molcraft 0.1.0rc9__py3-none-any.whl

molcraft/layers.py

@@ -0,0 +1,1910 @@
+import warnings
+import keras 
+import tensorflow as tf
+import functools
+import inspect
+from keras.src.models import functional
+
+from molcraft import tensors
+from molcraft import ops 
+
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class GraphLayer(keras.layers.Layer):
+    """Base graph layer.
+
+    Subclasses must implement a forward pass via **propagate(graph)**. 
+    
+    Subclasses may create dense layers and weights in **build(graph_spec)**.
+
+    Note: `GraphLayer` currently only supports `GraphTensor` input.
+
+    The list of arguments below are only relevant if the derived layer 
+    invokes 'get_dense_kwargs`, `get_dense`  or `get_einsum_dense`. 
+
+    Arguments:
+        use_bias (bool):
+            Whether bias should be used in dense layers. Default to `True`.
+        kernel_initializer (keras.initializers.Initializer, str):
+            Initializer for the kernel weight matrix of the dense layers.
+            Default to `glorot_uniform`.
+        bias_initializer (keras.initializers.Initializer, str):
+            Initializer for the bias weight vector of the dense layers.
+            Default to `zeros`.
+        kernel_regularizer (keras.regularizers.Regularizer, None):
+            Regularizer function applied to the kernel weight matrix.
+            Default to `None`.
+        bias_regularizer (keras.regularizers.Regularizer, None):
+            Regularizer function applied to the bias weight vector.
+            Default to `None`.
+        activity_regularizer (keras.regularizers.Regularizer, None):
+            Regularizer function applied to the output of the dense layers.
+            Default to `None`.
+        kernel_constraint (keras.constraints.Constraint, None):
+            Constraint function applied to the kernel weight matrix.
+            Default to `None`.
+        bias_constraint (keras.constraints.Constraint, None):
+            Constraint function applied to the bias weight vector.
+            Default to `None`.
+    """
+
+    def __init__(
+        self,
+        use_bias: bool = True,
+        kernel_initializer: keras.initializers.Initializer | str = "glorot_uniform",
+        bias_initializer: keras.initializers.Initializer | str = "zeros",
+        kernel_regularizer: keras.regularizers.Regularizer | None = None,
+        bias_regularizer: keras.regularizers.Regularizer | None = None,
+        activity_regularizer: keras.regularizers.Regularizer | None = None,
+        kernel_constraint: keras.constraints.Constraint | None = None,
+        bias_constraint: keras.constraints.Constraint | None = None,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        self._use_bias = use_bias
+        self._kernel_initializer = keras.initializers.get(kernel_initializer)
+        self._bias_initializer = keras.initializers.get(bias_initializer)
+        self._kernel_regularizer = keras.regularizers.get(kernel_regularizer)
+        self._bias_regularizer = keras.regularizers.get(bias_regularizer)
+        self._activity_regularizer = keras.regularizers.get(activity_regularizer)
+        self._kernel_constraint = keras.constraints.get(kernel_constraint)
+        self._bias_constraint = keras.constraints.get(bias_constraint)
+        self._custom_build_config = {}
+        self._propagate_kwargs = _propagate_kwargs(self.propagate)
+        self.built = False
+
+    def __init_subclass__(cls, **kwargs):
+        super().__init_subclass__(**kwargs)
+        subclass_build = cls.build
+
+        @functools.wraps(subclass_build)
+        def build_wrapper(self: GraphLayer, spec: tensors.GraphTensor.Spec | None):
+            GraphLayer.build(self, spec)
+            subclass_build(self, spec)
+            if not self.built and isinstance(self, keras.Model):
+                symbolic_inputs = Input(spec)
+                self.built = True
+                self(symbolic_inputs)
+
+        cls.build = build_wrapper
+
+    def propagate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        """Forward pass.
+
+        Must be implemented by subclass.
+
+        Arguments:
+            tensor:
+                A `GraphTensor` instance.
+        """
+        raise NotImplementedError(
+            'The forward pass of the layer is not implemented. '
+            'Please implement `propagate`.'
+        )
+
+    def build(self, tensor_spec: tensors.GraphTensor.Spec) -> None:
+        """Builds the layer.
+
+        May use built-in methods such as `get_weight`, `get_dense` and `get_einsum_dense`.
+
+        Optionally implemented by subclass.
+
+        Arguments:
+            tensor_spec:
+                A `GraphTensor.Spec` instance corresponding to the `GraphTensor` 
+                passed to `propagate`.
+        """ 
+        if isinstance(tensor_spec, tensors.GraphTensor.Spec):
+            self._custom_build_config['spec'] = _serialize_spec(tensor_spec)
+
+    def call(
+        self, 
+        graph: dict[str, dict[str, tf.Tensor]],
+        training: bool | None = None,
+    ) -> dict[str, dict[str, tf.Tensor]]:
+        graph_tensor = tensors.from_dict(graph)
+        if not self._propagate_kwargs:
+            outputs = self.propagate(graph_tensor)
+        else:
+            outputs = self.propagate(graph_tensor, training=training)
+        if isinstance(outputs, tensors.GraphTensor):
+            return tensors.to_dict(outputs)
+        return outputs
+
+    def __call__(
+        self, 
+        graph: dict[str, dict[str, tf.Tensor]] | tensors.GraphTensor, 
+        **kwargs
+    ) -> tf.Tensor | dict[str, dict[str, tf.Tensor]] | tensors.GraphTensor:
+        if not self.built:
+            spec = _spec_from_inputs(graph)
+            self.build(spec)
+
+        is_graph_tensor = isinstance(graph, tensors.GraphTensor)
+        if is_graph_tensor:
+            graph = tensors.to_dict(graph)
+        else:
+            graph = {field: dict(data) for (field, data) in graph.items()}
+
+        if isinstance(self, functional.Functional):
+            # As a functional model is strict for what input can be 
+            # passed to it, we need to temporarily pop some of the 
+            # input and add it back afterwards.
