relayrl_types 0.7.2

use std::collections::HashMap;
use std::convert::TryInto;
use std::path::PathBuf;
use std::sync::Arc;

#[cfg(feature = "ndarray-backend")]
use crate::data::tensor::NdArrayDType;
#[cfg(feature = "tch-backend")]
use crate::data::tensor::TchDType;

use crate::data::action::RelayRLData;
use crate::data::tensor::{
    AnyBurnTensor, BackendMatcher, BoolBurnTensor, DType, DeviceType, FloatBurnTensor,
    IntBurnTensor,
};
use burn_tensor::{Shape, backend::Backend};

use crate::model::{ModelError, ModelModule};

/// Converts a dictionary of auxiliary data into a HashMap with String keys and RelayRLData values.
///
/// This function works with generic Rust types (i32, f64, TensorData) instead of Torch-specific IValue.
/// It's designed to work with Burn tensors and the relayrl_types TensorData abstraction.
///
/// # Arguments
///
/// * `dict` - A reference to a HashMap with String keys and generic RelayRLData values.
///
/// # Returns
///
/// An Option containing a HashMap with String keys and RelayRLData values.
pub fn convert_generic_dict(
    dict: &HashMap<String, RelayRLData>,
) -> Option<HashMap<String, RelayRLData>> {
    Some(dict.clone())
}

/// Validates a Burn model by checking that it can perform a forward pass with dummy tensors.
///
/// This function creates dummy tensors whose shapes match the metadata, runs a forward pass, and
/// verifies that the produced action tensor matches the expected shape. Supports input and output
/// ranks from 1 to 9 (independently).
pub fn validate_module<B: Backend + BackendMatcher<Backend = B> + 'static>(
    module: &ModelModule<B>,
) -> Result<(), ModelError> {
    let device = module.resolve_device();

    let input_shape = &module.metadata.input_shape;
    let output_shape = &module.metadata.output_shape;

    if !(1..=9).contains(&input_shape.len()) || !(1..=9).contains(&output_shape.len()) {
        return Err(ModelError::UnsupportedRank(format!(
            "Unsupported ranks: input {} output {}",
            input_shape.len(),
            output_shape.len()
        )));
    }

    match input_shape.len() {
        1 => validate_with_input::<B, 1>(module, &device, input_shape, output_shape),
        2 => validate_with_input::<B, 2>(module, &device, input_shape, output_shape),
        3 => validate_with_input::<B, 3>(module, &device, input_shape, output_shape),
        4 => validate_with_input::<B, 4>(module, &device, input_shape, output_shape),
        5 => validate_with_input::<B, 5>(module, &device, input_shape, output_shape),
        6 => validate_with_input::<B, 6>(module, &device, input_shape, output_shape),
        7 => validate_with_input::<B, 7>(module, &device, input_shape, output_shape),
        8 => validate_with_input::<B, 8>(module, &device, input_shape, output_shape),
        9 => validate_with_input::<B, 9>(module, &device, input_shape, output_shape),
        _ => unreachable!(),
    }
}

fn validate_with_input<B: Backend + BackendMatcher<Backend = B> + 'static, const D_IN: usize>(
    module: &ModelModule<B>,
    device: &<B as Backend>::Device,
    input_shape: &[usize],
    output_shape: &[usize],
) -> Result<(), ModelError> {
    match output_shape.len() {
        1 => call_validate::<B, D_IN, 1>(module, device, input_shape, output_shape),
        2 => call_validate::<B, D_IN, 2>(module, device, input_shape, output_shape),
        3 => call_validate::<B, D_IN, 3>(module, device, input_shape, output_shape),
        4 => call_validate::<B, D_IN, 4>(module, device, input_shape, output_shape),
        5 => call_validate::<B, D_IN, 5>(module, device, input_shape, output_shape),
        6 => call_validate::<B, D_IN, 6>(module, device, input_shape, output_shape),
        7 => call_validate::<B, D_IN, 7>(module, device, input_shape, output_shape),
        8 => call_validate::<B, D_IN, 8>(module, device, input_shape, output_shape),
        9 => call_validate::<B, D_IN, 9>(module, device, input_shape, output_shape),
        _ => Err(ModelError::UnsupportedRank(format!(
            "Unsupported ranks: input {} output {}",
            input_shape.len(),
            output_shape.len()
        ))),
    }
}

