Source code for stratmc.inference

import os
import sys
import warnings
from datetime import date
from pathlib import Path
from time import asctime

import arviz as az
import numpy as np
import numpyro
import pandas as pd
import pymc as pm
import pymc.sampling.jax
import xarray as xr
from scipy.stats import gaussian_kde
from tqdm.notebook import tqdm

numpyro.enable_x64()

import pytensor

os.environ["PYTENSOR_FLAGS"] = "mode=FAST_RUN,device=cpu,floatX=float64"

warnings.filterwarnings("ignore", ".*The group X_new is not defined in the InferenceData scheme.*")
warnings.filterwarnings("ignore", ".*X_new group is not defined in the InferenceData scheme.*")

from stratmc.data import clean_data, drop_chains, save_object
from stratmc.tests import check_inference


[docs] def get_trace(model, gp, ages, sample_df, ages_df, proxies = ['d13c'], approximate = False, name="", chains = 8, draws = 1000, tune = 2000, prior_draws = 1000, target_accept = 0.9, sampler = 'numpyro', nuts_kwargs = None, jitter = 0.001, seed = None, save = True, postprocessing_backend = None, generate_custom_initvals = True, initval_seed = None, save_custom_initvals = True, initvals = None, sample_predictive = True, chain_method = 'parallel', **kwargs): """ Sample the prior and posterior distributions for a :class:`pymc.model.core.Model` returned by :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`. By default, uses :py:func:`pymc.sampling.jax.sample_numpyro_nuts() <pymc.sampling.jax.sample_numpyro_nuts>` to sample the posterior; change ``sampler`` to 'blackjax' to use :py:func:`pymc.sampling.jax.sample_blackjax_nuts() <pymc.sampling.jax.sample_blackjax_nuts>`. After the posterior has been sampled, runs :py:meth:`check_inference() <stratmc.tests.check_inference>` in :py:mod:`stratmc.tests` to check that superposition is never violated in the posterior. Any chains with superposition violations are removed from the trace with :py:meth:`drop_chains() <stratmc.data.drop_chains>` before it is returned (if ``save = True``, both the original and 'cleaned' traces are saved to the ``traces`` subfolder), and a warning is issued. See :py:meth:`check_inference() <stratmc.tests.check_inference>` for details; superposition issues are rare, and typically are related to minor violations of detrital or intrusive age constraints. Problems during sampling, including frequent divergences or minor violations of limiting age constraints, might be resolved by increasing the number of ``tune`` steps and/or increasing ``target_accept`` (which decreases the step size). Parameters ---------- model: pymc.Model :class:`pymc.model.core.Model` object returned by :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`. gp: pymc.gp.Latent Gaussian process prior (:class:`pymc.gp.Latent` or :class:`pymc.gp.HSGP`) returned by :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`. ages: numpy.array(float) array of ages at which to sample the posterior distribution of the proxy signal. sample_df: pandas.DataFrame :class:`pandas.DataFrame` containing proxy data for all sections. ages_df: pandas.DataFrame :class:`pandas.DataFrame` containing age constraints for all sections. sections:: list(str) or numpy.array(str), optional List of sections to include in the inference model. Defaults to all sections in ``sample_df``. proxies: str or list(str), optional List of proxies included in the model. Defaults to 'd13c'. approximate: bool, optional Set to ``True`` if the Hilbert space GP approximation (:class:`pymc.gp.HSGP`) was used in :py:meth:`build_model() <stratmc.model.build_model>`; defaults to ``False``. name: str, optional Prefix for the saved NetCDF file with the inference results (suffix is timestamp for function call). chains: int, optional Number of Markov chains to sample in parallel; defaults to 8. draws: int, optional Number of samples per chain to draw from the posterior; defaults to 1000. tune: int, optional Number of iterations to tune; defaults to 1000. prior_draws: int, optional Number of samples to draw from the prior; defaults to 1000. target_accept: float, optional Between 0 and 1 (exclusive). During tuning, the sampler adapts the proposals such that the average acceptance probability is equal to ``target_accept``; higher values for ``target_accept`` typically lead to smaller step sizes. Defaults to 0.9. generate_custom_initvals: bool, optional Whether to generate custom initial values for each chain with :py:meth:`make_initial_values_per_chain() <stratmc.inference.make_initial_values_per_chain>` prior to sampling; defaults to ``True``. Recommended to improve exploration of the posterior, and required to avoid superposition violations for models with intermediate detrial or intrusive age constraints. initval_seed: int, optional Random seed for initial value generator. save_custom_initvals: bool, optional Whether to save the initial value dictionary generated by :py:meth:`make_initial_values_per_chain() <stratmc.inference.make_initial_values_per_chain>` (to 'initial_values' subfolder in the current directory); defaults to ``True``. initvals: dict, optional Diciontary of custom initial values generated by :py:meth:`make_initial_values_per_chain() <stratmc.inference.make_initial_values_per_chain>`. If not provided, initial values will be generated with prior to sampling if ``generate_custom_initvals`` is ``True``; if ``generated_custom_initvals`` is ``False``, the default initial point of ``model`` will be used. sampler: str, optional Which NUTS algorithm to use to sample the posterior ('numpyro' or 'blackjax'); defaults to 'numpyro'. nuts_kwargs: dict, optional Dictionary of keyword arguments passed to NumPyro NUTS sampler (see :py:func:`pymc.sampling.jax.sample_numpyro_nuts() <pymc.sampling.jax.sample_numpyro_nuts>` and :class:`numpyro.infer.hmc.NUTS`) or blackjax NUTS sampler (see :py:func:`pymc.sampling.jax.sample_blackjax_nuts() <pymc.sampling.jax.sample_blackjax_nuts>`). sample_predictive: bool, optional Whether to draw prior and posterior predictive samples; defaults to ``True``. jitter: float, optional Value of ``jitter`` passed to :meth:`pymc.gp.Latent.conditional`. Defaults to 0.001. Changing this value may help if a linear algebra error is encountered during posterior predictive sampling. postprocessing_backend: str, optional Use the 'cpu' or 'gpu' for postprocessing. Defaults to ``None``. chain_method: str, optional Method for drawing samples ('parallel' or 'vectorized'); defaults to 'parallel'. The 'vectorized' method should be used for sampling with a GPU (requires installing JAX with GPU support). seed: int, optional Random seed for sampler. save: bool, optional Whether to save the trace (to 'traces' subfolder in the current directory); defaults to ``True``. Returns ------- full_trace: arviz.InferenceData An :class:`arviz.InferenceData` object containing the full set of prior and posterior samples. """ if type(proxies) == str: proxies = list([proxies]) if 'sections' in kwargs: sections = list(kwargs['sections']) else: sections = np.unique(sample_df.dropna(subset = proxies, how = 'all')['section']) if save: # if folders for saving traces don't already exist in the current directory, make them dir_path = Path("traces/") dir_path_intermediate = Path("traces/temp/") if not dir_path.is_dir(): print('Creating directory for saved traces in ' + str(os.getcwd())) dir_path.mkdir() if not dir_path_intermediate.is_dir(): dir_path_intermediate.mkdir() if save_custom_initvals: dir_path = Path("initial_values/") if not dir_path.is_dir(): print('Creating directory for initial value dictionaries in ' + str(os.getcwd())) dir_path.mkdir() if len(asctime().split(" ")[3]) > 1: ind = 3 else: ind = 4 tstamp = ( str(date.today()) + "_at_" + asctime().split(" ")[ind].replace(":", "_") ) # this version of pymc automatically sets log_likelihood to false if not provided idata_kwargs = {'log_likelihood': True} if initvals is None: if generate_custom_initvals: initvals = make_initial_values_per_chain(sample_df, ages_df, model, n_chains = chains, proxies = proxies, seed = initval_seed, sections = sections) if save_custom_initvals: save_object(initvals, 'initial_values/' + 'initial_value_dictionary_' + tstamp) with model: if sampler == 'numpyro': full_trace = pm.sampling.jax.sample_numpyro_nuts(draws, tune=tune, chains=chains, target_accept=target_accept, postprocessing_vectorize='scan', postprocessing_backend = postprocessing_backend, random_seed = seed, idata_kwargs = idata_kwargs, nuts_kwargs = nuts_kwargs, initvals = initvals, chain_method = chain_method, jitter = False) elif sampler == 'blackjax': full_trace = pm.sampling.jax.sample_blackjax_nuts(draws, tune=tune, chains=chains, target_accept=target_accept, postprocessing_vectorize='scan', postprocessing_backend = postprocessing_backend, random_seed = seed, idata_kwargs = idata_kwargs, nuts_kwargs = nuts_kwargs, initvals = initvals, chain_method = chain_method, jitter = False) # nutpie sampler -- slower than numpyro and blackjax because some operations in the model (custom SortOp and AdvancedSetSubtensor) don't have a nopython implementation, which causes Numba to fall back to 'object mode' (slower python implementation) elif sampler == 'nutpie': nuts_kwargs['target_accept'] = target_accept nutpie_kwargs = dict(backend="numba", # numba or jax gradient_backend="pytensor") # jax or pytensor full_trace = pm.sample(draws=draws, tune=tune, chains=chains, nuts_sampler = 'nutpie', nuts=nuts_kwargs, compile_kwargs = nutpie_kwargs, idata_kwargs = idata_kwargs, random_seed = seed, ) if save: # save the trace with posterior samples -- if an error is encountered during sample_posterior_predictive, can re-load this file so we don't have to start over full_trace.to_netcdf("traces/temp/" + str(name) + '_' + tstamp + ".nc") if sample_predictive: v2r = ['f_pred_' + proxy for proxy in proxies] for proxy in proxies: if approximate: f_pred = gp[proxy].conditional('f_pred_' + proxy, Xnew=ages) else: f_pred = gp[proxy].conditional('f_pred_' + proxy, Xnew=ages, jitter = jitter) posterior_predictive = pm.sample_posterior_predictive( full_trace, var_names=v2r, return_inferencedata = True, random_seed = seed ) prior = pm.sample_prior_predictive(random_seed = seed, draws = prior_draws) full_trace.extend(prior) full_trace.extend(posterior_predictive) dataset = xr.Dataset() dataset = xr.Dataset({"X_new": ("time", ages.ravel())}) full_trace.add_groups(dataset) if save: full_trace.to_netcdf("traces/" + str(name) + '_' + tstamp + ".nc") # delete the temporary backup os.remove("traces/temp/" + str(name) + '_' + tstamp + ".nc") bad_chains, _, _ = check_inference(full_trace, sample_df, ages_df, quiet = True, sections = sections) if len(bad_chains) > 0: warnings.warn(f"Superposition violated in chains {bad_chains}. These chains were removed from the trace; the original trace (with the bad chains) is saved in the `traces` folder. To investigate the cause of the superposition violation, load the original trace and run the functions in the `stratmc.tests` module with `quiet = False`.") full_trace = drop_chains(full_trace, bad_chains) # save the clean version if save: full_trace.to_netcdf("traces/" + 'clean_' + str(name) + '_' + tstamp + ".nc") return full_trace
[docs] def make_initial_values_per_chain(sample_df, ages_df, model, n_chains, proxies = ['d13c'], seed = None, **kwargs): """ Generate valid (transformed) initial values for MCMC chains. Output can be passed to the ``initvals`` argument of :py:meth:`get_trace() <stratmc.inference.get_trace>`. Parameters ---------- sample_df: pandas.DataFrame :class:`pandas.DataFrame` with proxy data used during the inference step (as input to :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`). ages_df: pandas.DataFrame :class:`pandas.DataFrame` containing age constraints used during the inference step (as input to :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`). model: pymc.Model :class:`pymc.model.core.Model` object returned by :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`. n_chains: int Number of MCMC chains to generate initial values for. proxies: list(str) List of proxies included in the inference. sections: list(str) or numpy.array(str), optional List of sections included in the inference; only required if not all sections in ``sample_df`` were included. seed: int, optional Random seed used to sample the prior. Returns ------- initval_dicts: list(dict) List of dictionaries (one per chain) with valid initial values for each variable in ``model``; keys are variable names. """ if 'sections' in kwargs: sections = list(kwargs['sections']) else: sections = list(np.unique(sample_df['section'])) sample_df, ages_df = clean_data(sample_df, ages_df, proxies, sections) # ignore sections that have no observations included in the inference (samples w/ no observations were marked `Exclude? = True` in clean_data) data_sections = list(np.unique(sample_df[~sample_df['Exclude?']]['section'])) for section in list(sections): if section not in data_sections: sections.remove(section) # clean again (avoids issues w/ offset and noise groups) sample_df, ages_df = clean_data(sample_df, ages_df, proxies, sections) # check that for all constraints with 'shared == True', the constraint actually is used >1 time (if not, set shared = False) for shared_age_name in ages_df[ages_df['shared?']==True]['name']: if ages_df[(ages_df['shared?'] == True) & (ages_df['name']==shared_age_name)].shape[0] < 2: idx = ages_df[(ages_df['shared?'] == True) & (ages_df['name']==shared_age_name)].index ages_df.loc[idx, 'shared?'] = False ## step 1: draw from prior (1 draw per chain) with warnings.catch_warnings(): warnings.simplefilter("ignore", category=UserWarning) with model: print('Drawing initial values from model prior') prior = pm.sample_prior_predictive(random_seed = seed, draws = n_chains) ## step 2: store initial values for each variable listed in model.initial_point() # grab list of variables with initial values # note - use rvs_to_initial_values instead of rvs_to_values (which includes observed 'proxy_pred' variables) init_var_names = [] for i in range(len(list(model.rvs_to_initial_values.keys()))): init_var_names.append(list(model.rvs_to_initial_values.keys())[i].name) # list of dictionaries of transformed initial values initval_dicts = [] for i in range(n_chains): chain_dict = {} for var in init_var_names: prior_vals = az.extract(prior.prior)[var].values if len(prior_vals.shape) == 1: # draw i chain_dict[var] = prior_vals[i] else: # all dims, draw i chain_dict[var] = prior_vals[:, i] for section in sections: # check whether the initial ages violate superpostiion, and change if necessary chain_dict = get_valid_initvals_per_chain(chain_dict, sample_df, ages_df, sections) initval_dicts.append(chain_dict) return initval_dicts
[docs] def get_valid_initvals_per_chain(initval_dict, sample_df, ages_df, sections): """ Helper function for generating valid initial values for a single MCMC chain. Called in :py:meth:`make_initial_values_per_chain() <stratmc.inference.make_initial_values_per_chain>`. Parameters ---------- initval_dict: dict Dictionary of proposed initial values. sample_df: pandas.DataFrame :class:`pandas.DataFrame` with proxy data used during the inference step (as input to :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`). ages_df: pandas.DataFrame :class:`pandas.DataFrame` containing age constraints used during the inference step (as input to :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`). sections: list(str) or numpy.array(str) List of sections included in the inference. Returns ------- initval_dict: list of dict Dictionary with initial values modified to respect superposition; keys are variable names. """ # figure out which variable names go with each section for section in sections: section_df = sample_df[sample_df['section']==section] # separate dataframes for intermediate detrital or intrusive constraints and depositional age constraints section_age_df = ages_df[(ages_df['section']==section) & (~ages_df['intermediate detrital?']) & (~ages_df['intermediate intrusive?']) & (~ages_df['depositional?'])] intermediate_detrital_section_ages_df = ages_df[(ages_df['section']==section) & (ages_df['intermediate detrital?'])] intermediate_intrusive_section_ages_df = ages_df[(ages_df['section']==section) & (ages_df['intermediate intrusive?'])] section_ages = section_age_df['age'].values age_heights = section_age_df['height'].values # if all age constraint priors are Gaussian and not shared, then superposition was enforced with an ordered transform if all(section_age_df['distribution_type']=='Normal') and all(section_age_df['shared?']