+            symbolic_input = self._symbolic_input
+            excluded = {}
+            for outer_field in ['context', 'node', 'edge']:
+                excluded[outer_field] = {}
+                for inner_field in list(graph[outer_field]):
+                    if inner_field not in symbolic_input[outer_field]:
+                        excluded[outer_field][inner_field] = (
+                            graph[outer_field].pop(inner_field)
+                        )
+            
+            tf.nest.assert_same_structure(self.input, graph)
+
+        outputs = super().__call__(graph, **kwargs)
+
+        if not tensors.is_graph(outputs):
+            return outputs
+        
+        graph = outputs
+        if isinstance(self, functional.Functional):
+            for outer_field in ['context', 'node', 'edge']:
+                for inner_field in list(excluded[outer_field]):
+                    graph[outer_field][inner_field] = (
+                        excluded[outer_field].pop(inner_field)
+                    )
+
+        if is_graph_tensor:
+            return tensors.from_dict(graph)
+
+        return graph
+
+    def get_build_config(self) -> dict:
+        if self._custom_build_config:
+            return self._custom_build_config
+        return super().get_build_config()
+    
+    def build_from_config(self, config: dict) -> None:
+        serialized_spec = config.get('spec')
+        if serialized_spec is not None:
+            spec = _deserialize_spec(serialized_spec)
+            self.build(spec)
+        else:
+            super().build_from_config(config)
+
+    def get_weight(
+        self,
+        shape: tf.TensorShape,
+        **kwargs,
+    ) -> tf.Variable:
+        common_kwargs = self.get_dense_kwargs()
+        weight_kwargs = {
+            'initializer': common_kwargs['kernel_initializer'],
+            'regularizer': common_kwargs['kernel_regularizer'],
+            'constraint': common_kwargs['kernel_constraint']
+        }
+        weight_kwargs.update(kwargs)
+        return self.add_weight(shape=shape, **weight_kwargs)
+    
+    def get_dense(
+        self, 
+        units: int, 
+        **kwargs
+    ) -> keras.layers.Dense:
+        common_kwargs = self.get_dense_kwargs()
+        common_kwargs.update(kwargs)
+        return keras.layers.Dense(units, **common_kwargs)
+    
+    def get_einsum_dense(
+        self, 
+        equation: str, 
+        output_shape: tf.TensorShape, 
+        **kwargs
+    ) -> keras.layers.EinsumDense:
+        common_kwargs = self.get_dense_kwargs()
+        common_kwargs.update(kwargs)
+        use_bias = common_kwargs.pop('use_bias', False)
+        if use_bias and not 'bias_axes' in common_kwargs:
+            common_kwargs['bias_axes'] = equation.split('->')[-1][1:] or None
+        return keras.layers.EinsumDense(equation, output_shape, **common_kwargs)
+    
+    def get_dense_kwargs(self) -> dict:
+        common_kwargs = dict(
+            use_bias=self._use_bias,
+            kernel_regularizer=self._kernel_regularizer,
+            bias_regularizer=self._bias_regularizer,
+            activity_regularizer=self._activity_regularizer,
+            kernel_constraint=self._kernel_constraint,
+            bias_constraint=self._bias_constraint,
+        )
+        kernel_initializer = self._kernel_initializer.__class__.from_config(
+            self._kernel_initializer.get_config()
+        )
+        bias_initializer = self._bias_initializer.__class__.from_config(
+            self._bias_initializer.get_config()
+        )
+        common_kwargs["kernel_initializer"] = kernel_initializer
+        common_kwargs["bias_initializer"] = bias_initializer
+        return common_kwargs
+    
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({
+            "use_bias": self._use_bias,
+            "kernel_initializer": 
+                keras.initializers.serialize(self._kernel_initializer),
+            "bias_initializer": 
+                keras.initializers.serialize(self._bias_initializer),
+            "kernel_regularizer": 
+                keras.regularizers.serialize(self._kernel_regularizer),
+            "bias_regularizer": 
+                keras.regularizers.serialize(self._bias_regularizer),
+            "activity_regularizer":
+                keras.regularizers.serialize(self._activity_regularizer),
+            "kernel_constraint": 
+                keras.constraints.serialize(self._kernel_constraint),
+            "bias_constraint": 
+                keras.constraints.serialize(self._bias_constraint),
+        })
+        return config
+
+    @property
+    def _symbolic_output(self) -> dict[dict[str, keras.KerasTensor]]:
+        output = self.output
+        if tensors.is_graph(output):
+            # GraphModel
+            return output
+        symbolic_input = self._symbolic_input
+        symbolic_output = self.compute_output_spec(symbolic_input)
+        return tf.nest.pack_sequence_as(symbolic_output, output)
+
+    @property
+    def _symbolic_input(self) -> dict[dict[str, keras.KerasTensor]]:
+        input = self.input
+        if tensors.is_graph(input):
+            # GraphModel or initial GraphLayer of GraphModel
+            return input
+        spec = _deserialize_spec(self.get_build_config()['spec'])
+        spec_dict = {k: dict(v) for k, v in spec.__dict__.items()}
+        return tf.nest.pack_sequence_as(spec_dict, input)
+
+    @property
+    def output_spec(self) -> tensors.GraphTensor.Spec | tf.TensorSpec:
+        if not self.built:
+            return None
+        if isinstance(self, functional.Functional):
+            output_spec = self.output
+        else:
+            serialized_spec = self.get_build_config()['spec']
+            deserialized_spec = _deserialize_spec(serialized_spec)
+            input_spec = Input(deserialized_spec)
+            output_spec = self.compute_output_spec(input_spec)
+        if not tensors.is_graph(output_spec):
+            return tf.TensorSpec(output_spec.shape, output_spec.dtype)
+        spec_dict = tf.nest.map_structure(
+            lambda t: tf.TensorSpec(t.shape, t.dtype), output_spec
+        )
+        return tensors.GraphTensor.Spec(**spec_dict)
+
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class GraphConv(GraphLayer):
+
+    """Base graph neural network layer.
+
+    This layer implements the three basic steps of a graph neural network layer, each of which 
+    can (optionally) be overridden by the `GraphConv` subclass:
+
+    1. **message(graph)**, which computes the *messages* to be passed to target nodes;
+    2. **aggregate(graph)**, which *aggregates* messages to target nodes;
+    3. **update(graph)**, which further *updates* (target) nodes.
+
+    Note: for skip connection to work, the `GraphConv` subclass requires final node feature 
+    output dimension to be equal to `units`. 
+
+    Arguments:
+        units (int):
+            Dimensionality of the output space.
+        activation (keras.layers.Activation, str, None):
+            Activation function to be accessed via `self.activation`, and used for the 
+            `message()` and `update()` methods, if not overriden. Default to `relu`.
+        use_bias (bool):
+            Whether bias should be used in the dense layers. Default to `True`.
+        normalize (bool, str):
+            Whether a normalization layer should be obtain by `get_norm()`. Default to `False`.
+        skip_connect (bool):
+            Whether node feature input should be added to the node feature output. Default to `True`.
+        kernel_initializer (keras.initializers.Initializer, str):
+            Initializer for the kernel weight matrix of the dense layers.
+            Default to `glorot_uniform`.
+        bias_initializer (keras.initializers.Initializer, str):
+            Initializer for the bias weight vector of the dense layers.
+            Default to `zeros`.
+        kernel_regularizer (keras.regularizers.Regularizer, None):
+            Regularizer function applied to the kernel weight matrix.
+            Default to `None`.
+        bias_regularizer (keras.regularizers.Regularizer, None):
+            Regularizer function applied to the bias weight vector.
+            Default to `None`.
+        activity_regularizer (keras.regularizers.Regularizer, None):
+            Regularizer function applied to the output of the dense layers.
+            Default to `None`.
+        kernel_constraint (keras.constraints.Constraint, None):
+            Constraint function applied to the kernel weight matrix.
+            Default to `None`.
+        bias_constraint (keras.constraints.Constraint, None):
+            Constraint function applied to the bias weight vector.
+            Default to `None`.
+    """
+        
+    def __init__(
+        self, 
+        units: int = None, 
+        activation: str | keras.layers.Activation | None = 'relu',
+        use_bias: bool = True,
+        normalize: bool = False,
+        skip_connect: bool = True,
+        **kwargs
+    ) -> None:
+        super().__init__(use_bias=use_bias, **kwargs)
+        self._units = units
+        self._normalize = normalize
+        self._skip_connect = skip_connect
+        self._activation = keras.activations.get(activation)
+
+    def __init_subclass__(cls, **kwargs):
+        super().__init_subclass__(**kwargs)
+        subclass_build = cls.build
+
+        @functools.wraps(subclass_build)
+        def build_wrapper(self, spec):
+            GraphConv.build(self, spec)
+            return subclass_build(self, spec)
+
+        cls.build = build_wrapper
+
+    @property 
+    def units(self):
+        return self._units 
+    
+    @property
+    def activation(self):
+        return self._activation
+    
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        if not self.units:
+            raise ValueError(
+                f'`self.units` needs to be a positive integer. Found: {self.units}.'
+            )
+        node_feature_dim = spec.node['feature'].shape[-1]
+        self._project_residual = (
+            self._skip_connect and (node_feature_dim != self.units)
+        )
+        if self._project_residual:
+            warnings.warn(
+                'Found incompatible dim between input and output. Applying '
+                'a projection layer to residual to match input and output dim.',
+            )
+            self._residual_dense = self.get_dense(
+                self.units, name='residual_dense'
+            )
+
+        self.has_edge_feature = 'feature' in spec.edge
+        self.has_node_coordinate = 'coordinate' in spec.node
+
+        has_overridden_message = self.__class__.message != GraphConv.message 
+        if not has_overridden_message:
+            self._message_intermediate_dense = self.get_dense(self.units)
+            self._message_norm = self.get_norm()
+            self._message_intermediate_activation = self.activation
+            self._message_final_dense = self.get_dense(self.units)
+
+        has_overridden_aggregate = self.__class__.message != GraphConv.aggregate
+        if not has_overridden_aggregate:
+            pass
+
+        has_overridden_update = self.__class__.update != GraphConv.update 
+        if not has_overridden_update:
+            self._update_intermediate_dense = self.get_dense(self.units)
+            self._update_norm = self.get_norm()
+            self._update_intermediate_activation = self.activation
+            self._update_final_dense = self.get_dense(self.units)
+
+    def propagate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        """Forward pass.
