✨ Create NeuraNetwork traits, WIP

src/gradient_solver/backprop.rs

@@ -1,8 +1,8 @@
 use num::ToPrimitive;
 
 use crate::{
-    derivable::NeuraLoss, layer::NeuraTrainableLayerBackprop, layer::NeuraTrainableLayerSelf,
+    derivable::NeuraLoss, layer::*,
-    network::NeuraOldTrainableNetworkBase,
+    network::*,
 };
 
 use super::*;
@@ -91,9 +91,46 @@ impl<
     }
 }
 
+trait BackpropRecurse<Input, Network, Gradient> {
+    fn recurse(&self, network: &Network, input: &Input) -> (Input, Gradient);
+}
+
+impl<Input, Loss: NeuraLoss<Input>> BackpropRecurse<Input, (), ()> for (&NeuraBackprop<Loss>, &Loss::Target) {
+    fn recurse(&self, _network: &(), input: &Input) -> (Input, ()) {
+        (self.0.loss.nabla(self.1, input), ())
+    }
+}
+
+impl<
+    Input: Clone,
+    Network: NeuraNetworkRec + NeuraNetwork<Input> + NeuraTrainableLayerBase<Input>,
+    Loss,
+    Target
+> BackpropRecurse<Input, Network, Network::Gradient> for (&NeuraBackprop<Loss>, &Target)
+where
+    // Verify that we can traverse recursively
+    for<'a> (&'a NeuraBackprop<Loss>, &'a Target): BackpropRecurse<Network::NodeOutput, Network::NextNode, <Network::NextNode as NeuraTrainableLayerBase<Network::NodeOutput>>::Gradient>,
+    // Verify that the current layer implements the right traits
+    Network::Layer: NeuraTrainableLayerSelf<Network::LayerInput> + NeuraTrainableLayerBackprop<Network::LayerInput>,
+    // Verify that the layer output can be cloned
+    <Network::Layer as NeuraLayer<Network::LayerInput>>::Output: Clone,
+    Network::NextNode: NeuraTrainableLayerBase<Network::NodeOutput>,
+{
+    fn recurse(&self, network: &Network, input: &Input) -> (Input, Network::Gradient) {
+        let layer_input = network.map_input(input);
+        let (layer_output, layer_intermediary) = network.get_layer().eval_training(layer_input.as_ref());
+        let output = network.map_output(input, &layer_output);
+
+        let (epsilon_in, gradient_rec) = self.recurse(network.get_next(), output.as_ref());
+
+        todo!()
+    }
+}
+
 #[cfg(test)]
 mod test {
     use approx::assert_relative_eq;
+    use nalgebra::dvector;
 
     use super::*;
     use crate::{
@@ -161,4 +198,12 @@ mod test {
             assert_relative_eq!(gradient1_actual.as_slice(), gradient1_expected.as_slice());
         }
     }
+
+    #[test]
+    fn test_recursive() {
+        let backprop = NeuraBackprop::new(Euclidean);
+        let target = dvector![0.0];
+
+        (&backprop, &target).recurse(&(), &dvector![0.0]);
+    }
 }

src/layer/mod.rs

@@ -71,6 +71,7 @@ pub trait NeuraTrainableLayerBase<Input>: NeuraLayer<Input> {
     /// Applies `δW_l` to the weights of the layer
     fn apply_gradient(&mut self, gradient: &Self::Gradient);
 
+    // TODO: move this into another trait
     fn eval_training(&self, input: &Input) -> (Self::Output, Self::IntermediaryRepr);
 
     /// Arbitrary computation that can be executed at the start of an epoch

src/network/mod.rs

@@ -5,6 +5,9 @@ use crate::{
 // pub mod residual;
 pub mod sequential;
 
+mod traits;
+pub use traits::*;
+
 // TODO: extract regularize from this, so that we can drop the trait constraints on NeuraSequential's impl
 pub trait NeuraOldTrainableNetworkBase<Input>: NeuraLayer<Input> {
     type Gradient: NeuraVectorSpace;

src/network/traits.rs

@@ -0,0 +1,44 @@
+use std::borrow::Cow;
+
+use super::*;
+
+/// This trait has to be non-generic, to ensure that no downstream crate can implement it for foreign types,
+/// as that would otherwise cause infinite recursion when dealing with `NeuraNetworkRec`.
+pub trait NeuraNetworkBase {
+    /// The type of the enclosed layer
+    type Layer;
+
+    fn get_layer(&self) -> &Self::Layer;
+}
+
+pub trait NeuraNetwork<NodeInput: Clone>: NeuraNetworkBase
+where
+    Self::Layer: NeuraLayer<Self::LayerInput>,
+    <Self::Layer as NeuraLayer<Self::LayerInput>>::Output: Clone
+{
+    /// The type of the input to `Self::Layer`
+    type LayerInput: Clone;
+
+    /// The type of the output of this node
+    type NodeOutput: Clone;
+
+    /// Maps the input of network node to the enclosed layer
+    fn map_input<'a>(&'_ self, input: &'a NodeInput) -> Cow<'a, Self::LayerInput>;
+    /// Maps the output of the enclosed layer to the output of the network node
+    fn map_output<'a>(&'_ self, input: &'_ NodeInput, layer_output: &'a <Self::Layer as NeuraLayer<Self::LayerInput>>::Output) -> Cow<'a, Self::NodeOutput>;
+
+    /// Maps a gradient in the format of the node's output into the format of the enclosed layer's output
+    fn map_gradient_in<'a>(&'_ self, input: &'_ NodeInput, gradient_in: &'a Self::NodeOutput) -> Cow<'a, <Self::Layer as NeuraLayer<Self::LayerInput>>::Output>;
+    /// Maps a gradient in the format of the enclosed layer's input into the format of the node's input
+    fn map_gradient_out<'a>(&'_ self, input: &'_ NodeInput, gradient_in: &'_ Self::NodeOutput, gradient_out: &'a Self::LayerInput) -> Cow<'a, NodeInput>;
+}
+
+pub trait NeuraNetworkRec: NeuraNetworkBase {
+    /// The type of the children network, it does not need to implement `NeuraNetworkBase`,
+    /// although many functions will expect it to be either `()` or an implementation of `NeuraNetworkRec`.
+    type NextNode;
+
+    fn get_next(&self) -> &Self::NextNode;
+
+
+}