==False): label = str(section) +'_' ordered_age_dist_name = label + 'flip_radiometric_age' # ensures that radiometric ages are in order -- ordered transform not applied during forward (i.e., prior predictive) sampling initval_dict = ordered_transform_in_prior(initval_dict, ordered_age_dist_name) # otherwise, superposition was enforced using potentials applied to a group of independent prior distributions else: age_dist_names = [] for i in np.arange(len(section_ages)): label = str(section) + '_' + str(i) + '_' constraint_shared = section_age_df['shared?'].values[i] # for shared constraints, link to existing distribution if constraint_shared == True: shared_constraint_name = section_age_df['name'].values[i] age_dist_names.append(shared_constraint_name) # if the constraint is not shared, build a new distribution else: age_dist_names.append(label + 'radiometric_age') # make sure that superposition between depositional age constraints is respected initval_dict = superposition_from_dict(initval_dict, age_dist_names, section_age_df) for interval in np.arange(0, len(age_heights)-1).tolist(): label = str(section)+'_'+ str(interval) +'_' above = section_df['height']>=age_heights[interval] below = section_df['height']<age_heights[interval+1] interval_df = section_df[above & below] interval_heights = interval_df['height'].values above = intermediate_detrital_section_ages_df['height']>age_heights[interval] below = intermediate_detrital_section_ages_df['height']<age_heights[interval+1] detrital_interval_df = intermediate_detrital_section_ages_df[above & below] above = intermediate_intrusive_section_ages_df['height']>age_heights[interval] below = intermediate_intrusive_section_ages_df['height']<age_heights[interval+1] intrusive_interval_df = intermediate_intrusive_section_ages_df[above & below] if len(interval_heights) > 0: # make list of detrital age distribution names detrital_age_dist_names = [] detrital_age_dist_names_radio = [] if detrital_interval_df.shape[0] > 0: # enforce detrital ages -- note that initial values will be set base --> top for i in np.arange(detrital_interval_df.shape[0]): # construct DZ age prior constraint_shared = detrital_interval_df['shared?'].values[i] if constraint_shared == True: shared_constraint_name = detrital_interval_df['name'].values[i] intermediate_detrital_age_dist_name = shared_constraint_name else: intermediate_detrital_age_dist_name = label + 'detrital_age_' + str(i) # if there are overlying samples in interval, enforce maximum age (add name to list) if len(interval_df[interval_df['height']>=detrital_interval_df['height'].values[i]]['height'].values)>0: detrital_age_dist_names.append(intermediate_detrital_age_dist_name) detrital_age_dist_names_radio.append(intermediate_detrital_age_dist_name) # if there are no samples above the current detrital constraint, only add its name to list of detritals that apply to overlying depositional age constraint else: detrital_age_dist_names_radio.append(intermediate_detrital_age_dist_name) # make list of intrusive age distribution names intrusive_age_dist_names = [] intrusive_age_dist_names_radio = [] if intrusive_interval_df.shape[0] > 0: # enforce intrusive ages -- note that initial values will be set base --> top for i in np.arange(intrusive_interval_df.shape[0]): # if there are underlying samples in interval, enforce maximum age # construct intrusive age prior constraint_shared = intrusive_interval_df['shared?'].values[i] if constraint_shared == True: shared_constraint_name = intrusive_interval_df['name'].values[i] intermediate_intrusive_age_dist_name = shared_constraint_name else: intermediate_intrusive_age_dist_name = label + 'intrusive_age_' + str(i) # if there are underlying samples in interval, enforce minimum age (add name to list) if len(interval_df[interval_df['height']<=intrusive_interval_df['height'].values[i]]['height'].values)>0: intrusive_age_dist_names.append(intermediate_intrusive_age_dist_name) intrusive_age_dist_names_radio.append(intermediate_intrusive_age_dist_name) # if there are no samples below the current intrusive constraint, only add its name to list of detritals that apply to overlying depositional age constraint else: intrusive_age_dist_names_radio.append(intermediate_intrusive_age_dist_name) # if there are limiting age constraints in the section, check them. if not, we can leave the initvals as-is if (detrital_interval_df.shape[0] > 0) or (intrusive_interval_df.shape[0] > 0): if all(section_age_df['distribution_type']=='Normal') and all(section_age_df['shared?']==False): # THESE WILL BE UPSIDE DOWN -- inside of the intrusive potential, use this name, but flip the initial values # with np.flip() THEN index with [interval] and [interval+1] base_age_dist_name = str(section) +'_' + 'flip_radiometric_age' upper_age_dist_name = str(section) +'_' + 'flip_radiometric_age' shared_radiometric_age_dist = True base_age_idx = interval top_age_idx = interval + 1 else: base_age_dist_name = age_dist_names[interval] upper_age_dist_name = age_dist_names[interval + 1] shared_radiometric_age_dist = False base_age_idx = None top_age_idx = None # enforce superposition between basal age constraint and intrusive ages if len(intrusive_age_dist_names_radio) > 0: # NOTE: base_age_dist has to be flipped before using index initval_dict = superposition_depositional_and_limiting_ages_from_dict(initval_dict, base_age_dist_name, [], intrusive_age_dist_names_radio, depositional_age_idx = base_age_idx) if len(detrital_age_dist_names_radio) > 0: # enforce superposition between top age constriant and detrital ages # NOTE: upper_age_dist has to be flipped before using index inside of function initval_dict = superposition_depositional_and_limiting_ages_from_dict(initval_dict, upper_age_dist_name, detrital_age_dist_names_radio, [], depositional_age_idx = top_age_idx) if (len(detrital_age_dist_names) > 0) or (len(intrusive_age_dist_names) > 0): if len(interval_heights) > 1: age_label = 'unsorted_random_ages' else: age_label = 'random_ages' initval_dict = get_valid_initial_sample_ages_from_dict(initval_dict, detrital_age_dist_names, intrusive_age_dist_names, base_age_dist_name, upper_age_dist_name, label + age_label, interval_df['height'].values, detrital_interval_df['height'].values, intrusive_interval_df['height'].values, interval, sf1_name = label + 'scaling_factor_1', sf2_name = label + 'scaling_factor_2', shared_radiometric_age_dist = shared_radiometric_age_dist) return initval_dict
[docs] def ordered_transform_in_prior(initval_dict, ordered_dist_name): """ Helper function to enforces ordered transform in dictionary of prior draws. Parameters ---------- initval_dict: dict Dictionary of proposed initial values. ordered_dist_name: Names of distribution with ordered transform in model. Returns ------- initval_dict: dict Dictionary with sorted (increasing) initial values for ``ordered_dist_name``. """ if not all(np.diff(initval_dict[ordered_dist_name]) >= 0): initval_dict[ordered_dist_name] = np.sort(initval_dict[ordered_dist_name]) return initval_dict
[docs] def superposition_from_dict(initval_dict, age_dist_names, section_age_df): """ Modify dictionary of initial values such that superposition between depositional age constraints is respected. Replicates logic used to set model initial point in :py:meth:`superposition() <stratmc.model.superposition>` in :py:mod:`stratmc.model`. Parameters ---------- initval_dict: dict Dictionary of proposed initial values. age_dist_names: Names of radiometric age distributions in ``model`` (must be in stratigraphic order - lowest to highest). section_age_df: section_age_df: pandas.DataFrame :class:`pandas.DataFrame` containing age constraints for current section. Returns ------- initval_dict: dict Dictionary with initial values modified to respect superposition; keys are variable names. """ for i in np.arange(0, len(section_age_df['height'])-1).tolist(): lower_label = age_dist_names[i] upper_label = age_dist_names[i + 1] # note that model.initial_point returns transformed values, while initival values provided to sampler are untransformed (so ages can be provided as-is) base_age_initval = initval_dict[lower_label] upper_age_initval = initval_dict[upper_label] age_diff = base_age_initval - upper_age_initval # if model is starting out of superposition, make upper sample younger by 1 if age_diff < 0: initval_dict[upper_label] = base_age_initval - 1 return initval_dict
[docs] def superposition_depositional_and_limiting_ages_from_dict(initval_dict, depositional_age_name, detrital_age_dist_names, intrusive_age_dist_names, depositional_age_idx = None): """ Modify dictionary of initial values such that superposition between limiting and depositional age constraints is respected. Replicates logic used to set model initial point in :py:meth:`superposition_depositional_and_limiting_ages() <stratmc.model.superposition_depositional_and_limiting_ages>` in :py:mod:`stratmc.model`. Parameters ---------- initval_dict: dict Dictionary of proposed initial values. depositional_age_name: str Name of distribution for target depositional age constraint in ``model``. detrital_age_dist_names: list(str) Names of underlying detrital age constraint distributions in ``model``. intrusive_age_dist_names: list(str) Names of overlying intrusive age constraint distributions in ``model``. depositional_age_idx: int, optional Position of ``depositional_age`` in model variable ``depositional_age_name``. Only required if ``depositional_age`` is one of multiple ages modeled using a single multidimensional distribution. Returns ------- initval_dict: dict Dictionary with initial values modified to respect superposition; keys are variable names. """ # grab initial value of depositional age constraint depositional_age_initval = initval_dict[depositional_age_name] if depositional_age_idx is not None: depositional_age_initval = np.flip(depositional_age_initval)[depositional_age_idx] # loop over detrital age constraints, adding potentials to enforce superposition with depositional age for detrital_var in detrital_age_dist_names: # check initval superposition for detrital age. if the detrital age is violated (i.e., the depositional age is older), set the detrital age initval to be 1 Myr older than the current depositional age initval # note - not modifying depositional age initval, because initvals have already been set in superposition_from_dict() to enforce superposition between depositional ages detrital_age_initval = initval_dict[detrital_var] if depositional_age_initval > detrital_age_initval: initval_dict[detrital_var] = depositional_age_initval + 1 # loop over intrusive age constraints, adding potentials to enforce superposition with depositional age for intrusive_var in intrusive_age_dist_names: # check initval superposition for intrusive age. if the intrusive age is violated (i.e., the depositional age is younger), set the intrusive age initval to be 1 Myr younger than the current depositional age initval intrusive_age_initval = initval_dict[intrusive_var] if depositional_age_initval < intrusive_age_initval: initval_dict[intrusive_var] = depositional_age_initval - 1 return initval_dict