fn call_validate<
    B: Backend + BackendMatcher<Backend = B> + 'static,
    const D_IN: usize,
    const D_OUT: usize,
>(
    module: &ModelModule<B>,
    device: &<B as Backend>::Device,
    input_shape: &[usize],
    output_shape: &[usize],
) -> Result<(), ModelError> {
    let input_array: [usize; D_IN] = slice_to_array::<D_IN>(input_shape)?;
    let output_array: [usize; D_OUT] = slice_to_array::<D_OUT>(output_shape)?;
    let input_shape = Shape::from(input_array);
    let output_shape = Shape::from(output_array);

    validate_model_shapes::<B, D_IN, D_OUT>(module, device, &input_shape, &output_shape)
}

fn slice_to_array<const N: usize>(shape: &[usize]) -> Result<[usize; N], ModelError> {
    shape.try_into().map_err(|_| {
        ModelError::InvalidMetadata(format!(
            "Expected dimension of length {N}, but got {}",
            shape.len()
        ))
    })
}

fn validate_model_shapes<
    B: Backend + BackendMatcher<Backend = B> + 'static,
    const D_IN: usize,
    const D_OUT: usize,
>(
    module: &ModelModule<B>,
    device: &<B as Backend>::Device,
    input_shape: &Shape,
    output_shape: &Shape,
) -> Result<(), ModelError> {
    let obs: Arc<AnyBurnTensor<B, D_IN>> = match &module.metadata.input_dtype {
        #[cfg(feature = "ndarray-backend")]
        DType::NdArray(nd) => match nd {
            NdArrayDType::F16 | NdArrayDType::F32 | NdArrayDType::F64 => {
                Arc::new(AnyBurnTensor::Float(FloatBurnTensor::empty(
                    input_shape,
                    &DType::NdArray(nd.clone()),
                    device,
                )))
            }
            NdArrayDType::I8 | NdArrayDType::I16 | NdArrayDType::I32 | NdArrayDType::I64 => {
                Arc::new(AnyBurnTensor::Int(IntBurnTensor::empty(
                    input_shape,
                    &DType::NdArray(nd.clone()),
                    device,
                )))
            }
            NdArrayDType::Bool => Arc::new(AnyBurnTensor::Bool(BoolBurnTensor::empty(
                input_shape,
                &DType::NdArray(nd.clone()),
                device,
            ))),
        },
        #[cfg(feature = "tch-backend")]
        DType::Tch(tch) => match tch {
            TchDType::F16 | TchDType::Bf16 | TchDType::F32 | TchDType::F64 => {
                Arc::new(AnyBurnTensor::Float(FloatBurnTensor::empty(
                    input_shape,
                    &DType::Tch(tch.clone()),
                    device,
                )))
            }
            TchDType::I8 | TchDType::I16 | TchDType::I32 | TchDType::I64 | TchDType::U8 => {
                Arc::new(AnyBurnTensor::Int(IntBurnTensor::empty(
                    input_shape,
                    &DType::Tch(tch.clone()),
                    device,
                )))
            }
            TchDType::Bool => Arc::new(AnyBurnTensor::Bool(BoolBurnTensor::empty(
                input_shape,
                &DType::Tch(tch.clone()),
                device,
            ))),
        },
    };

    let mask: Arc<AnyBurnTensor<B, D_OUT>> = match &module.metadata.output_dtype {
        #[cfg(feature = "ndarray-backend")]
        DType::NdArray(nd) => match nd {
            NdArrayDType::F16 | NdArrayDType::F32 | NdArrayDType::F64 => {
                Arc::new(AnyBurnTensor::Float(FloatBurnTensor::empty(
                    output_shape,
                    &DType::NdArray(nd.clone()),
                    device,
                )))
            }
            NdArrayDType::I8 | NdArrayDType::I16 | NdArrayDType::I32 | NdArrayDType::I64 => {
                Arc::new(AnyBurnTensor::Int(IntBurnTensor::empty(
                    output_shape,
                    &DType::NdArray(nd.clone()),
                    device,
                )))
            }
            NdArrayDType::Bool => Arc::new(AnyBurnTensor::Bool(BoolBurnTensor::empty(
                output_shape,
                &DType::NdArray(nd.clone()),
                device,
            ))),
        },
        #[cfg(feature = "tch-backend")]
        DType::Tch(tch) => match tch {
            TchDType::F16 | TchDType::Bf16 | TchDType::F32 | TchDType::F64 => {
                Arc::new(AnyBurnTensor::Float(FloatBurnTensor::empty(
                    output_shape,
                    &DType::Tch(tch.clone()),
                    device,
                )))
            }
            TchDType::I8 | TchDType::I16 | TchDType::I32 | TchDType::I64 | TchDType::U8 => {
                Arc::new(AnyBurnTensor::Int(IntBurnTensor::empty(
                    output_shape,
                    &DType::Tch(tch.clone()),
                    device,
                )))
            }
            TchDType::Bool => Arc::new(AnyBurnTensor::Bool(BoolBurnTensor::empty(
                output_shape,
                &DType::Tch(tch.clone()),
                device,
            ))),
        },
    };