+
+        Invokes `message(graph)`, `aggregate(graph)` and `update(graph)` in sequence.
+
+        Arguments:
+            tensor:
+                A `GraphTensor` instance.
+        """
+        if self._skip_connect:
+            residual = tensor.node['feature']
+            if self._project_residual:
+                residual = self._residual_dense(residual)
+
+        message = self.message(tensor)
+        add_message = not isinstance(message, tensors.GraphTensor)
+        if add_message:
+            message = tensor.update({'edge': {'message': message}})
+        elif not 'message' in message.edge:
+            raise ValueError('Could not find `message` in `edge` output.')
+        
+        aggregate = self.aggregate(message)
+        add_aggregate = not isinstance(aggregate, tensors.GraphTensor)
+        if add_aggregate:
+            aggregate = tensor.update({'node': {'aggregate': aggregate}})
+        elif not 'aggregate' in aggregate.node:
+            raise ValueError('Could not find `aggregate` in `node` output.')
+            
+        update = self.update(aggregate)
+        if not isinstance(update, tensors.GraphTensor):
+            update = tensor.update({'node': {'feature': update}})
+        elif not 'feature' in update.node:
+            raise ValueError('Could not find `feature` in `node` output.')
+        
+        if update.node['feature'].shape[-1] != self.units:
+            raise ValueError('Updated node `feature` is not equal to `self.units`.')
+
+        if add_message and add_aggregate:
+            update = update.update({'node': {'aggregate': None}, 'edge': {'message': None}})
+        elif add_message:
+            update = update.update({'edge': {'message': None}})
+        elif add_aggregate:
+            update = update.update({'node': {'aggregate': None}})
+
+        if not self._skip_connect:
+            return update
+        
+        feature = update.node['feature']
+
+        if self._skip_connect:
+            feature += residual 
+
+        return update.update({'node': {'feature': feature}})
+
+    def message(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        """Compute messages.
+
+        This method may be overridden by subclass.
+
+        Arguments:
+            tensor:
+                The inputted `GraphTensor` instance.
+        """
+        message = keras.ops.concatenate(
+            [
+                tensor.gather('feature', 'source'),
+                tensor.gather('feature', 'target'),
+            ],
+            axis=-1
+        )
+        if self.has_edge_feature:
+            message = keras.ops.concatenate(
+                [
+                    message, 
+                    tensor.edge['feature']
+                ], 
+                axis=-1
+            )
+        if self.has_node_coordinate:
+            euclidean_distance = ops.euclidean_distance(
+                tensor.gather('coordinate', 'target'),
+                tensor.gather('coordinate', 'source'),
+                axis=-1,
+                keepdims=True
+            )
+            message = keras.ops.concatenate(
+                [
+                    message,
+                    euclidean_distance
+                ],
+                axis=-1
+            )
+        message = self._message_intermediate_dense(message)
+        message = self._message_norm(message)
+        message = self._message_intermediate_activation(message)
+        message = self._message_final_dense(message)
+        return tensor.update({'edge': {'message': message}})
+
+    def aggregate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        """Aggregates messages.
+
+        This method may be overridden by subclass.
+
+        Arguments:
+            tensor:
+                A `GraphTensor` instance containing a message.
+        """
+        previous = tensor.node['feature']
+        aggregate = tensor.aggregate('message', mode='mean')
+        aggregate = keras.ops.concatenate([aggregate, previous], axis=-1)
+        return tensor.update(
+            {
+                'node': {
+                    'aggregate': aggregate, 
+                },
+                'edge': {
+                    'message': None,
+                }
+            }
+        )
+
+    def update(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        """Updates nodes. 
+
+        This method may be overridden by subclass.
+
+        Arguments:
+            tensor:
+                A `GraphTensor` instance containing aggregated messages 
+                (updated node features).
+        """
+        aggregate = tensor.node['aggregate']
+        node_feature = self._update_intermediate_dense(aggregate)
+        node_feature = self._update_norm(node_feature)
+        node_feature = self._update_intermediate_activation(node_feature)
+        node_feature = self._update_final_dense(node_feature)
+        return tensor.update(
+            {
+                'node': {
+                    'feature': node_feature,
+                    'aggregate': None,
+                },
+            }
+        )
+
+    def get_norm(self, **kwargs):
+        if not self._normalize:
+            return keras.layers.Identity()
+        elif str(self._normalize).lower().startswith('layer'):
+            return keras.layers.LayerNormalization(**kwargs)
+        else:
+            return keras.layers.BatchNormalization(**kwargs)
+        
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({
+            'units': self.units,
+            'activation': keras.activations.serialize(self._activation),
+            'normalize': self._normalize,
+            'skip_connect': self._skip_connect,
+        })
+        return config
+    
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class GIConv(GraphConv):
+
+    """Graph isomorphism network layer.
+
+    >>> graph = molcraft.tensors.GraphTensor(
+    ...     context={
+    ...         'size': [2]
+    ...     },
+    ...     node={
+    ...         'feature': [[1.], [2.]]
+    ...     },
+    ...     edge={
+    ...         'source': [0, 1],
+    ...         'target': [1, 0],
+    ...     }
+    ... )
+    >>> conv = molcraft.layers.GIConv(units=4)
+    >>> conv(graph)
+        GraphTensor(
+            context={
+                'size': <tf.Tensor: shape=[1], dtype=int32>
+            },
+            node={
+                'feature': <tf.Tensor: shape=[2, 4], dtype=float32>
+            },
+            edge={
+                'source': <tf.Tensor: shape=[2], dtype=int32>,
+                'target': <tf.Tensor: shape=[2], dtype=int32>
+            }
+        )
+    """
+
+    def __init__(
+        self,
+        units: int,
+        activation: keras.layers.Activation | str | None = 'relu',
+        use_bias: bool = True,
+        normalize: bool = False,
+        skip_connect: bool = True,
+        update_edge_feature: bool = True,
+        **kwargs,
+    ):
+        super().__init__(
+            units=units, 
+            activation=activation,
+            use_bias=use_bias, 
+            normalize=normalize, 
+            skip_connect=skip_connect,
+            **kwargs
+        )
+        self._update_edge_feature = update_edge_feature
+
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        node_feature_dim = spec.node['feature'].shape[-1]
+
+        self.epsilon = self.add_weight(
+            name='epsilon', 
+            shape=(), 
+            initializer='zeros',
+            trainable=True,
+        )
+
+        if self.has_edge_feature:
+            edge_feature_dim = spec.edge['feature'].shape[-1]
+
+            if not self._update_edge_feature:
+                if (edge_feature_dim != node_feature_dim):
+                    warnings.warn(
+                        'Found edge and node feature dim to be incompatible. Applying a '
+                        'projection layer to edge features to match the dim of the node features.',
+                    )
+                    self._update_edge_feature = True 
+
+            if self._update_edge_feature:
+                self._edge_dense = self.get_dense(node_feature_dim)
+        else:
+            self._update_edge_feature = False
+
+    def message(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        message = tensor.gather('feature', 'source')
+        edge_feature = tensor.edge.get('feature')
+        if self._update_edge_feature:
+            edge_feature = self._edge_dense(edge_feature)
+        if self.has_edge_feature:
+            message += edge_feature
+            message = keras.ops.relu(message)
+        return tensor.update(
+            {
+                'edge': {
+                    'message': message,
+                    'feature': edge_feature
+                }
+            }
+        )
+
+    def aggregate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        node_feature = tensor.aggregate('message', mode='mean')
+        node_feature += (1 + self.epsilon) * tensor.node['feature']
+        return tensor.update(
+            {
+                'node': {
+                    'aggregate': node_feature,
+                },
+                'edge': {
+                    'message': None,
+                }
+            }
+        )
+    
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({
+            'update_edge_feature': self._update_edge_feature
+        })
+        return config
+
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class GAConv(GraphConv):
+
+    """Graph attention network layer.