[docs] def get_valid_initial_sample_ages_from_dict(initval_dict, detrital_age_dist_names, intrusive_age_dist_names, maximum_age_dist_name, minimum_age_dist_name, sample_age_dist_name, sample_heights, detrital_heights, intrusive_heights, interval, sf1_name, sf2_name, shared_radiometric_age_dist): """ Modify dictionary of initial values such that all detrital and intrusive age constraints are respected. Replicates logic used to set model initial point in :py:meth:`get_valid_initial_ages() <stratmc.model.get_valid_initial_ages>` in :py:mod:`stratmc.model`. Parameters ---------- initval_dict: dict Dictionary of proposed initial values. detrital_age_dist_names: list(str) List of names for detrital age constraint distributions in ``model``. intrusive_age_dist_names: list(str) List of names for intrusive age constraint distributions in ``model``. maximum_age_dist_name: str Name of distribution for underlying maximum age constraint in ``model``. minimum_age_dist_name: str Name of distribution for overlying minimum age constraint in ``model``. sample_age_dist_name: str Name of sample age distribution (unsorted and unscaled) in the pymc.Model object. sample_heights: np.array Array of heights for samples in the current interval. detrital_heights: list(float) Heights of detrital age constraints. intrusive_heights: list(float) Heights of intrusive age constraints. interval: int Current interval number. sf1_name: str Name of the distribution associated with scaling factor 1 in ``model``. sf2_name: str Name of the distribution associated with scaling factor 2 in ``model``. shared_radiometric_age_dist: bool, optional Whether the radiometric age distributions are part of a single object (versus initiated as separate distributions). Defaults to ``True``. Returns ---------- initval_dict: dict Dictionary of valid initial values; keys are model variables. """ ## step 1: reset sf1 and sf2 such that it's possible for all of the age constraints to be respected (i.e., ages that are younger than the oldest detrital age, and older than the youngest intrusive age, must be inside of the 'box' of possible ages) # get initial values for sample ages # note - samples from prior are untransformed (no need to apply backward transform) sample_age_initval = initval_dict[sample_age_dist_name] # sort initial values (lower likelihood that we'll have to alter the initial values, and values will be sorted inside of model anyway) sample_age_initval = np.sort(sample_age_initval) # smallest to largest = oldest to youngest (base to top) # get initial value for underlying minimum age constraint base_age_initval = initval_dict[maximum_age_dist_name] if shared_radiometric_age_dist: if len(base_age_initval) > 1: # radiometric ages modeled using multidimensional distribution with ordered transform are upside down --> flip before indexing base_age_initval = np.flip(base_age_initval)[interval] # get initial value for overlying minimum age constraint upper_age_initval = initval_dict[minimum_age_dist_name] if shared_radiometric_age_dist: if len(upper_age_initval) > 1: # radiometric ages modeled using multidimensional distribution with ordered transform are upside down --> flip before indexing upper_age_initval = np.flip(upper_age_initval)[interval+1] # set initial values for sample ages # get initial values for scaling factors sf1_initval = initval_dict[sf1_name] sf2_initval = initval_dict[sf2_name] # calculate the bounds of the current 'age box', given sf1 and sf2 current_max = base_age_initval - (1 - sf2_initval) * (base_age_initval - upper_age_initval) * (1 - sf1_initval) current_min = base_age_initval - ((base_age_initval - upper_age_initval) * sf1_initval) - (1 - sf2_initval) * (base_age_initval - upper_age_initval) * (1 - sf1_initval) scaled_dz_initvals = [] scaled_intrusive_initvals = [ ] # iterate over detrital ages # get initial value for detrital age constraints for detrital_age_dist_name in detrital_age_dist_names: dz_age_initval = initval_dict[detrital_age_dist_name] # calculate DZ age on [0, 1] scale defined by bounds scaled_dz_initvals.append((current_max - dz_age_initval)/(current_max - current_min)) # iterate over intrusive ages for intrusive_age_dist_name in intrusive_age_dist_names: intrusive_age_initval = initval_dict[intrusive_age_dist_name] # calculate intrusive age on [0, 1] scale defined by bounds scaled_intrusive_initvals.append((current_max - intrusive_age_initval)/(current_max - current_min)) scaled_dz_initvals = np.array(scaled_dz_initvals) scaled_intrusive_initvals = np.array(scaled_intrusive_initvals) ## step 2: check that the initial sample ages satisfy all of the limiting age constraints. if not, re-draw so they do # if any limiting age constraints are precluded by the current scale and shift parameters, reset until all age constraints can be respected # if a detrital age is > 1, then it's impossible for sample ages to be younger # if an intrusive age is < 0, then it's impossible for sample age to be younger if (any(scaled_dz_initvals > 1)) or (any(scaled_intrusive_initvals < 0)): # re-draw scale and shift parameters from U[0, 1] until all age constraints can be satisfied while (any(scaled_dz_initvals > 1)) or (any(scaled_intrusive_initvals < 0)): sf1_initval = np.random.uniform(0, 1, 1) sf2_initval = np.random.uniform(0, 1, 1) current_max = base_age_initval - (1 - sf2_initval) * (base_age_initval - upper_age_initval) * (1 - sf1_initval) current_min = base_age_initval - ((base_age_initval - upper_age_initval) * sf1_initval) - (1 - sf2_initval) * (base_age_initval - upper_age_initval) * (1 - sf1_initval) for i, detrital_age_dist_name in enumerate(detrital_age_dist_names): dz_age_initval = initval_dict[detrital_age_dist_name] scaled_dz_initvals[i] = (current_max - dz_age_initval)/(current_max - current_min) for i, intrusive_age_dist_name in enumerate(intrusive_age_dist_names): intrusive_age_initval = initval_dict[intrusive_age_dist_name] scaled_intrusive_initvals[i] = (current_max - intrusive_age_initval)/(current_max - current_min) # calculate final scaled initial values (necessary if while statement was satisfied before all values were re-calculated) for i, detrital_age_dist_name in enumerate(detrital_age_dist_names): dz_age_initval = initval_dict[detrital_age_dist_name] scaled_dz_initvals[i] = (current_max - dz_age_initval)/(current_max - current_min) for i, intrusive_age_dist_name in enumerate(intrusive_age_dist_names): intrusive_age_initval = initval_dict[intrusive_age_dist_name] scaled_intrusive_initvals[i] = (current_max - intrusive_age_initval)/(current_max - current_min) # set scale/shift initial values in dictionary initval_dict[sf1_name] = sf1_initval initval_dict[sf2_name] = sf2_initval ## step 3: make sure the initial sample age values respect all limiting age constraints. otherwise, mcmc sampler will start in a low-probability space with a flat likelihood # only detrital ages: set initial values, base top top (oldest to youngest) # note - oldest to youngest should always be base top (a higher and older DZ would be useless) if (len(detrital_age_dist_names) > 0) and (len(intrusive_age_dist_names) == 0): for i, detrital_age_dist_name in enumerate(detrital_age_dist_names): overlying_sample_idx = np.where(sample_heights >= detrital_heights[i])[0] # if scaled initial value is > 0, use it to truncate the age box # is scaled initial value is <= 0, then all samples already must be younger than the detrital age --> do nothing if (scaled_dz_initvals[i] > 0): # only re-draw if current (sorted) initvals don't satisfy the constraint (must be younger = larger initval) if not all(sample_age_initval[overlying_sample_idx] > scaled_dz_initvals[i]): scaled_age_initial_points = np.random.uniform(scaled_dz_initvals[i], 1, len(overlying_sample_idx)) sample_age_initval[overlying_sample_idx] = np.sort(scaled_age_initial_points) # only intrusive ages: set initial values youngest to oldest (top to base) # note - youngest to oldest should always be top to base (a lower and younger intrusive would be useless) elif (len(detrital_age_dist_names) == 0) and (len(intrusive_age_dist_names) > 0): for i, intrusive_age_dist_name in enumerate(np.flip(intrusive_age_dist_names)): underlying_sample_idx = np.where(sample_heights <= np.flip(intrusive_heights)[i])[0] # if scaled initival value is <1, use it to truncate the age box # scaled initial value could be >=1, which means all the values in the age box already respect the constraint --> do nothing (initial values will already respect constraint) if np.flip(scaled_intrusive_initvals)[i] < 1: # only re-draw if current (sorted) initvals don't satisfy the constraint (must be older = smaller initval) if not all(sample_age_initval[underlying_sample_idx] < np.flip(scaled_intrusive_initvals)[i]): scaled_age_initial_points = np.random.uniform(0, np.flip(scaled_intrusive_initvals)[i], len(underlying_sample_idx)) sample_age_initval[underlying_sample_idx] = np.sort(scaled_age_initial_points) # combination of detrital and an intrusive ages: elif (len(detrital_age_dist_names) > 0) and (len(intrusive_age_dist_names) > 0): # iterate over intrusive ages from top to base # note - iterating over indices in reverse (top to base/high to low), so no need to flip intrusive variables inside of loop for idx in np.flip(np.arange(len(intrusive_heights))): # grab detrital ages below the current intrusive age detrital_idx = np.where(detrital_heights <= intrusive_heights[idx])[0] # set initvals in chunks bracketed by [detrital age, intrusive age], starting with the lowermost detrital age for d_idx in detrital_idx: # grab indices of samples in between the current intrusive and detrital constraints sample_idx = np.where((sample_heights <= intrusive_heights[idx]) & (sample_heights >= detrital_heights[d_idx]))[0] # if initial sample ages don't fall in between these constraints, re-draw initvals if not (all(sample_age_initval[sample_idx] > scaled_dz_initvals[d_idx]) and all(sample_age_initval[sample_idx] < scaled_intrusive_initvals[idx])): scaled_age_initial_points = np.random.uniform(np.max(np.concatenate([scaled_dz_initvals[d_idx], np.array([0])])), np.min(np.concatenate([scaled_intrusive_initvals[idx], np.array([1])])), len(sample_idx)) sample_age_initval[sample_idx] = np.sort(scaled_age_initial_points) # if there aren't detrital ages below the current intrusive age, set initial values using only the intrusive constraint if len(detrital_idx) == 0: underlying_sample_idx = np.where(sample_heights <= intrusive_heights[idx])[0] if scaled_intrusive_initvals[idx] < 1: # only re-draw if current values don't satisfy age constraint if not all(sample_age_initval[underlying_sample_idx] < scaled_intrusive_initvals[idx]): scaled_age_initial_points = np.random.uniform(0, # no limit on how old scaled_intrusive_initvals[idx], # older than intrusive len(underlying_sample_idx)) sample_age_initval[underlying_sample_idx] = np.sort(scaled_age_initial_points) # set all samples below the lowest detrital age to be older than the intrusive age, and also older than all the samples above the detrital age. otherwise, if initvals are actually younger than the overlying samples, then the intrusive age could be violated when the initvals are sorted # note - doing for every intrusive age b/c we don't know where the lowest detrital age is located (ensures that samples between the lowest DZ and any underlying intrusives have valid initial ages) if len(detrital_idx) > 0: # grab samples below the highest intrusive age and above the lowest detrital age # note -- not checking above the uppermost intrusive because we haven't yet checked the initvals for these samples (their ages will be set last, and must be younger than all underlying samples) sample_idx_above = np.where((sample_heights <= intrusive_heights[-1]) & (sample_heights >= detrital_heights[detrital_idx[0]]))[0] # grab samples below the lowest DZ sample_idx_below = np.where(sample_heights < detrital_heights[detrital_idx[0]])[0] if len(sample_idx_below) > 0: # if all of the samples below the lowest DZ aren't older than all overlying samples, re-draw if not all(sample_age_initval[sample_idx_below] < np.min(sample_age_initval[sample_idx_above])): # ages need to be both older than the lowermost intrusive age, and older than the oldest sample above the detrital age scaled_age_initial_points = np.random.uniform(0, # no limit on max age np.min(sample_age_initval[sample_idx_above]), # min age = max age (lowest initval) of overlying sample group. intrusive ages enforced as a byproduyct, since all of the underlying sample ages must be older len(sample_idx_below)) sample_age_initval[sample_idx_below] = np.sort(scaled_age_initial_points) # first, set everything above the highest intrusive to be younger than all underlying samples (which now have been valid initvals). then, work way up and enforce any overlying DZs # necessary because if these initvals are older than the underlying samples, sorting would push our previously set initial values stratigraphically higher (relative to the age constraints), potentially making them invalid # note: not a problem if we only have intrusive limiting ages # grab samples above uppermost intrusive overlying_sample_idx = np.where(sample_heights > np.max(intrusive_heights))[0] # grab underlying samples -- all of these initvals were already set in previous loop sample_below_idx = np.where(sample_heights <= np.max(intrusive_heights))[0] if len(overlying_sample_idx) > 0: # if all overlying samples aren't already younger (larger initvals) than the youngest underlying sample, re-draw values if not all(sample_age_initval[overlying_sample_idx] > np.max(sample_age_initval[sample_below_idx])): scaled_age_initial_points = np.random.uniform(np.max(sample_age_initval[sample_below_idx]), # youngest underlying sample = max age 1, # no limit on how young len(overlying_sample_idx)) sample_age_initval[overlying_sample_idx] = np.sort(scaled_age_initial_points) # iterate over detrital ages that are stratigraphically higher than the uppermost (youngest) intrusive age. these initial values can be set in the same way as the 'dz-only' samples (set initial values base --> top) # note - samples in between highest intrusive and overlying DZ already taken care of by previous loop detrital_only_idx = np.where(detrital_heights > np.max(intrusive_heights))[0] for idx in detrital_only_idx: overlying_sample_idx = np.where(sample_heights >= detrital_heights[idx])[0] sample_below_idx = np.where(sample_heights < detrital_heights[idx])[0] # if scaled initial value is > 0, use to truncate the age box # is scaled initial value is <= 0, then all samples already must be younger than the detrital age --> do nothing if scaled_dz_initvals[idx] > 0: # print(scaled_dz_initvals[idx]) # if current initvals aren't younger (larger) than both the detrital constraint and all underlying ages, re-draw if not all((sample_age_initval[overlying_sample_idx] > scaled_dz_initvals[idx]) & (sample_age_initval[overlying_sample_idx] > np.max(sample_age_initval[sample_below_idx]))): # also need to be younger than all underlying samples lower_bound = np.max(np.concatenate([scaled_dz_initvals[idx], np.max(sample_age_initval[sample_below_idx]).ravel()])) scaled_age_initial_points = np.random.uniform(lower_bound, # younger than detrital age or youngest underlying sample (whichever is younger) 1, # no limit on how young len(overlying_sample_idx)) sample_age_initval[overlying_sample_idx] = np.sort(scaled_age_initial_points) initval_dict[sample_age_dist_name] = sample_age_initval return initval_dict
