Move FreeFormFlows to experimental module

In initial test, they do not seem to outperform other models in out
standard setting. Until we find a good example, or a configuration where
we see benefits, we move them to the experimental module to indicate
they are not ready for prime time in BayesFlow yet.

Author: vpratz

bayesflow/experimental/__init__.py

@@ -1,5 +1,6 @@
 from .cif import CIF
 from .continuous_time_consistency_model import ContinuousTimeConsistencyModel
+from .free_form_flow import FreeFormFlow
 
 from ..utils._docs import _add_imports_to_all
 

bayesflow/networks/free_form_flow/__init__.py renamed to bayesflow/experimental/free_form_flow/__init__.py

@@ -14,7 +14,7 @@
     deserialize_value_or_type,
 )
 
-from ..inference_network import InferenceNetwork
+from bayesflow.networks import InferenceNetwork
 
 
 @serializable(package="networks.free_form_flow")

bayesflow/networks/free_form_flow/free_form_flow.py renamed to bayesflow/experimental/free_form_flow/free_form_flow.py

@@ -2,7 +2,6 @@
 from .coupling_flow import CouplingFlow
 from .deep_set import DeepSet
 from .flow_matching import FlowMatching
-from .free_form_flow import FreeFormFlow
 from .inference_network import InferenceNetwork
 from .mlp import MLP
 from .lstnet import LSTNet

bayesflow/networks/__init__.py

@@ -50,14 +50,14 @@ def coupling_flow_subnet(subnet):
 
 @pytest.fixture()
 def free_form_flow():
-    from bayesflow.networks import FreeFormFlow
+    from bayesflow.experimental import FreeFormFlow
 
     return FreeFormFlow()
 
 
 @pytest.fixture()
 def free_form_flow_subnet(subnet):
-    from bayesflow.networks import FreeFormFlow
+    from bayesflow.experimental import FreeFormFlow
 
     return FreeFormFlow(encoder_subnet=subnet, decoder_subnet=subnet)