+
+    >>> graph = molcraft.tensors.GraphTensor(
+    ...     context={
+    ...         'size': [2]
+    ...     },
+    ...     node={
+    ...         'feature': [[1.], [2.]]
+    ...     },
+    ...     edge={
+    ...         'source': [0, 1],
+    ...         'target': [1, 0],
+    ...     }
+    ... )
+    >>> conv = molcraft.layers.GAConv(units=4, heads=2)
+    >>> conv(graph)
+        GraphTensor(
+            context={
+                'size': <tf.Tensor: shape=[1], dtype=int32>
+            },
+            node={
+                'feature': <tf.Tensor: shape=[2, 4], dtype=float32>
+            },
+            edge={
+                'source': <tf.Tensor: shape=[2], dtype=int32>,
+                'target': <tf.Tensor: shape=[2], dtype=int32>
+            }
+        )
+    """
+
+    def __init__(
+        self,
+        units: int,
+        heads: int = 8,
+        activation: keras.layers.Activation | str | None = "relu",
+        use_bias: bool = True,
+        normalize: bool = False,
+        skip_connect: bool = True,
+        update_edge_feature: bool = True,
+        **kwargs,
+    ) -> None:
+        super().__init__(
+            units=units, 
+            activation=activation,
+            normalize=normalize, 
+            use_bias=use_bias,
+            skip_connect=skip_connect,
+            **kwargs
+        )
+        self._heads = heads
+        if self.units % self.heads != 0:
+            raise ValueError(f"units need to be divisible by heads.")
+        self._head_units = self.units // self.heads 
+        self._update_edge_feature = update_edge_feature
+
+    @property 
+    def heads(self):
+        return self._heads 
+    
+    @property 
+    def head_units(self):
+        return self._head_units 
+
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        self._update_edge_feature = self.has_edge_feature and self._update_edge_feature
+        if self._update_edge_feature:
+            self._edge_dense = self.get_einsum_dense(
+                'ijh,jkh->ikh', (self.head_units, self.heads)
+            )
+        self._node_dense = self.get_einsum_dense(
+            'ij,jkh->ikh', (self.head_units, self.heads)
+        )
+        self._feature_dense = self.get_einsum_dense(
+            'ij,jkh->ikh', (self.head_units, self.heads)
+        )
+        self._attention_dense = self.get_einsum_dense(
+            'ijh,jkh->ikh', (1, self.heads)
+        )
+
+    def message(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        attention_feature = keras.ops.concatenate(
+            [
+                tensor.gather('feature', 'source'),
+                tensor.gather('feature', 'target')
+            ], 
+            axis=-1
+        )
+        if self.has_edge_feature:
+            attention_feature = keras.ops.concatenate(
+                [
+                    attention_feature, 
+                    tensor.edge['feature']
+                ], 
+                axis=-1
+            )
+
+        attention_feature = self._feature_dense(attention_feature)
+
+        edge_feature = tensor.edge.get('feature')
+
+        if self._update_edge_feature:
+            edge_feature = self._edge_dense(attention_feature)
+            edge_feature = keras.ops.reshape(edge_feature, (-1, self.units))
+        
+        attention_feature = keras.ops.leaky_relu(attention_feature)
+        attention_score = self._attention_dense(attention_feature)
+        attention_score = ops.edge_softmax(
+            score=attention_score, edge_target=tensor.edge['target']
+        )
+        node_feature = self._node_dense(tensor.node['feature'])
+        message = ops.gather(node_feature, tensor.edge['source'])
+        message = ops.edge_weight(message, attention_score)
+        return tensor.update(
+            {
+                'edge': {
+                    'message': message,
+                    'feature': edge_feature,
+                }
+            }
+        )
+    
+    def aggregate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        node_feature = tensor.aggregate('message', mode='sum')
+        node_feature = keras.ops.reshape(node_feature, (-1, self.units))
+        return tensor.update(
+            {
+                'node': {
+                    'aggregate': node_feature
+                },
+                'edge': {
+                    'message': None,
+                }
+            }
+        )
+    
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({
+            "heads": self._heads,
+            'update_edge_feature': self._update_edge_feature,
+        })
+        return config
+    
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class MPConv(GraphConv):
+
+    """Message passing neural network layer.
+
+    Also supports 3D molecular graphs.
+
+    >>> graph = molcraft.tensors.GraphTensor(
+    ...     context={
+    ...         'size': [2]
+    ...     },
+    ...     node={
+    ...         'feature': [[1.], [2.]]
+    ...     },
+    ...     edge={
+    ...         'source': [0, 1],
+    ...         'target': [1, 0],
+    ...     }
+    ... )
+    >>> conv = molcraft.layers.MPConv(units=4)
+    >>> conv(graph)
+        GraphTensor(
+            context={
+                'size': <tf.Tensor: shape=[1], dtype=int32>
+            },
+            node={
+                'feature': <tf.Tensor: shape=[2, 4], dtype=float32>
+            },
+            edge={
+                'source': <tf.Tensor: shape=[2], dtype=int32>,
+                'target': <tf.Tensor: shape=[2], dtype=int32>
+            }
+        )
+    """
+
+    def __init__(
+        self, 
+        units: int = 128, 
+        activation: keras.layers.Activation | str | None = 'relu', 
+        use_bias: bool = True,
+        normalize: bool = False,
+        skip_connect: bool = True,
+        **kwargs
+    ) -> None:
+        super().__init__(
+            units=units, 
+            activation=activation,
+            use_bias=use_bias,
+            normalize=normalize, 
+            skip_connect=skip_connect,
+            **kwargs
+        )
+
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        node_feature_dim = spec.node['feature'].shape[-1]
+        self.update_fn = keras.layers.GRUCell(self.units)
+        self._project_previous_node_feature = node_feature_dim != self.units
+        if self._project_previous_node_feature:
+            warnings.warn(
+                'Inputted node feature dim does not match updated node feature dim, '
+                'which is required for the GRU update. Applying a projection layer to '
+                'the inputted node features prior to the GRU update, to match dim '
+                'of the updated node feature dim.'
+            )
+            self._previous_node_dense = self.get_dense(self.units)
+
+    def aggregate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        """Aggregates messages.
+
+        This method may be overridden by subclass.
+
+        Arguments:
+            tensor:
+                A `GraphTensor` instance containing a message.
+        """
+        aggregate = tensor.aggregate('message', mode='mean')
+        return tensor.update(
+            {
+                'node': {
+                    'aggregate': aggregate, 
+                },
+                'edge': {
+                    'message': None,
+                }
+            }
+        )
+    
+    def update(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        previous = tensor.node['feature']
+        aggregate = tensor.node['aggregate']
+        if self._project_previous_node_feature:
+            previous = self._previous_node_dense(previous)
+        updated_node_feature, _ = self.update_fn(
+            inputs=aggregate, states=previous
+        )
+        return tensor.update(
+            {
+                'node': {
+                    'feature': updated_node_feature,
+                    'aggregate': None,
+                }
+            }
+        )
+
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({})
+        return config 
+    
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class GTConv(GraphConv):
+
+    """Graph transformer layer.
+
+    Also supports 3D molecular graphs.
+
+    >>> graph = molcraft.tensors.GraphTensor(
+    ...     context={
+    ...         'size': [2]
+    ...     },
+    ...     node={
+    ...         'feature': [[1.], [2.]]
