Unexpected Behavior with Additional Keys in Prior Dictionary

[arrjon]

If the prior generates a dict with keys, which are not used neither as inference variables nor as conditions, e.g. additional_key, the user has to specify that these keys should be dropped in the adapter; otherwise, the fitting will fail with the following error:

TypeError: ContinuousApproximator.compute_metrics() got an unexpected keyword argument 'additional_key'

From a user’s point of view, this is unexpected. A more user-friendly approach would be to handle this situation smoothly by issuing a warning about unused keys instead of causing an error.

[LarsKue]

@arrjon Can you provide a minimal reproducible example?

[arrjon]

here a minimal example

import os
os.environ['KERAS_BACKEND'] = 'jax'

import numpy as np
import bayesflow as bf

def prior_bf():
    a = np.random.normal(0, 1)
    b = a * 100
    return dict(a=a, b=b)

def model_bf(b):
    sim = np.random.normal(b, 1)
    return dict(sim=sim)

simulator = bf.simulators.make_simulator([prior_bf, model_bf])

adapter = (
    bf.adapters.Adapter()
    .to_array()
    .convert_dtype(from_dtype="float64", to_dtype="float32")
    .rename("a", "inference_variables")
    .rename("sim", "inference_conditions")
    #.drop('b')
)

workflow = bf.BasicWorkflow(
    adapter=adapter,
    inference_network=bf.networks.CouplingFlow(),
    simulator=simulator
)

history = workflow.fit_online(epochs=1, batch_size=32, num_batches_per_epoch=100)

TypeError: ContinuousApproximator.compute_metrics() got an unexpected keyword argument 'b'

[LarsKue]

@stefanradev93 Is this already sufficiently addressed?

[stefanradev93]

Address in 822ad89