[docs] def extend_age_model(full_trace, sample_df, ages_df, new_proxies, new_proxy_df = None, **kwargs): """ Extend age models to a different set of proxy observations using posterior sample ages from an existing inference. For instance, extend an age model built using C isotope data to new S isotope data collected from different stratigraphic horizons in the same sections. Note that the age of stratigraphic horizons that were included in ``sample_df`` (but marked ``Exclude? = True``) is already passively tracked within the model; this function is only required to estimate the age of observations that were not in ``sample_df`` when the inference was run. To estimate ages for new measurements of the same proxy, place the new data in a different column (e.g., 'd13c_new`). Parameters ---------- full_trace: arviz.InferenceData An :class:`arviz.InferenceData` object containing the full set of prior and posterior samples from :py:meth:`get_trace() <stratmc.inference.get_trace>` in :py:mod:`stratmc.inference`. sample_df: pandas.DataFrame :class:`pandas.DataFrame` with proxy data used during the inference step (as input to :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`). ages_df: pandas.DataFrame :class:`pandas.DataFrame` containing age constraints used during the inference step (as input to :py:meth:`build_model() <stratmc.model.build_model>` in :py:mod:`stratmc.model`). new_proxies: str or list(str) New proxy(s) to construct age models for. new_proxy_df: pandas.DataFrame, optional :class:`pandas.DataFrame` containing new proxy observations. Optional; if not provided, uses ``sample_df`` (assumes that observations for the new proxy are in the same DataFrame as the original proxy observations). sections: list(str) or numpy.array(str), optional List of sections included in the inference; only required if not all sections in ``sample_df`` were included. Returns ------- interpolated_df: pandas.DataFrame :class:`pandas.DataFrame` with interpolated age draws and sample age summary statistics (maximum likelihood estimate, median, and 68% and 95% confidence intervals) for each new proxy observation. """ # get list of proxies included in model from full_trace variables = [ l for l in list(full_trace["posterior"].data_vars.keys()) if (f"{'gp_ls_'}" in l) and (f"{'unshifted'}" not in l) ] proxies = [] for var in variables: proxies.append(var[6:]) if 'sections' in kwargs: sections = list(kwargs['sections']) else: sections = np.unique(sample_df.dropna(subset = proxies, how = 'all')['section']) if type(new_proxies) == str: new_proxies = list([new_proxies]) if new_proxy_df is None: new_proxy_df = sample_df.copy() # only keep samples that were included in the inference sample_df, ages_df = clean_data(sample_df, ages_df, proxies, sections) new_proxy_df, _ = clean_data(new_proxy_df, ages_df, new_proxies, sections) new_proxy_df = new_proxy_df[~new_proxy_df['Exclude?']] interp_df = pd.DataFrame(columns = list(new_proxy_df.columns) + ['interp']) common_sections = [section for section in sections if section in np.unique(new_proxy_df['section'].values)] print('Interpolating section age models') for section in tqdm(common_sections): # height of samples included in inference section_df = sample_df[sample_df['section']==section] sample_heights = np.concatenate([section_df[~np.isnan(section_df[proxy])]['height'].values for proxy in proxies]) sample_heights = np.sort(np.unique(sample_heights)) # heights at which to interpolate age models new_section_df = new_proxy_df[new_proxy_df['section']==section] new_heights = np.concatenate([new_section_df[~np.isnan(new_section_df[proxy])]['height'].values for proxy in new_proxies]) new_heights = np.sort(np.unique(new_heights)) new_heights = np.sort(new_heights) # heights of radiometric age constraints section_ages_df = ages_df[(ages_df['section']==section) & (~ages_df['Exclude?']) & (~ages_df['intermediate detrital?']) & (~ages_df['intermediate intrusive?'])] age_heights = section_ages_df['height'].values # heights at which to interpolate age models # just keep all heights, even if included in the original inference - improves workflow for plotting interpolated proxy signals interp_heights = [h for h in new_heights] # if h not in sample_heights] cols = list(np.unique(list(new_proxy_df.columns))) interp_section_df = pd.DataFrame(columns = cols + ['interp']) for h in interp_heights: idx = new_proxy_df[new_proxy_df['height']==h].index.tolist() # this has to be merge because we've added new columns interp_section_df = pd.concat([interp_section_df, new_proxy_df.loc[idx]]) interp_section_df['interp'] = 'y' interp_section_df.reset_index(inplace = True, drop = True) # sample age posterior - shape (samples x draws) sample_age_post = az.extract(full_trace.posterior)[str(section) + '_ages'].values age_constraint_post = az.extract(full_trace.posterior)[str(section) + '_radiometric_age'].values if sample_age_post.shape[0] != len(sample_heights): sys.exit(f"Number of data points for {section} does not match the number of data points in the trace. Check that input data and list of proxies match.") if age_constraint_post.shape[0] != len(age_heights): sys.exit(f"Number of data points for {section} does not match the number of data points in the trace. Check that input data and list of proxies match.") # combine position data for all samples + age constraints included in the inference all_heights = np.concatenate([sample_heights, age_heights]) sorted_idx = np.argsort(all_heights) # construct age and height vectors using posteriors for 1) samples in section, and 2) age constraints ages_comb = np.vstack([sample_age_post, age_constraint_post]) age_paths = ages_comb[sorted_idx, :] interp_section_ages = np.ones((len((interp_section_df.index.tolist())), age_paths.shape[1])) * np.nan # for each posterior draw, extend age model to new stratigraphic horizons for i in np.arange(age_paths.shape[1]): interp_section_ages[:, i] = np.interp(interp_section_df['height'], all_heights[sorted_idx], age_paths[:, i]) # store draws for each horizon interp_section_df['age_draws'] = np.nan interp_section_df['age_draws'] = interp_section_df['age_draws'].astype(object) for i in interp_section_df.index.tolist(): # shape samples x draws interp_section_df.at[i, 'age_draws'] = list(interp_section_ages[i, :]) interp_df = pd.concat([interp_df, interp_section_df]) interp_df.sort_values(by = ['section', 'height'], inplace = True) interp_df.reset_index(inplace = True, drop = True) interp_df['mle'] = np.nan interp_df['2.5'] = np.nan interp_df['16'] = np.nan interp_df['50'] = np.nan interp_df['84'] = np.nan interp_df['97.5'] = np.nan for i in interp_df.index.tolist(): current_ages = interp_df['age_draws'].loc[i] # mle dy = np.linspace(np.min(current_ages), np.max(current_ages), 2000) max_like = dy[np.argmax(gaussian_kde(current_ages, bw_method = 1)(dy))] interp_df.loc[i, 'mle'] = max_like # median interp_df.loc[i, '50'] = np.percentile(current_ages, 50) # 2.5% interp_df.loc[i, '2.5'] = np.percentile(current_ages, 2.5) # 16% interp_df.loc[i, '16'] = np.percentile(current_ages, 16) # 84% interp_df.loc[i, '84'] = np.percentile(current_ages, 84) # 97.5% interp_df.loc[i, '97.5'] = np.percentile(current_ages, 97.5) return interp_df
[docs] def interpolate_proxy(interp_df, proxy, ages): """ Use interpolated sample ages from :py:meth:`extend_age_model() <stratmc.inference.extend_age_model>` to calculate proxy values at a given set of ages (e.g., to plot 68 and 95% confidence intervals over time for a new proxy using :py:meth:`interpolated_proxy_inference() <stratmc.plotting.interpolated_proxy_inference>` in :py:mod:`stratmc.plotting`). Parameters ---------- interp_df: pandas.DataFrame :class:`pandas.DataFrame` with proxy data and interpolated ages from :py:meth:`extend_age_model() <stratmc.inference.extend_age_model>`. proxy: str Tracer to interpolate. ages: list(float) or numpy.array(float) Target ages at which to interpolate proxy values. Returns ------- interpolated_proxy_df: pandas.DataFrame :class:`pandas.DataFrame` with interpolated proxy values and summary statistics (maximum likelihood estimate, median, and 68% and 95% confidence intervals) at each age in ``ages``. """ age_paths = np.vstack(np.asarray(interp_df['age_draws'])) proxy_vec = interp_df[proxy] # iterate over draws print('Interpolating proxy values') for i in tqdm(np.arange(age_paths.shape[1])): current_order = np.argsort(age_paths[:, i]) current_path = age_paths[current_order, i] current_interp_proxy = np.interp(ages, current_path, proxy_vec[current_order]) if i == 0: interp_proxy = current_interp_proxy else: interp_proxy = np.vstack((interp_proxy, current_interp_proxy)) interpolated_proxy_df = pd.DataFrame(columns = ['age', proxy + '_draws', 'mle', '2.5', '16', '50', '84', '97.5']) interpolated_proxy_df[proxy + '_draws'] = interpolated_proxy_df[proxy + '_draws'].astype(object) # iterate over ages print('Calculating summary statistics') for i in tqdm(np.arange(len(ages))): # shape = draws x ages current_proxy = interp_proxy[:, i] dy = np.linspace(np.min(current_proxy), np.max(current_proxy), 200) max_like = dy[np.argmax(gaussian_kde(current_proxy, bw_method = 1)(dy))] current_mle = max_like temp_df = pd.DataFrame({'age': ages[i], proxy + '_draws': [current_proxy], 'mle': current_mle, '2.5': np.percentile(current_proxy, 2.5), '16': np.percentile(current_proxy, 16), '50': np.percentile(current_proxy, 50), '84': np.percentile(current_proxy, 84), '97.5': np.percentile(current_proxy, 97.5), }) interpolated_proxy_df = pd.concat([interpolated_proxy_df, temp_df]) return interpolated_proxy_df