+    ...     },
+    ...     edge={
+    ...         'source': [0, 1],
+    ...         'target': [1, 0],
+    ...     }
+    ... )
+    >>> conv = molcraft.layers.GTConv(units=4, heads=2)
+    >>> conv(graph)
+        GraphTensor(
+            context={
+                'size': <tf.Tensor: shape=[1], dtype=int32>
+            },
+            node={
+                'feature': <tf.Tensor: shape=[2, 4], dtype=float32>
+            },
+            edge={
+                'source': <tf.Tensor: shape=[2], dtype=int32>,
+                'target': <tf.Tensor: shape=[2], dtype=int32>,
+            }
+        )
+    """
+
+    def __init__(
+        self,
+        units: int,
+        heads: int = 8,
+        activation: keras.layers.Activation | str | None = "relu",
+        use_bias: bool = True,
+        normalize: bool = False,
+        skip_connect: bool = True,
+        attention_dropout: float = 0.0,
+        **kwargs,
+    ) -> None:
+        super().__init__(
+            units=units, 
+            activation=activation,
+            normalize=normalize, 
+            use_bias=use_bias,
+            skip_connect=skip_connect,
+            **kwargs
+        )
+        self._heads = heads
+        if self.units % self.heads != 0:
+            raise ValueError(f"units need to be divisible by heads.")
+        self._head_units = self.units // self.heads 
+        self._attention_dropout = attention_dropout
+
+    @property 
+    def heads(self):
+        return self._heads 
+    
+    @property 
+    def head_units(self):
+        return self._head_units 
+    
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        """Builds the layer.
+        """
+        self._query_dense = self.get_einsum_dense(
+            'ij,jkh->ikh', (self.head_units, self.heads)
+        )
+        self._key_dense = self.get_einsum_dense(
+            'ij,jkh->ikh', (self.head_units, self.heads)
+        )
+        self._value_dense = self.get_einsum_dense(
+            'ij,jkh->ikh', (self.head_units, self.heads)
+        )
+        self._output_dense = self.get_dense(self.units)
+        self._softmax_dropout = keras.layers.Dropout(self._attention_dropout) 
+
+        if self.has_edge_feature:
+            self._attention_bias_dense_1 = self.get_einsum_dense('ij,jkh->ikh', (1, self.heads))
+
+        if self.has_node_coordinate:
+            node_feature_dim = spec.node['feature'].shape[-1]
+            num_kernels = self.units
+            self._gaussian_loc = self.add_weight(
+                shape=[num_kernels], initializer='zeros', dtype='float32', trainable=True
+            ) 
+            self._gaussian_scale = self.add_weight(
+                shape=[num_kernels], initializer='ones', dtype='float32', trainable=True
+            )
+            self._centrality_dense = self.get_dense(units=node_feature_dim)
+            self._attention_bias_dense_2 = self.get_einsum_dense('ij,jkh->ikh', (1, self.heads))
+
+    def message(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        node_feature = tensor.node['feature']
+
+        if self.has_node_coordinate:
+            euclidean_distance = ops.euclidean_distance(
+                tensor.gather('coordinate', 'target'),
+                tensor.gather('coordinate', 'source'),
+                axis=-1,
+                keepdims=True
+            )
+            gaussian = ops.gaussian(
+                euclidean_distance, self._gaussian_loc, self._gaussian_scale
+            )
+            centrality = keras.ops.segment_sum(gaussian, tensor.edge['target'], tensor.num_nodes)
+            node_feature += self._centrality_dense(centrality)
+
+        query = self._query_dense(node_feature)
+        key = self._key_dense(node_feature)
+        value = self._value_dense(node_feature)
+
+        query = ops.gather(query, tensor.edge['source'])
+        key = ops.gather(key, tensor.edge['target'])
+        value = ops.gather(value, tensor.edge['source'])
+
+        attention_score = keras.ops.sum(query * key, axis=1, keepdims=True)
+        attention_score /= keras.ops.sqrt(float(self.head_units))
+            
+        if self.has_edge_feature:
+            attention_score += self._attention_bias_dense_1(tensor.edge['feature'])
+
+        if self.has_node_coordinate:
+            attention_score += self._attention_bias_dense_2(gaussian)
+
+        attention = ops.edge_softmax(attention_score, tensor.edge['target'])
+        attention = self._softmax_dropout(attention)
+
+        if self.has_node_coordinate:
+            displacement = ops.displacement(
+                tensor.gather('coordinate', 'target'),
+                tensor.gather('coordinate', 'source'),
+                normalize=True,
+                axis=-1,
+                keepdims=True,
+            )
+            attention *= keras.ops.expand_dims(displacement, axis=-1)
+            attention = keras.ops.expand_dims(attention, axis=2)
+            value = keras.ops.expand_dims(value, axis=1)
+
+        message = ops.edge_weight(value, attention)
+        
+        return tensor.update(
+            {
+                'edge': {
+                    'message': message,
+                },
+            }
+        )
+
+    def aggregate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        node_feature = tensor.aggregate('message', mode='sum')
+        if self.has_node_coordinate:
+            shape = (tensor.num_nodes, -1, self.units)
+        else:
+            shape = (tensor.num_nodes, self.units)
+        node_feature = keras.ops.reshape(node_feature, shape)
+        node_feature = self._output_dense(node_feature)
+        if self.has_node_coordinate:
+            node_feature = keras.ops.sum(node_feature, axis=1)
+        return tensor.update(
+            {
+                'node': {
+                    'aggregate': node_feature,
+                },
+                'edge': {
+                    'message': None,
+                }
+            }
+        )
+    
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({
+            "heads": self._heads,
+            'attention_dropout': self._attention_dropout,
+        })
+        return config
+    
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class EGConv(GraphConv):
+
+    """Equivariant graph neural network layer 3D.
+
+    Only supports 3D molecular graphs.
+
+    >>> graph = molcraft.tensors.GraphTensor(
+    ...     context={
+    ...         'size': [2]
+    ...     },
+    ...     node={
+    ...         'feature': [[1.], [2.]],
+    ...         'coordinate': [[0.1, -0.1, 0.5], [1.2, -0.5, 2.1]],
+    ...     },
+    ...     edge={
+    ...         'source': [0, 1],
+    ...         'target': [1, 0],
+    ...     }
+    ... )
+    >>> conv = molcraft.layers.EGConv(units=4)
+    >>> conv(graph)
+        GraphTensor(
+            context={
+                'size': <tf.Tensor: shape=[1], dtype=int32>
+            },
+            node={
+                'feature': <tf.Tensor: shape=[2, 4], dtype=float32>,
+                'coordinate': <tf.Tensor: shape=[2, 3], dtype=float32>
+            },
+            edge={
+                'source': <tf.Tensor: shape=[2], dtype=int32>,
+                'target': <tf.Tensor: shape=[2], dtype=int32>
+            }
+        )
+    """
+
+    def __init__(
+        self, 
+        units: int = 128, 
+        activation: keras.layers.Activation | str | None = 'silu', 
+        use_bias: bool = True,
+        normalize: bool = False,
+        skip_connect: bool = True,
+        **kwargs
+    ) -> None:
+        super().__init__(
+            units=units, 
+            activation=activation,
+            use_bias=use_bias,
+            normalize=normalize, 
+            skip_connect=skip_connect,
+            **kwargs
+        )
+
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        if not self.has_node_coordinate:
+            raise ValueError(
+                'Could not find `coordinate`s in node, '
+                'which is required for Conv3D layers.'
