rustyml 0.17.0

//! RepeatVector layer that repeats a feature vector into a sequence, and parks the input shape
//! for backpropagation
//!
//! The layer inserts a step axis after the batch axis. Its backward pass sums gradients
//! over that new axis.

use crate::error::Error;
use crate::neural_network::layers::ParamCounts;
use crate::neural_network::layers::validation::start_build;
use crate::neural_network::layers::{
    built_layer_shape_functions, no_trainable_parameters_layer_functions,
};
use crate::neural_network::traits::{LayerBase, UnaryLayer};
use crate::neural_network::{Ctx, Shape, Tensor};
use ndarray::{Axis, IxDyn};

/// Repeats each feature vector `n` times along a new step axis
///
/// The input shape is `[batch_size, features]` and the output shape is
/// `[batch_size, n, features]`. Every step holds the same vector. The batch axis passes through
/// unchanged, so 1 layer instance serves every batch size
///
/// The layer holds no parameter. Its use is the decoder side of an encoder-decoder model. A
/// recurrent layer returns its last hidden state as a rank-2 tensor by default, and a
/// recurrent layer needs a rank-3 input. This layer bridges the 2, so an `LSTM` can feed another
/// `LSTM`
///
/// # Examples
///
/// ```rust
/// use rustyml::neural_network::Shape;
/// use rustyml::neural_network::sequential::SequentialBuilder;
/// use rustyml::neural_network::layers::*;
/// use rustyml::neural_network::optimizers::*;
/// use rustyml::neural_network::losses::*;
/// use ndarray::Array2;
///
/// // A rank-2 input: 2 samples of 4 features each.
/// let x = Array2::ones((2, 4)).into_dyn();
///
/// let mut model = SequentialBuilder::new()
///     .add(RepeatVector::new(3).unwrap())
///     .build(&Shape::known(x.shape()))
///     .unwrap();
/// model.compile(SGD::new(0.01, 0.0, false, 0.0).unwrap(), MeanSquaredError::new());
///
/// // View model structure.
/// model.summary();
///
/// let sequence = model.predict(&x).unwrap();
///
/// // 1 vector of 4 features becomes 3 steps of the same 4 features.
/// assert_eq!(sequence.shape(), &[2, 3, 4]);
/// ```
#[derive(Debug)]
pub struct RepeatVector {
    /// Number of steps the output holds
    n: usize,
    /// Shape the layer was built for, batch axis first. `None` before the build
    built: Option<Shape>,
}

impl RepeatVector {
    /// Creates a new RepeatVector layer
    ///
    /// # Parameters
    ///
    /// - `n` - Number of steps the output holds. Must be greater than 0
    ///
    /// # Returns
    ///
    /// - `Result<Self, Error>` - New `RepeatVector` layer instance
    ///
    /// # Errors
    ///
    /// - `Error::InvalidParameter` - If `n` is 0
    pub fn new(n: usize) -> Result<Self, Error> {
        if n == 0 {
            return Err(Error::invalid_parameter(
                "n",
                "is 0, and the output must hold at least 1 step",
            ));
        }

        Ok(RepeatVector { n, built: None })
    }

    /// Checks the rank and the element count of a tensor entering the layer
    ///
    /// # Errors
    ///
    /// - `Error::InvalidInput` - If the rank is not 2
    /// - `Error::EmptyInput` - If either axis has an extent of 0
    fn validate(&self, input: &Tensor) -> Result<(), Error> {
        if input.ndim() != 2 {
            return Err(Error::invalid_input(format!(
                "RepeatVector layer expects a 2D input [batch_size, features], got a {}D tensor",
                input.ndim()
            )));
        }
        if input.is_empty() {
            return Err(Error::empty_input("input tensor"));
        }
        Ok(())
    }

    /// Writes `n` copies of every feature vector into a new tensor
    fn repeat(&self, input: &Tensor) -> Tensor {
        let shape = input.shape();
        let mut output = Tensor::zeros(IxDyn(&[shape[0], self.n, shape[1]]));
        // The `[batch, 1, features]` view broadcasts across the new step axis
        output.assign(&input.view().insert_axis(Axis(1)));
        output
    }
}

impl LayerBase for RepeatVector {
    fn layer_type(&self) -> &str {
        "RepeatVector"
    }

    built_layer_shape_functions!();

    no_trainable_parameters_layer_functions!();
}

impl UnaryLayer for RepeatVector {
    /// Records the shape the layer repeats. The layer holds no array, so nothing is allocated.
    /// The shape algebra checks the rank
    fn build(&mut self, input: &Shape) -> Result<(), Error> {
        let Some(built) = start_build(&self.built, "RepeatVector", input)? else {
            return Ok(());
        };
        self.compute_output_shape(&built)?;
        self.built = Some(built);
        Ok(())
    }

    fn forward(&self, input: &Tensor, ctx: &mut Ctx) -> Result<Tensor, Error> {
        self.validate(input)?;

        if ctx.is_training() {
            ctx.push_cache("RepeatVector", input.shape().to_vec());
        }

        Ok(self.repeat(input))
    }

    fn backward(&self, grad_output: &Tensor, ctx: &mut Ctx) -> Result<Tensor, Error> {
        let input_shape: Vec<usize> = ctx.pop_cache("RepeatVector")?;

        let expected = [input_shape[0], self.n, input_shape[1]];
        if grad_output.shape() != expected {
            return Err(Error::shape_mismatch(expected, grad_output.shape()));
        }

        // Every step reads the same input vector, so the input gradient is the sum over the
        // step axis
        Ok(grad_output.sum_axis(Axis(1)))
    }

    /// The layer inserts a step axis of `n` copies between the batch axis and the features
    fn compute_output_shape(&self, input: &Shape) -> Result<Shape, Error> {
        input.check_rank("RepeatVector", 2)?;
        let (batch, tail) = input.split_batch("RepeatVector")?;
        Ok(Shape::from_batch(batch, &[self.n, tail[0]]))
    }
}