[BUG] serialization of the broadcast transformation in the adapter fails

[arrjon]

Describe the bug
The serialization of the broadcast transformation in the adapter does not work. After saving and loading a model, the broadcasting fails. With some debugging, I found that the problem is that self.expand is loaded as a list instead of a tuple. This error might also show up somewhere else, but I am not sure.

To Reproduce
Minimal steps to reproduce the behavior:

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

import numpy as np
import keras
import bayesflow as bf
#%%
def prior():
    return dict(alpha=np.random.uniform(0, 1), beta=np.random.uniform(0, 1))

def simulator(alpha, beta):
    # Simulate some data based on the alpha parameter
    data = np.random.normal(loc=alpha, scale=beta, size=100)
    inference_conditions = np.random.normal(loc=0, scale=1, size=(100, 1, 1))
    return dict(data=data, inference_conditions=inference_conditions)

adapter = (
    bf.adapters.Adapter()
    .to_array()
    .broadcast("data", to="inference_conditions", expand=(2,3))
    .squeeze(keys="inference_conditions", axis=(2,3))
    .concatenate(['alpha', 'beta'], into='inference_variables')
    )

bf_sim = bf.make_simulator([prior, simulator])
#%%
print(adapter.forward(bf_sim.sample(5))['inference_variables'].shape, adapter.forward(bf_sim.sample(5))['inference_conditions'].shape)
#%%
print(adapter.forward(bf_sim.sample(5))['data'].shape)
#%%
workflow = bf.BasicWorkflow(
    simulator=bf_sim,
    adapter=adapter,
    inference_network=bf.networks.CouplingFlow()
)
#%%
history = workflow.fit_online(
    epochs=1,
    batch_size=32
)
workflow.approximator.save(filepath='test.keras')
#%%
workflow.plot_default_diagnostics(test_data=bf_sim.sample(100), num_samples=100)
#%%
workflow.approximator = keras.saving.load_model(filepath='test.keras')
#%%
workflow.plot_default_diagnostics(test_data=bf_sim.sample(100), num_samples=100)

Expected behavior
Saving and loading of the adapter should not be problem. If a transform was working before, it should also work after the loading.

Traceback
ValueError: operands could not be broadcast together with remapped shapes [original->remapped]: (100,100) and requested shape (100,100,1,100)

Environment

• OS: MacOS
• Python Version: 3.11
• Backend: fails for all of them
• BayesFlow Version: 2.0.4 (dev branch from today)

[vpratz]

Fixed by #530, thanks for reporting and fixing @arrjon!