burn-nn 0.22.0

use burn_core as burn;

use crate::PaddingConfig2d;
use burn::config::Config;
use burn::module::Module;
use burn::module::{Content, DisplaySettings, ModuleDisplay};
use burn::tensor::Tensor;
use burn::tensor::ops::AvgPoolOptions;

use burn::tensor::module::avg_pool2d;

/// Configuration to create a [2D avg pooling](AvgPool2d) layer using the [init function](AvgPool2dConfig::init).
#[derive(Config, Debug)]
pub struct AvgPool2dConfig {
    /// The size of the kernel.
    pub kernel_size: [usize; 2],
    /// The strides.
    #[config(default = "kernel_size")]
    pub strides: [usize; 2],
    /// The padding configuration.
    ///
    /// Supports symmetric and asymmetric padding. `Same` padding with even kernel sizes
    /// will automatically use asymmetric padding to preserve input dimensions.
    #[config(default = "PaddingConfig2d::Valid")]
    pub padding: PaddingConfig2d,
    /// If the padding is counted in the denominator when computing the average.
    #[config(default = "true")]
    pub count_include_pad: bool,
    /// If true, use ceiling instead of floor for output size calculation.
    #[config(default = "false")]
    pub ceil_mode: bool,
}

/// Applies a 2D avg pooling over input tensors.
///
/// Should be created with [AvgPool2dConfig](AvgPool2dConfig).
///
/// # Remarks
///
/// The zero-padding values will be included in the calculation
/// of the average. This means that the zeros are counted as
/// legitimate values, and they contribute to the denominator
/// when calculating the average. This is equivalent to
/// `torch.nn.AvgPool2d` with `count_include_pad=True`.
#[derive(Module, Debug)]
#[module(custom_display)]
pub struct AvgPool2d {
    /// Stride of the pooling.
    pub stride: [usize; 2],
    /// Size of the kernel.
    pub kernel_size: [usize; 2],
    /// Padding configuration.
    #[module(skip)]
    pub padding: PaddingConfig2d,
    /// If the padding is counted in the denominator when computing the average.
    pub count_include_pad: bool,
    /// If true, use ceiling instead of floor for output size calculation.
    pub ceil_mode: bool,
}

impl ModuleDisplay for AvgPool2d {
    fn custom_settings(&self) -> Option<DisplaySettings> {
        DisplaySettings::new()
            .with_new_line_after_attribute(false)
            .optional()
    }

    fn custom_content(&self, content: Content) -> Option<Content> {
        content
            .add("kernel_size", &alloc::format!("{:?}", self.kernel_size))
            .add("stride", &alloc::format!("{:?}", self.stride))
            .add_debug_attribute("padding", &self.padding)
            .add("count_include_pad", &self.count_include_pad)
            .add("ceil_mode", &self.ceil_mode)
            .optional()
    }
}

impl AvgPool2dConfig {
    /// Initialize a new [avg pool 2d](AvgPool2d) module.
    pub fn init(&self) -> AvgPool2d {
        AvgPool2d {
            stride: self.strides,
            kernel_size: self.kernel_size,
            padding: self.padding.clone(),
            count_include_pad: self.count_include_pad,
            ceil_mode: self.ceil_mode,
        }
    }
}

impl AvgPool2d {
    /// Applies the forward pass on the input tensor.
    ///
    /// See [avg_pool2d](burn::tensor::module::avg_pool2d) for more information.
    ///
    /// # Shapes
    ///
    /// - input: `[batch_size, channels, height_in, width_in]`
    /// - output: `[batch_size, channels, height_out, width_out]`
    pub fn forward(&self, input: Tensor<4>) -> Tensor<4> {
        let [_batch_size, _channels_in, height_in, width_in] = input.dims();
        let (padding_height, padding_width) = self.padding.calculate_padding_2d_pairs(
            height_in,
            width_in,
            &self.kernel_size,
            &self.stride,
        );

        avg_pool2d(
            input,
            AvgPoolOptions::new(self.kernel_size)
                .with_stride(self.stride)
                .with_padding_pairs([padding_height, padding_width])
                .with_count_include_pad(self.count_include_pad)
                .with_ceil_mode(self.ceil_mode),
        )
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use burn::tensor::{Device, TensorData, Tolerance};
    use rstest::rstest;

    #[test]
    fn same_with_even_kernel_uses_asymmetric_padding() {
        let device = Default::default();
        let config = AvgPool2dConfig::new([2, 2])
            .with_strides([1, 1])
            .with_padding(PaddingConfig2d::Same);
        let pool = config.init();

        // Input: [batch=1, channels=2, height=5, width=5]
        let input = Tensor::<4>::ones([1, 2, 5, 5], &device);
        let output = pool.forward(input);

        // Same padding should preserve spatial dimensions
        assert_eq!(output.dims(), [1, 2, 5, 5]);
    }

    #[test]
    fn display() {
        let config = AvgPool2dConfig::new([3, 3]);

        let layer = config.init();

        assert_eq!(
            alloc::format!("{layer}"),
            "AvgPool2d {kernel_size: [3, 3], stride: [3, 3], padding: Valid, count_include_pad: true, ceil_mode: false}"
        );
    }