[docs] def age_range_to_height(full_trace, sample_df, ages_df, lower_age, upper_age, **kwargs): """ Use the posterior age model for each section to find the stratigraphic interval (with uncertainty) corresponding to a given age range. If ``sections`` is not provided, returns height estimates for every section that overlaps the target age range. To visualize the stratigraphic intervals, use :py:meth:`section_age_range() <stratmc.plotting.section_age_range>` in :py:mod:`stratmc.plotting`. Parameters ---------- full_trace: arviz.InferenceData An :class:`arviz.InferenceData` object containing the full set of prior and posterior samples from :py:meth:`get_trace() <stratmc.inference.get_trace>` in :py:mod:`stratmc.inference`. sample_df: pandas.DataFrame :class:`pandas.DataFrame` containing proxy data for all sections. ages_df: pandas.DataFrame :class:`pandas.DataFrame` containing age constraints for all sections. lower_age: float Lower bound (youngest age) of the target age interval. upper_age: float Upper bound (oldest age) of the target age interval. sections: list(str) or numpy.array(str), optional List of sections included in the original inference; only required if not all sections in ``sample_df`` were included. Returns ------- height_range_df: pandas.DataFrame Summary statistics for the base and top height of the target age interval (maximum likelihood estimate, median, and 68% and 95% confidence intervals) for each section. """ # get list of proxies included in model from full_trace variables = [ l for l in list(full_trace["posterior"].data_vars.keys()) if (f"{'gp_ls_'}" in l) and (f"{'unshifted'}" not in l) ] proxies = [] for var in variables: proxies.append(var[6:]) if 'sections' in kwargs: sections = list(kwargs['sections']) else: sections = np.unique(sample_df.dropna(subset = proxies, how = 'all')['section']) sample_df, ages_df = clean_data(sample_df, ages_df, proxies, sections) ages_new = full_trace.X_new.X_new.values above = ages_new >= lower_age below = ages_new <= upper_age target_ind = np.where(above & below)[0] age_target = ages_new[target_ind] height_paths = {} keep_sections = [] print('Mapping age models to sections') for section in tqdm(sections): # all sample + age constraint heights section_df = sample_df[sample_df['section']==section] sample_heights = np.concatenate([section_df[~np.isnan(section_df[proxy])]['height'].values for proxy in proxies]) sample_heights = np.unique(sample_heights) sample_heights = np.sort(sample_heights) # heights of radiometric age constraints section_ages_df = ages_df[(ages_df['section']==section) & (~ages_df['Exclude?']) & (~ages_df['intermediate detrital?']) & (~ages_df['intermediate intrusive?'])] age_heights = section_ages_df['height'].values # sample age posterior - shape (samples x draws) sample_age_post = az.extract(full_trace.posterior)[str(section) + '_ages'].values age_constraint_post = az.extract(full_trace.posterior)[str(section) + '_radiometric_age'].values # combine position data for all samples + age constraints included in the inference all_heights = np.concatenate([sample_heights, age_heights]) sorted_idx = np.argsort(all_heights) all_heights_sort = all_heights[sorted_idx] lowest_age = np.min(sample_age_post.ravel()) highest_age = np.max(sample_age_post.ravel()) # only include section if it overlaps the target age range target_range = pd.Interval(np.min(age_target), np.max(age_target), closed = 'both') section_range = pd.Interval(lowest_age, highest_age, closed = 'both') result = target_range.overlaps(section_range) if result: keep_sections.append(section) # construct age and height vectors for the current draw using posteriors for 1) samples in section, and 2) age constraints for i in np.arange(sample_age_post.shape[1]): sample_age_vec = sample_age_post[:, i] constraint_age_vec = age_constraint_post[:, i] age_vec = np.concatenate([sample_age_vec, constraint_age_vec]) age_vec_sort = np.flip(age_vec[sorted_idx]) # young to old (low to high) # interpolate - x is age (must be strictly increasing), y is age interp_heights = np.interp(age_target, age_vec_sort, np.flip(all_heights_sort.astype(float))) # young/high to old/low interp_heights = np.asarray(interp_heights).reshape(len(age_target), 1) if i == 0: height_paths[section] = interp_heights else: height_paths[section] = np.hstack((height_paths[section], interp_heights)) # dataframe with: max likelihood, 2.5, 16, 50, 84, 97.5 percentiles for the top and bottom of the target interval for each section height_range_df = pd.DataFrame(columns = ['section', 'base_mle', 'base_2.5', 'base_16', 'base_50', 'base_84', 'base_97.5', 'top_mle', 'top_2.5', 'top_16', 'top_50', 'top_84', 'top_97.5']) for section in keep_sections: section_stats = {} section_stats['section'] = section section_df = sample_df[sample_df['section']==section] # 0 is the top height, -1 is the bottom height section_stats['top_2.5'] = np.percentile(height_paths[section][0], 2.5) section_stats['base_2.5'] = np.percentile(height_paths[section][-1], 2.5) section_stats['top_97.5'] = np.percentile(height_paths[section][0], 97.5) section_stats['base_97.5'] = np.percentile(height_paths[section][-1], 97.5) section_stats['top_16'] = np.percentile(height_paths[section][0], 16) section_stats['base_16'] = np.percentile(height_paths[section][-1], 16) section_stats['top_84'] = np.percentile(height_paths[section][0], 100-16) section_stats['base_84'] = np.percentile(height_paths[section][-1], 100-16) section_stats['top_50'] = np.percentile(height_paths[section][0], 50) section_stats['base_50'] = np.percentile(height_paths[section][-1], 50) # only calculate MLE if there's more than 1 unique value (edge case where interpolations all hit top/bottom of section) if len(np.unique(height_paths[section][-1])) > 1: dy = np.linspace(np.min(height_paths[section][-1]), np.max(height_paths[section][-1]), 200) section_stats['base_mle'] = dy[np.argmax(gaussian_kde(height_paths[section][-1], bw_method = 1)(dy))] else: section_stats['base_mle'] = np.unique(height_paths[section][-1]) if len(np.unique(height_paths[section][0])) > 1: dy = np.linspace(np.min(height_paths[section][0]), np.max(height_paths[section][0]), 200) section_stats['top_mle'] = dy[np.argmax(gaussian_kde(height_paths[section][0], bw_method = 1)(dy))] else: section_stats['top_mle'] = np.unique(height_paths[section][0]) section_stats_df = pd.DataFrame(section_stats, index = [0]) height_range_df = pd.concat([height_range_df.astype(section_stats_df.dtypes), section_stats_df], ignore_index = True) height_range_df.set_index('section', inplace = True) return height_range_df
[docs] def map_ages_to_section(full_trace, sample_df, ages_df, include_radiometric_ages = False, **kwargs): """ Helper function for :py:meth:`section_proxy_signal() <stratmc.plotting.section_proxy_signal>` in :py:mod:`stratmc.plotting`. Maps the ``ages`` array passed to :py:meth:`get_trace() <stratmc.inference.get_trace>` to height in each section using the most likely posterior age models. Parameters ---------- full_trace: arviz.InferenceData An :class:`arviz.InferenceData` object containing the full set of prior and posterior samples from :py:meth:`get_trace() <stratmc.inference.get_trace>` in :py:mod:`stratmc.inference`. sample_df: pandas.DataFrame :class:`pandas.DataFrame` containing proxy data for all sections. ages_df: pandas.DataFrame :class:`pandas.DataFrame` containing age constraints for all sections. include_radiometric_ages: bool, optional Whether to consider radiometric ages when calculating the most likely posterior age model for each section; defaults to ``False``. sections: list(str) or numpy.array(str), optional List of sections included in the inference; only required if not all sections in ``sample_df`` were included. Returns ------- age_model_df: pandas.DataFrame :class:`pandas.DataFrame` with interpolated heights at each age in the ``ages`` vector that was passed to :py:meth:`get_trace() <stratmc.inference.get_trace>`. """ # get list of proxies included in model from full_trace variables = [ l for l in list(full_trace["posterior"].data_vars.keys()) if (f"{'gp_ls_'}" in l) and (f"{'unshifted'}" not in l) ] proxies = [] for var in variables: proxies.append(var[6:]) if 'sections' in kwargs: sections = list(kwargs['sections']) else: sections = np.unique(sample_df.dropna(subset = proxies, how = 'all')['section']) # only keep samples that were included in the inference sample_df, ages_df = clean_data(sample_df, ages_df, proxies, sections) ages_new = full_trace.X_new.X_new.values age_model_df = pd.DataFrame(columns = ['section', 'age', 'interpolated height']) # get posterior age model for each section age_paths = {} for section in sections: # all sample + age constraint heights section_df = sample_df[sample_df['section']==section] sample_heights = section_df['height'].values sample_heights = np.sort(np.unique(sample_heights)) # heights of radiometric age constraints section_ages_df = ages_df[(ages_df['section']==section) & (~ages_df['Exclude?']) & (~ages_df['intermediate detrital?']) & (~ages_df['intermediate intrusive?'])] age_heights = section_ages_df['height'].values # sample age posterior - shape = (samples x draws) sample_age_post = az.extract(full_trace.posterior)[str(section) + '_ages'].values # get posterior age draws for section if include_radiometric_ages: age_constraint_post = az.extract(full_trace.posterior)[str(section) + '_radiometric_age'].values # combine position data for all samples + age constraints included in the inference all_heights = np.concatenate([sample_heights, age_heights]) sorted_idx = np.argsort(all_heights) all_heights_sort = all_heights[sorted_idx] ages_comb = np.vstack([sample_age_post, age_constraint_post]) age_paths[section] = ages_comb[sorted_idx, :] else: sorted_idx = np.argsort(sample_heights) all_heights_sort = sample_heights[sorted_idx] age_paths[section] = sample_age_post[sorted_idx, :] # calculate MLE age model # shape of age_paths: (observations, draws) max_like = np.zeros(age_paths[section].shape[0]) for i in np.arange(age_paths[section].shape[0]): sample_ages = age_paths[section][i, :] dx = np.linspace(np.min(sample_ages), np.max(sample_ages), 1000) max_like[i] = dx[np.argmax(gaussian_kde(sample_ages, bw_method = 1)(dx))] lowest_age = np.min(max_like) highest_age = np.max(max_like) below = ages_new >= lowest_age above = ages_new <= highest_age target_age_ind = np.where(below & above)[0] target_age_vec = ages_new[target_age_ind] # interpolate the MLE age model interp_height = np.interp(target_age_vec, np.flip(max_like), np.flip(all_heights_sort).astype(np.float64)) section_df_temp = pd.DataFrame({'section': [section] * interp_height.shape[0], 'age': target_age_vec, 'interpolated height': interp_height, }) age_model_df = pd.concat([age_model_df, section_df_temp], ignore_index = True) return age_model_df
[docs] def count_samples(full_trace, time_grid = None): """ Helper function for :py:meth:`proxy_data_density() <stratmc.plotting.proxy_data_density>` in :py:mod:`stratmc.plotting`. Counts the number of observations within discrete time bins (based on the posterior sample ages). Parameters ---------- full_trace: arviz.InferenceData An :class:`arviz.InferenceData` object containing the full set of prior and posterior samples from :py:meth:`get_trace() <stratmc.inference.get_trace>` in :py:mod:`stratmc.inference`. time_grid: np.array, optional Time bin edges; if not provided, defaults to the ``ages`` array passed to :py:meth:`get_trace() <stratmc.inference.get_trace>`. Returns ------- sample_counts: np.array Number of observations in each time bin, summed over all posterior draws such that the average number of observations is ``sample_counts/n``. time_grid: np.array Time bin edges. n: int Number of posterior draws in ``full_trace``. """ sample_ages_post = az.extract(full_trace.posterior)['ages'].values if time_grid is None: time_grid = full_trace.X_new.X_new.values sample_counts, time_grid = np.histogram(sample_ages_post, bins = time_grid) n = sample_ages_post.shape[1] return sample_counts, time_grid, n
[docs] def find_gaps(full_trace, time_grid = None): """ Helper function for :py:meth:`proxy_data_gaps() <stratmc.plotting.proxy_data_gaps>` in :py:mod:`stratmc.plotting`. Counts the number of draws from the posterior where there are no observations within discrete time bins (based on the posterior sample ages). Parameters ---------- full_trace: arviz.InferenceData An :class:`arviz.InferenceData` object containing the full set of prior and posterior samples from :py:meth:`get_trace() <stratmc.inference.get_trace>` in :py:mod:`stratmc.inference`. time_grid: np.array, optional Time bin edges; if not provided, defaults to the ``ages`` array passed to :py:meth:`get_trace() <stratmc.inference.get_trace>`. Returns ------- age_gaps: np.array of int Number of posterior draws where there are no observations; each entry corresponds to an age bin (corresponding to ``grid_centers`` and ``grid_widths``). grid_centers: np.array Time bin centers. grid_widths: np.array Time bin widths. n: int Number of posterior draws in ``full_trace``. """ sample_ages_post = az.extract(full_trace.posterior)['ages'].values if time_grid is None: time_grid = full_trace.X_new.X_new.values # for each grid point age_gaps = np.zeros(len(time_grid) - 1) grid_centers = np.diff(time_grid)/2 + time_grid[:-1] grid_widths = np.diff(time_grid) print('Calculating gaps in the data') for i in tqdm(np.arange(len(time_grid) - 1)): for j in np.arange(sample_ages_post.shape[1]): if all((sample_ages_post[:, j] <= time_grid[i]) | (sample_ages_post[:, j] > time_grid[i + 1])): age_gaps[i] += 1 n = sample_ages_post.shape[1] return age_gaps, grid_centers, grid_widths, n
[docs] def calculate_lengthscale_stability(full_trace, **kwargs): """ Compute the posterior standard deviation of the :class:`pymc.gp.cov.ExpQuad <pymc.gp.cov.ExpQuad>` covariance kernel lengthscale as additional chains are considered (i.e., for 1 to `N` chains). When the posterior has been sufficiently explored, the standard deviation will stabilize; if it has not stabilized, then additional chains should be run. Helper function for :py:meth:`lengthscale_stability() <stratmc.plotting.lengthscale_stability>` in :py:mod:`stratmc.plotting`. To consider chains from multiple traces associated with the same inference model, first combine the traces (saved as NetCDF files) using :py:meth:`combine_traces() <stratmc.data.combine_traces>` in :py:mod:`stratmc.data`. Parameters ---------- full_trace: arviz.InferenceData An :class:`arviz.InferenceData` object containing the full set of prior and posterior samples from :py:meth:`get_trace() <stratmc.inference.get_trace>` in :py:mod:`stratmc.inference`. proxy: str, optional Name of the proxy; only required if multiple proxies were included in the inference model. Returns ------- gp_ls_std: np.array of float Posterior standard deviation of the covariance kernel lengthscale posterior; entries correspond to number of chains considered (first entry is 1 chain, last entry is all `N` chains). """ # get list of proxies included in model from full_trace variables = [ l for l in list(full_trace["posterior"].data_vars.keys()) if (f"{'gp_ls_'}" in l) and (f"{'unshifted'}" not in l) ] proxies = [] for var in variables: proxies.append(var[6:]) if 'proxy' in kwargs: proxy = kwargs['proxy'] else: proxy = proxies[0] gp_ls_post = full_trace.posterior['gp_ls_' + str(proxy)].values gp_ls_std = [] # calculate variance when considering 1 to N chains for i in np.arange(gp_ls_post.shape[0]): gp_ls_std.append((np.std(gp_ls_post[0:i + 1, :, :]))) return gp_ls_std
[docs] def calculate_proxy_signal_stability(full_trace, **kwargs): """ Compute the residuals between the median inferred proxy signal when all *N* chains are considered compared to when 1 to *N-1* chains are considered. When the posterior has been sufficiently explored, the residuals will stabilize and approach zero; if they have not stabilized, then additional chains should be run. Helper function for :py:meth:`proxy_signal_stability() <stratmc.plotting>` in :py:mod:`stratmc.plotting`. To consider chains from multiple traces associated with the same inference model, first combine the traces (saved as NetCDF files) using :py:meth:`combine_traces() <stratmc.data.combine_traces>` in :py:mod:`stratmc.data`. Parameters ---------- full_trace: arviz.InferenceData An :class:`arviz.InferenceData` object containing the full set of prior and posterior samples from :py:meth:`get_trace() <stratmc.inference.get_trace>` in :py:mod:`stratmc.inference`. proxy: str, optional Name of the proxy; only required if multiple proxies were included in the inference model. Returns ------- median_residuals: np.array of float Residuals between the median inferred proxy value (at each time in ``ages`` passed to :py:meth:`get_trace() <stratmc.inference.get_trace>`) calculated using 1 to `N-1` chains versus all `N` chains. Shape is (chains x ages). """ # get list of proxies included in model from full_trace variables = [ l for l in list(full_trace["posterior"].data_vars.keys()) if (f"{'gp_ls_'}" in l) and (f"{'unshifted'}" not in l) ] proxies = [] for var in variables: proxies.append(var[6:]) if 'proxy' in kwargs: proxy = kwargs['proxy'] else: proxy = proxies[0] post_proxy = full_trace.posterior_predictive['f_pred_' + proxy].values median_proxy = [] for i in np.arange(post_proxy.shape[0]): selected_chains = np.arange(0, i + 1) current_data = post_proxy[selected_chains, :, :].reshape(len(selected_chains) * post_proxy.shape[1], post_proxy.shape[2]) median_proxy.append(np.median(current_data, axis = 0)) median_residuals = [] for i in np.arange(post_proxy.shape[0]): median_residuals.append(np.abs(median_proxy[-1] - median_proxy[i])) median_residuals = np.array(median_residuals) return median_residuals