+            )
+        self._message_feedforward_intermediate = self.get_dense(
+            self.units, activation=self.activation
+        )
+        self._message_feedforward_final = self.get_dense(
+            self.units, activation=self.activation
+        )
+
+        self._coord_feedforward_intermediate = self.get_dense(
+            self.units, activation=self.activation
+        ) 
+        self._coord_feedforward_final = self.get_dense(
+            1, use_bias=False, activation='tanh'
+        )
+
+    def message(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        relative_node_coordinate = keras.ops.subtract(
+            tensor.gather('coordinate', 'target'), 
+            tensor.gather('coordinate', 'source')
+        ) 
+        squared_distance = keras.ops.sum(
+            keras.ops.square(relative_node_coordinate), 
+            axis=-1, 
+            keepdims=True
+        )
+    
+        # For numerical stability (i.e., to prevent NaN losses), this implementation of `EGConv3D` 
+        # either needs to apply a `tanh` activation to the output of `self._coord_feedforward_final`, 
+        # or normalize `relative_node_cordinate` as follows:
+        #
+        # norm = keras.ops.sqrt(squared_distance) + keras.backend.epsilon()
+        # relative_node_coordinate /= norm
+        #
+        # For now, this implementation does the former.
+
+        feature = keras.ops.concatenate(
+            [
+                tensor.gather('feature', 'target'), 
+                tensor.gather('feature', 'source'), 
+                squared_distance, 
+            ], 
+            axis=-1
+        )
+        if self.has_edge_feature:
+            feature = keras.ops.concatenate(
+                [
+                    feature,
+                    tensor.edge['feature']
+                ], 
+                axis=-1
+            )
+        message = self._message_feedforward_final(
+            self._message_feedforward_intermediate(feature)
+        )
+
+        relative_node_coordinate = keras.ops.multiply(
+            relative_node_coordinate, 
+            self._coord_feedforward_final(
+                self._coord_feedforward_intermediate(message)
+            )
+        )
+        return tensor.update(
+            {
+                'edge': {
+                    'message': message,
+                    'relative_node_coordinate': relative_node_coordinate
+                }
+            }
+        )
+
+    def aggregate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        coordinate = tensor.node['coordinate']
+        coordinate += tensor.aggregate('relative_node_coordinate', mode='mean')
+
+        # Original implementation seems to apply sum aggregation, which does not
+        # seem work well for this implementation of `EGConv3D`, as it causes 
+        # large output values and large initial losses. The magnitude of the 
+        # aggregated values of a sum aggregation depends on the number of 
+        # neighbors, which may be many and may differ from node to node (or 
+        # graph to graph). Therefore, a mean mean aggregation is performed 
+        # instead:
+        aggregate = tensor.aggregate('message', mode='mean')
+        aggregate = keras.ops.concatenate([aggregate, tensor.node['feature']], axis=-1)
+        # Simply added to silence warning ('no gradients for variables ...')
+        aggregate += (0.0 * keras.ops.sum(coordinate))
+
+        return tensor.update(
+            {
+                'node': {
+                    'aggregate': aggregate, 
+                    'coordinate': coordinate,
+                },
+                'edge': {
+                    'message': None,
+                    'relative_node_coordinate': None
+                }
+            }
+        ) 
+    
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({})
+        return config 
+    
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class Readout(GraphLayer):
+
+    """Readout layer.
+    """
+
+    def __init__(self, mode: str | None = None, **kwargs):
+        kwargs['kernel_initializer'] = None 
+        kwargs['bias_initializer'] = None
+        super().__init__(**kwargs)
+        self.mode = mode
+        if str(self.mode).lower().startswith('sum'):
+            self._reduce_fn = keras.ops.segment_sum
+        elif str(self.mode).lower().startswith('max'):
+            self._reduce_fn = keras.ops.segment_max 
+        elif str(self.mode).lower().startswith('super'):
+            self._reduce_fn = keras.ops.segment_sum
+        else:
+            self._reduce_fn = ops.segment_mean
+
+    def propagate(self, tensor: tensors.GraphTensor) -> tf.Tensor:
+        node_feature = tensor.node['feature']
+        if str(self.mode).lower().startswith('super'):
+            node_feature = keras.ops.where(
+                tensor.node['super'][:, None], node_feature, 0.0
+            )
+        return self._reduce_fn(
+            node_feature, tensor.graph_indicator, tensor.num_subgraphs
+        )
+
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config['mode'] = self.mode 
+        return config 
+    
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class NodeEmbedding(GraphLayer):
+
+    """Node embedding layer.
+
+    Embeds nodes based on its initial features.
+    """
+
+    def __init__(
+        self, 
+        dim: int | None = None, 
+        intermediate_dim: int | None = None,
+        intermediate_activation: str | keras.layers.Activation | None = 'relu',
+        normalize: bool = False,
+        embed_context: bool = False,
+        num_wildcards: int | None = None,
+        **kwargs
+    ) -> None:
+        super().__init__(**kwargs)
+        self.dim = dim
+        self._intermediate_dim = intermediate_dim
+        self._intermediate_activation = keras.activations.get(
+            intermediate_activation
+        )
+        self._normalize = normalize
+        self._embed_context = embed_context
+        self._num_wildcards = num_wildcards
+
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        feature_dim = spec.node['feature'].shape[-1]
+        if not self.dim:
+            self.dim = feature_dim
+        if not self._intermediate_dim:
+            self._intermediate_dim = self.dim * 2 
+        self._node_dense = self.get_dense(
+            self._intermediate_dim, activation=self._intermediate_activation
+        )
+        self._has_wildcard = 'wildcard' in spec.node
+        self._has_super = 'super' in spec.node
+        has_context_feature = 'feature' in spec.context
+        if not has_context_feature:
+            self._embed_context = False 
+        if self._has_super and not self._embed_context:
+            self._super_feature = self.get_weight(
+                shape=[self._intermediate_dim], name='super_node_feature'
+            )
+        if self._embed_context:
+            self._context_dense = self.get_dense(
+                self._intermediate_dim, activation=self._intermediate_activation
+            )
+        if self._has_wildcard:
+            if self._num_wildcards is None:
+                warnings.warn(
+                    "Found wildcards in input, but `num_wildcards` is set to `None`. "
+                    "Automatically setting `num_wildcards` to 1. Please specify `num_wildcards>1` "
+                    "if the layer should distinguish between different types of wildcards."
+                )
+                self._num_wildcards = 1
+            self._wildcard_features = self.get_weight(
+                shape=[self._num_wildcards, self._intermediate_dim], 
+                name='wildcard_node_features'
+            )
+        if not self._normalize:
+            self._norm = keras.layers.Identity()
+        elif str(self._normalize).lower().startswith('layer'):
+            self._norm = keras.layers.LayerNormalization()
+        else:
+            self._norm = keras.layers.BatchNormalization()
+        self._dense = self.get_dense(self.dim)
+
+    def propagate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        feature = self._node_dense(tensor.node['feature'])
+
+        if self._has_super and not self._embed_context:
+            super_mask = keras.ops.expand_dims(tensor.node['super'], 1)
+            super_feature = self._intermediate_activation(self._super_feature)
+            feature = keras.ops.where(super_mask, super_feature, feature)
+
+        if self._embed_context:
+            context_feature = self._context_dense(tensor.context['feature'])
+            feature = ops.scatter_update(feature, tensor.node['super'], context_feature)
+            tensor = tensor.update({'context': {'feature': None}})
+
+        if self._has_wildcard:
+            wildcard = tensor.node['wildcard']
+            wildcard_indices = keras.ops.where(wildcard > 0)[0]
+            wildcard = ops.gather(wildcard, wildcard_indices)
+            wildcard_feature = ops.gather(
+                self._wildcard_features, keras.ops.mod(wildcard-1, self._num_wildcards)
+            )
+            wildcard_feature = self._intermediate_activation(wildcard_feature)
+            feature = ops.scatter_update(feature, wildcard_indices, wildcard_feature)
+
+        feature = self._norm(feature)
+        feature = self._dense(feature)
+
+        return tensor.update({'node': {'feature': feature}})
+
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({
+            'dim': self.dim,
+            'intermediate_dim': self._intermediate_dim,
+            'intermediate_activation': keras.activations.serialize(
+                self._intermediate_activation
+            ),
+            'normalize': self._normalize,
+            'embed_context': self._embed_context,
+            'num_wildcards': self._num_wildcards,
+        })
+        return config
+    
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class EdgeEmbedding(GraphLayer):
+
+    """Edge embedding layer.
