Add projectors to DeepSet (#453)

* v2.0.2 (#447)

* [no ci] notebook tests: increase timeout, fix platform/backend dependent code

Torch is very slow, so I had to increase the timeout accordingly.

* Enable use of summary networks with functional API again (#434)

* summary networks: add tests for using functional API

* fix build functions for use with functional API

* [no ci] docs: add GitHub and Discourse links, reorder navbar

* [no ci] docs: acknowledge scikit-learn website

* [no ci] docs: capitalize navigation headings

* More tests (#437)

* fix docs of coupling flow

* add additional tests

* Automatically run slow tests when main is involved. (#438)

In addition, this PR limits the slow test to Windows and Python 3.10.
The choices are somewhat arbitrary, my thought was to test the setup not
covered as much through use by the devs.

* Update dispatch

* Update dispatching distributions

* Improve workflow tests with multiple summary nets / approximators

* Fix zombie find_distribution import

* Add readme entry [no ci]

* Update README: NumFOCUS affiliation, awesome-abi list (#445)

* fix is_symbolic_tensor

* remove multiple batch sizes, remove multiple python version tests, remove update-workflows branch from workflow style tests, add __init__ and conftest to test_point_approximators (#443)

* implement compile_from_config and get_compile_config (#442)

* implement compile_from_config and get_compile_config

* add optimizer build to compile_from_config

* Fix Optimal Transport for Compiled Contexts (#446)

* remove the is_symbolic_tensor check because this would otherwise skip the whole function for compiled contexts

* skip pyabc test

* fix sinkhorn and log_sinkhorn message formatting for jax by making the warning message worse

* update dispatch tests for more coverage

* Update issue templates (#448)

* Hotfix Version 2.0.1 (#431)

* fix optimal transport config (#429)

* run linter

* [skip-ci] bump version to 2.0.1

* Update issue templates

* Robustify kwargs passing inference networks, add class variables

* fix convergence method to debug for non-log sinkhorn

* Bump optimal transport default to False

* use logging.info for backend selection instead of logging.debug

* fix model comparison approximator

* improve docs and type hints

* improve One-Sample T-Test Notebook:
- use torch as default backend
- reduce range of N so users of jax won't be stuck with a slow notebook
- use BayesFlow built-in MLP instead of keras.Sequential solution
- general code cleanup

* remove backend print

* [skip ci] turn all single-quoted strings into double-quoted strings

* turn all single-quoted strings into double-quoted strings

amend to trigger workflow

---------

Co-authored-by: Valentin Pratz <git@valentinpratz.de>
Co-authored-by: Valentin Pratz <112951103+vpratz@users.noreply.github.com>
Co-authored-by: stefanradev93 <stefan.radev93@gmail.com>
Co-authored-by: Marvin Schmitt <35921281+marvinschmitt@users.noreply.github.com>

* drafting feature

* Initialize projectors for invariant and equivariant DeepSet layers

* implement requested changes and improve activation

---------

Co-authored-by: Lars <lars@kuehmichel.de>
Co-authored-by: Valentin Pratz <git@valentinpratz.de>
Co-authored-by: Valentin Pratz <112951103+vpratz@users.noreply.github.com>
Co-authored-by: stefanradev93 <stefan.radev93@gmail.com>
Co-authored-by: Marvin Schmitt <35921281+marvinschmitt@users.noreply.github.com>

Author: 5 people

bayesflow/networks/deep_set/deep_set.py

@@ -30,7 +30,7 @@ def __init__(
         mlp_widths_invariant_inner: Sequence[int] = (64, 64),
         mlp_widths_invariant_outer: Sequence[int] = (64, 64),
         mlp_widths_invariant_last: Sequence[int] = (64, 64),
-        activation: str = "gelu",
+        activation: str = "silu",
         kernel_initializer: str = "he_normal",
         dropout: int | float | None = 0.05,
         spectral_normalization: bool = False,
@@ -72,7 +72,7 @@ def __init__(
         mlp_widths_invariant_last : Sequence[int], optional
             Widths of the MLP layers in the final invariant transformation. Default is (64, 64).
         activation : str, optional
-            Activation function used throughout the network, such as "gelu". Default is "gelu".
+            Activation function used throughout the network, such as "gelu". Default is "silu".
         kernel_initializer : str, optional
             Initialization strategy for kernel weights, such as "he_normal". Default is "he_normal".
         dropout : int, float, or None, optional

bayesflow/networks/deep_set/equivariant_layer.py

@@ -94,6 +94,7 @@ def __init__(
             kernel_initializer=kernel_initializer,
             spectral_normalization=spectral_normalization,
         )
+        self.out_fc_projector = keras.layers.Dense(mlp_widths_equivariant[-1], kernel_initializer=kernel_initializer)
 
         self.layer_norm = layers.LayerNormalization() if layer_norm else None
 
@@ -137,7 +138,10 @@ def call(self, input_set: Tensor, training: bool = False, **kwargs) -> Tensor:
         output_set = ops.concatenate([input_set, invariant_summary], axis=-1)
 
         # Pass through final equivariant transform + residual
-        output_set = input_set + self.equivariant_fc(output_set, training=training)
+        out_fc = self.equivariant_fc(output_set, training=training)
+        out_projected = self.out_fc_projector(out_fc)
+        output_set = input_set + out_projected
+
         if self.layer_norm is not None:
             output_set = self.layer_norm(output_set, training=training)
 

bayesflow/networks/deep_set/invariant_layer.py

@@ -74,6 +74,7 @@ def __init__(
             kernel_initializer=kernel_initializer,
             spectral_normalization=spectral_normalization,
         )
+        self.inner_projector = keras.layers.Dense(mlp_widths_inner[-1], kernel_initializer=kernel_initializer)
 
         self.outer_fc = MLP(
             mlp_widths_outer,
@@ -82,6 +83,7 @@ def __init__(
             kernel_initializer=kernel_initializer,
             spectral_normalization=spectral_normalization,
         )
+        self.outer_projector = keras.layers.Dense(mlp_widths_outer[-1], kernel_initializer=kernel_initializer)
 
         # Pooling function as keras layer for sum decomposition: inner( pooling( inner(set) ) )
         if pooling_kwargs is None:
@@ -106,8 +108,10 @@ def call(self, input_set: Tensor, training: bool = False, **kwargs) -> Tensor:
         """
 
         set_summary = self.inner_fc(input_set, training=training)
+        set_summary = self.inner_projector(set_summary)
         set_summary = self.pooling_layer(set_summary, training=training)
         set_summary = self.outer_fc(set_summary, training=training)
+        set_summary = self.outer_projector(set_summary)
         return set_summary
 
     @sanitize_input_shape