    #[rstest]
    #[case([2, 2])]
    #[case([1, 2])]
    fn default_strides_match_kernel_size(#[case] kernel_size: [usize; 2]) {
        let config = AvgPool2dConfig::new(kernel_size);

        assert_eq!(
            config.strides, kernel_size,
            "Expected strides ({:?}) to match kernel size ({:?}) in default AvgPool2dConfig::new constructor",
            config.strides, config.kernel_size
        );
    }

    #[test]
    fn asymmetric_padding_forward() {
        let device = Default::default();
        // Create avg pool with asymmetric padding: top=1, left=2, bottom=3, right=4
        let config = AvgPool2dConfig::new([3, 3])
            .with_strides([1, 1])
            .with_padding(PaddingConfig2d::Explicit(1, 2, 3, 4));
        let pool = config.init();

        // Input: [batch=1, channels=2, height=4, width=5]
        let input = Tensor::<4>::ones([1, 2, 4, 5], &device);
        let output = pool.forward(input);

        // Height: 4 + 1 + 3 = 8, output = (8 - 3) / 1 + 1 = 6
        // Width: 5 + 2 + 4 = 11, output = (11 - 3) / 1 + 1 = 9
        assert_eq!(output.dims(), [1, 2, 6, 9]);
    }

    #[test]
    fn asymmetric_padding_excludes_pad_from_average() {
        let device = Default::default();
        let input = Tensor::<1>::from_data(
            [
                1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0,
                16.0, 17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0, 25.0,
            ],
            &device,
        )
        .reshape([1, 1, 5, 5]);
        let pool = AvgPool2dConfig::new([2, 2])
            .with_strides([2, 2])
            .with_padding(PaddingConfig2d::Explicit(0, 0, 1, 1))
            .with_count_include_pad(false)
            .init();

        let output = pool.forward(input);
        let expected =
            TensorData::from([[[[4.0f32, 6.0, 7.5], [14.0, 16.0, 17.5], [21.5, 23.5, 25.0]]]]);

        output.to_data().assert_eq(&expected, true);
    }

    #[test]
    fn fully_padded_windows_are_finite_and_count_flag_is_preserved() {
        let device = Default::default();
        let input = Tensor::from_data([[[[4.0f32]]]], &device);

        let include_pad = AvgPool2dConfig::new([2, 2])
            .with_strides([1, 1])
            .with_padding(PaddingConfig2d::Explicit(2, 0, 0, 1))
            .with_count_include_pad(true)
            .init()
            .forward(input.clone());
        include_pad
            .to_data()
            .assert_eq(&TensorData::from([[[[0.0f32], [1.0]]]]), true);

        let exclude_pad = AvgPool2dConfig::new([2, 2])
            .with_strides([1, 1])
            .with_padding(PaddingConfig2d::Explicit(2, 0, 0, 1))
            .with_count_include_pad(false)
            .init()
            .forward(input);
        exclude_pad
            .to_data()
            .assert_eq(&TensorData::from([[[[0.0f32], [4.0]]]]), true);
    }

    #[test]
    fn asymmetric_padding_excludes_pad_from_gradients() {
        let device = Device::default().autodiff();
        let input = Tensor::ones([1, 1, 5, 5], &device).require_grad();
        let pool = AvgPool2dConfig::new([2, 2])
            .with_strides([2, 2])
            .with_padding(PaddingConfig2d::Explicit(0, 0, 1, 1))
            .with_count_include_pad(false)
            .init();

        let output = pool.forward(input.clone());
        let gradients = output.sum().backward();
        let expected = TensorData::from([[[
            [0.25f32, 0.25, 0.25, 0.25, 0.5],
            [0.25, 0.25, 0.25, 0.25, 0.5],
            [0.25, 0.25, 0.25, 0.25, 0.5],
            [0.25, 0.25, 0.25, 0.25, 0.5],
            [0.5, 0.5, 0.5, 0.5, 1.0],
        ]]]);

        input
            .grad(&gradients)
            .unwrap()
            .to_data()
            .assert_approx_eq::<f32>(&expected, Tolerance::default());
    }

    #[test]
    fn symmetric_explicit_padding_forward() {
        let device = Default::default();
        // Create avg pool with symmetric explicit padding: top=2, left=2, bottom=2, right=2
        let config = AvgPool2dConfig::new([3, 3])
            .with_strides([1, 1])
            .with_padding(PaddingConfig2d::Explicit(2, 2, 2, 2));
        let pool = config.init();

        // Input: [batch=1, channels=2, height=4, width=5]
        let input = Tensor::<4>::ones([1, 2, 4, 5], &device);
        let output = pool.forward(input);

        // Height: 4 + 2 + 2 = 8, output = (8 - 3) / 1 + 1 = 6
        // Width: 5 + 2 + 2 = 9, output = (9 - 3) / 1 + 1 = 7
        assert_eq!(output.dims(), [1, 2, 6, 7]);
    }
}