+
+    Embeds edges based on its initial features.
+    """
+
+    def __init__(
+        self, 
+        dim: int = None, 
+        intermediate_dim: int | None = None,
+        intermediate_activation: str | keras.layers.Activation | None = 'relu',
+        normalize: bool = False,
+        **kwargs
+    ) -> None:
+        super().__init__(**kwargs)
+        self.dim = dim
+        self._intermediate_dim = intermediate_dim
+        self._intermediate_activation = keras.activations.get(
+            intermediate_activation
+        )
+        self._normalize = normalize
+
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        feature_dim = spec.edge['feature'].shape[-1]
+        if not self.dim:
+            self.dim = feature_dim
+        if not self._intermediate_dim:
+            self._intermediate_dim = self.dim * 2 
+        self._edge_dense = self.get_dense(
+            self._intermediate_dim, activation=self._intermediate_activation
+        ) 
+        self._self_loop_feature = self.get_weight(
+            shape=[self._intermediate_dim], name='self_loop_edge_feature'
+        )
+        self._has_super = 'super' in spec.edge
+        if self._has_super:
+            self._super_feature = self.get_weight(
+                shape=[self._intermediate_dim], name='super_edge_feature'
+            )
+        if not self._normalize:
+            self._norm = keras.layers.Identity()
+        elif str(self._normalize).lower().startswith('layer'):
+            self._norm = keras.layers.LayerNormalization()
+        else:
+            self._norm = keras.layers.BatchNormalization()
+        self._dense = self.get_dense(self.dim)
+
+    def propagate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        feature = self._edge_dense(tensor.edge['feature'])
+
+        if self._has_super:
+            super_mask = keras.ops.expand_dims(tensor.edge['super'], 1)
+            super_feature = self._intermediate_activation(self._super_feature)
+            feature = keras.ops.where(super_mask, super_feature, feature)
+
+        self_loop_mask = keras.ops.expand_dims(tensor.edge['source'] == tensor.edge['target'], 1)
+        self_loop_feature = self._intermediate_activation(self._self_loop_feature)
+        feature = keras.ops.where(self_loop_mask, self_loop_feature, feature)
+        feature = self._norm(feature)
+        feature = self._dense(feature)
+        return tensor.update({'edge': {'feature': feature}})
+
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({
+            'dim': self.dim,
+            'intermediate_dim': self._intermediate_dim,
+            'intermediate_activation': keras.activations.serialize(
+                self._intermediate_activation
+            ),
+            'normalize': self._normalize,
+        })
+        return config
+    
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class AddContext(GraphLayer):
+
+    """Context adding layer.
+
+    Adds context to super nodes or nodes, depending on whether super nodes exist.
+    """
+
+    def __init__(
+        self, 
+        field: str = 'feature',
+        intermediate_dim: int | None = None,
+        intermediate_activation: str | keras.layers.Activation | None = 'relu',
+        drop: bool = False,
+        normalize: bool = False,
+        num_categories: int | None = None,
+        **kwargs
+    ) -> None:
+        super().__init__(**kwargs)
+        self._field = field
+        self._drop = drop
+        self._intermediate_dim = intermediate_dim
+        self._intermediate_activation = keras.activations.get(
+            intermediate_activation
+        )
+        self._normalize = normalize
+        self._num_categories = num_categories
+        
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        is_categorical = spec.context[self._field].dtype.is_integer
+        if is_categorical and not self._num_categories:
+            raise ValueError(
+                f'Found context ({self._field}) to be categorical (`int` dtype), but `num_categories`'
+                'to be `None`. Please specify `num_categories` for categorical context.'
+            )
+        elif not is_categorical and self._num_categories:
+            warnings.warn(
+                f'`num_categories` is set to {self._num_categories}, but found context to be '
+                'continuous (`float` dtype). Layer will cast context from `float` to `int` '
+                'before one-hot encoding it.'
+            )
+        feature_dim = spec.node['feature'].shape[-1]
+        self._has_super_node = 'super' in spec.node
+        if self._intermediate_dim is None:
+            self._intermediate_dim = feature_dim * 2
+        self._intermediate_dense = self.get_dense(
+            self._intermediate_dim, activation=self._intermediate_activation
+        )
+        self._final_dense = self.get_dense(feature_dim)
+        if not self._normalize:
+            self._intermediate_norm = keras.layers.Identity()
+        elif str(self._normalize).lower().startswith('layer'):
+            self._intermediate_norm = keras.layers.LayerNormalization()
+        else:
+            self._intermediate_norm = keras.layers.BatchNormalization()
+
+    def propagate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        context = tensor.context[self._field]
+        if self._num_categories:
+            if context.dtype.is_floating:
+                context = keras.ops.cast(context, dtype=tensor.edge['source'].dtype)
+            context = keras.utils.to_categorical(context, self._num_categories)
+        elif len(keras.ops.shape(context)) == 1:
+            context = keras.ops.expand_dims(context, axis=1)
+        context = self._intermediate_dense(context)
+        context = self._intermediate_norm(context)
+        context = self._final_dense(context)
+        if self._has_super_node:
+            node_feature = ops.scatter_add(
+                tensor.node['feature'], tensor.node['super'], context
+            )
+        else:
+            node_feature = (
+                tensor.node['feature'] + ops.gather(context, tensor.graph_indicator)
+            )
+        data = {'node': {'feature': node_feature}}
+        if self._drop:
+            data['context'] = {self._field: None}
+        return tensor.update(data)
+
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update({
+            'field': self._field,
+            'intermediate_dim': self._intermediate_dim,
+            'intermediate_activation': keras.activations.serialize(
+                self._intermediate_activation
+            ),
+            'drop': self._drop,
+            'normalize': self._normalize,
+            'num_categories': self._num_categories,
+        })
+        return config
+
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class GraphNetwork(GraphLayer):
+
+    """Graph neural network.
+
+    Sequentially calls graph layers (`GraphLayer`) and concatenates its output. 
+
+    Arguments:
+        layers (list):
+            A list of graph layers.
+    """
+
+    def __init__(self, layers: list[GraphLayer], **kwargs) -> None:
+        super().__init__(**kwargs)
+        self.layers = layers
+        self._update_edge_feature = False
+
+    def build(self, spec: tensors.GraphTensor.Spec) -> None:
+        units = self.layers[0].units 
+        node_feature_dim = spec.node['feature'].shape[-1]
+        self._update_node_feature = node_feature_dim != units 
+        if self._update_node_feature:
+            warnings.warn(
+                'Node feature dim does not match `units` of the first layer. '
+                'Applying a projection layer to node features to match `units`.',
+            )
+            self._node_dense = self.get_dense(units)
+        self._has_edge_feature = 'feature' in spec.edge 
+        if self._has_edge_feature:
+            edge_feature_dim = spec.edge['feature'].shape[-1]
+            self._update_edge_feature = edge_feature_dim != units
+            if self._update_edge_feature:
+                warnings.warn(
+                    'Edge feature dim does not match `units` of the first layer. '
+                    'Applying projection layer to edge features to match `units`.'
+                )
+                self._edge_dense = self.get_dense(units)
+
+    def propagate(self, tensor: tensors.GraphTensor) -> tensors.GraphTensor:
+        x = tensors.to_dict(tensor)
+        if self._update_node_feature:
+            x['node']['feature'] = self._node_dense(tensor.node['feature'])
+        if self._has_edge_feature and self._update_edge_feature:
+            x['edge']['feature'] = self._edge_dense(tensor.edge['feature'])
+        outputs = [x['node']['feature']]
+        for layer in self.layers:
+            x = layer(x)
+            outputs.append(x['node']['feature'])
+        return tensor.update(
+            {
+                'node': {
+                    'feature': keras.ops.concatenate(outputs, axis=-1)
+                }    
+            }
+        )
+    
+    def tape_propagate(
+        self,
+        tensor: tensors.GraphTensor,
+        tape: tf.GradientTape,
+        training: bool | None = None,
+    ) -> tuple[tensors.GraphTensor, list[tf.Tensor]]:
+        """Performs the propagation with a `GradientTape`.
