Return type:int