+
+        Performs the same forward pass as `propagate` but with a `GradientTape`
+        watching intermediate node features.
+
+        Arguments:
+            tensor (tensors.GraphTensor):
+                The graph input.
+        """
+        if isinstance(tensor, tensors.GraphTensor):
+            x = tensors.to_dict(tensor)
+        else:
+            x = tensor
+        if self._update_node_feature:
+            x['node']['feature'] = self._node_dense(tensor.node['feature'])
+        if self._update_edge_feature:
+            x['edge']['feature'] = self._edge_dense(tensor.edge['feature'])
+        tape.watch(x['node']['feature'])
+        outputs = [x['node']['feature']]
+        for layer in self.layers:
+            x = layer(x, training=training)
+            tape.watch(x['node']['feature'])
+            outputs.append(x['node']['feature'])
+
+        tensor = tensor.update(
+            {
+                'node': {
+                    'feature': keras.ops.concatenate(outputs, axis=-1)
+                }
+            }
+        )
+        return tensor, outputs
+    
+    def get_config(self) -> dict:
+        config = super().get_config()
+        config.update(
+            {
+                'layers': [
+                    keras.layers.serialize(layer) for layer in self.layers
+                ]
+            }
+        )
+        return config
+
+    @classmethod
+    def from_config(cls, config: dict) -> 'GraphNetwork':
+        config['layers'] = [
+            keras.layers.deserialize(layer) for layer in config['layers']
+        ]
+        return super().from_config(config)
+    
+
+@keras.saving.register_keras_serializable(package='molcraft')
+class GaussianParams(keras.layers.Dense):
+    '''Gaussian parameters.
+
+    Computes loc and scale via a dense layer. Should be implemented
+    as the last layer in a model and paired with `losses.GaussianNLL`.
+
+    The loc and scale parameters (resulting from this layer) are concatenated
+    together along the last axis, resulting in a single output tensor. 
+
+    Args:
+        events (int):
+            The number of events. If the model makes a single prediction per example,
+            then the number of events should be 1. If the model makes multiple predictions 
+            per example, then the number of events should be greater than 1. 
+            Default to 1.
+        kwargs:
+            See `keras.layers.Dense` documentation. `activation` will be applied
+            to `loc` only. `scale` is automatically softplus activated.
+    '''
+    def __init__(self, events: int = 1, **kwargs):
+        units = kwargs.pop('units', None)
+        activation = kwargs.pop('activation', None)
+        if units:
+            if units % 2 != 0:
+                raise ValueError(
+                    '`units` needs to be divisble by 2 as `units` = 2 x `events`.'
+                )
+        else:
+            units = int(events * 2)
+        super().__init__(units=units, **kwargs)
+        self.events = events
+        self.loc_activation = keras.activations.get(activation)
+
+    def call(self, inputs, **kwargs):
+        loc_and_scale = super().call(inputs, **kwargs)
+        loc = loc_and_scale[..., :self.events]
+        scale = loc_and_scale[..., self.events:]
+        scale = keras.ops.softplus(scale) + keras.backend.epsilon()
+        loc = self.loc_activation(loc)
+        return keras.ops.concatenate([loc, scale], axis=-1)
+
+    def get_config(self):
+        config = super().get_config()
+        config['events'] = self.events
+        config['units'] = None
+        config['activation'] = keras.activations.serialize(self.loc_activation)
+        return config
+    
+
+def Input(spec: tensors.GraphTensor.Spec) -> dict:
+    """Used to specify inputs to a functional model.
+
+    Example:
+
+    >>> import molcraft 
+    >>> import keras
+    >>> 
+    >>> featurizer = molcraft.featurizers.MolGraphFeaturizer()
+    >>> graph = featurizer([('N[C@@H](C)C(=O)O', 1.0), ('N[C@@H](CS)C(=O)O', 2.0)])
+    >>> 
+    >>> model = molcraft.models.GraphModel.from_layers(
+    ...     molcraft.layers.Input(graph.spec),
+    ...     molcraft.layers.NodeEmbedding(128),
+    ...     molcraft.layers.EdgeEmbedding(128),
+    ...     molcraft.layers.GraphTransformer(128),
+    ...     molcraft.layers.GraphTransformer(128),
+    ...     molcraft.layers.Readout('mean'),
+    ...     molcraft.layers.Dense(1)
+    ... ])
+    """
+    
+    # Currently, Keras (3.8.0) does not support extension types.
+    # So for now, this function will unpack the `GraphTensor.Spec` and 
+    # return a dictionary of nested tensor specs. However, the corresponding 
+    # nest of tensors will temporarily be converted to a `GraphTensor` by the 
+    # `GraphLayer`, to levarage the utility of a `GraphTensor` object. 
+    inputs = {}
+    for outer_field, data in spec.__dict__.items():
+        inputs[outer_field] = {}
+        for inner_field, nested_spec in data.items():
+            if inner_field in ['label', 'sample_weight']:
+                # Remove label and sample_weight from the symbolic input as
+                # a functional model is strict for what input can be passed.
+                continue
+            kwargs = {
+                'shape': nested_spec.shape[1:],
+                'dtype': nested_spec.dtype,
+                'name': f'{outer_field}_{inner_field}'
+            }
+            if isinstance(nested_spec, tf.RaggedTensorSpec):
+                # kwargs['ragged'] = True
+                raise ValueError(
+                    'Graph layers only supports graph input with nested `tf.Tensor` values.'
+                )
+            try:
+                inputs[outer_field][inner_field] = keras.Input(**kwargs)
+            except TypeError:
+                raise ValueError(
+                    "`keras.Input` does not currently support ragged tensors. For now, "
+                    "pass the `Spec` of a 'flat' `GraphTensor` to `GNNInput`." 
+                )
+    return inputs
+
+
+def _serialize_spec(spec: tensors.GraphTensor.Spec) -> dict:
+    serialized_spec = {}
+    for outer_field, data in spec.__dict__.items():
+        serialized_spec[outer_field] = {}
+        for inner_field, inner_spec in data.items():
+            if inner_field in ['label', 'sample_weight']:
+                continue
+            serialized_spec[outer_field][inner_field] = {
+                'shape': inner_spec.shape.as_list(), 
+                'dtype': inner_spec.dtype.name, 
+                'name': inner_spec.name,
+            }
+    return serialized_spec
+
+def _deserialize_spec(serialized_spec: dict) -> tensors.GraphTensor.Spec:
+    deserialized_spec = {}
+    for outer_field, data in serialized_spec.items():
+        deserialized_spec[outer_field] = {}
+        for inner_field, inner_spec in data.items():
+            deserialized_spec[outer_field][inner_field] = tf.TensorSpec(
+                inner_spec['shape'], inner_spec['dtype'], inner_spec['name']
+            )
+    return tensors.GraphTensor.Spec(**deserialized_spec)
+
+def _spec_from_inputs(inputs):
+    symbolic_inputs = keras.backend.is_keras_tensor(
+        tf.nest.flatten(inputs)[0]
+    )
+    if not symbolic_inputs:
+        nested_specs = tf.nest.map_structure(
+            tf.type_spec_from_value, inputs
+        )
+    else:
+        nested_specs = tf.nest.map_structure(
+            lambda t: tf.TensorSpec(t.shape, t.dtype), inputs
+        )
+    if isinstance(nested_specs, tensors.GraphTensor.Spec):
+        spec = nested_specs
+        return spec
+    return tensors.GraphTensor.Spec(**nested_specs)
+
+def _propagate_kwargs(func) -> bool:
+    signature = inspect.signature(func)
+    return any(
+        (param.kind == inspect.Parameter.VAR_KEYWORD) or (param.name == 'training')
+        for param in signature.parameters.values()
+    )