    let (action_tensor, _, _) = module.step::<D_IN, D_OUT>(obs, Some(mask));

    let action_dims: &Vec<usize> = &action_tensor.shape;
    let output_dims: &Vec<usize> = &output_shape.dims;

    for (a, o) in action_dims.iter().zip(output_dims.iter()) {
        if *a != *o {
            return Err(ModelError::InvalidOutputDimension(format!(
                "Model output shape mismatch: expected {:?}, got {:?}",
                output_dims, action_dims
            )));
        }
    }
    Ok(())
}

/// Serializes a model (`ModelModule`) into a vector of bytes.
///
/// The model is saved to a temporary file and then read back into a byte vector.
///
/// # Arguments
///
/// * `model` - A reference to the [`ModelModule`] (model) to be serialized.
///
/// # Returns
///
/// A vector of bytes representing the serialized model.
pub fn serialize_model_module<B: Backend + BackendMatcher<Backend = B>>(
    model: &ModelModule<B>,
    dir: PathBuf,
) -> Vec<u8> {
    // Save the full model directory (metadata.json + model file) to a temp dir,
    // then bundle as [4-byte metadata_len][metadata_bytes][model_bytes] so
    // deserialize_model_module can reconstruct the directory.
    let temp_dir = tempfile::Builder::new()
        .prefix("_model_save_")
        .tempdir_in(dir)
        .expect("Failed to create temp dir");
    let temp_path = temp_dir.path().to_path_buf();

    ModelModule::<B>::save(model, &temp_path).expect("Failed to save model");

    let meta_bytes =
        std::fs::read(temp_path.join("metadata.json")).expect("Failed to read metadata.json");
    let model_path = temp_path.join(&model.metadata.model_file);
    let model_bytes = std::fs::read(&model_path).expect("Failed to read model file");

    // Bundle: [4LE: meta_len][meta_bytes][model_bytes]
    let meta_len = meta_bytes.len() as u32;
    let mut bundle = Vec::with_capacity(4 + meta_bytes.len() + model_bytes.len());
    bundle.extend_from_slice(&meta_len.to_le_bytes());
    bundle.extend_from_slice(&meta_bytes);
    bundle.extend_from_slice(&model_bytes);
    bundle
}

/// Deserializes a vector of bytes into a model (`ModelModule`).
///
/// The function writes the provided bytes to a temporary file, flushes it, and then loads
/// the model from that file.
///
/// # Arguments
///
/// * `model_bytes` - A vector of bytes containing the serialized model.
///
/// # Returns
///
/// A [`ModelModule`] representing the deserialized model.
pub fn deserialize_model_module<B: Backend + BackendMatcher<Backend = B>>(
    bundle: Vec<u8>,
    _device: DeviceType,
) -> Result<ModelModule<B>, ModelError> {
    // Unpack bundle: [4LE: meta_len][meta_bytes][model_bytes]
    if bundle.len() < 4 {
        return Err(ModelError::InvalidMetadata("bundle too short".into()));
    }
    let meta_len = u32::from_le_bytes(bundle[..4].try_into().unwrap()) as usize;
    if bundle.len() < 4 + meta_len {
        return Err(ModelError::InvalidMetadata("bundle truncated".into()));
    }
    let meta_bytes = &bundle[4..4 + meta_len];
    let model_bytes = &bundle[4 + meta_len..];

    // Write to a temp dir so load_from_path can find metadata.json + model file.
    let temp_dir = tempfile::TempDir::new().expect("Failed to create temp dir");
    let temp_path = temp_dir.path();

    // Parse metadata to get the model_file name.
    let meta_json: serde_json::Value = serde_json::from_slice(meta_bytes)
        .map_err(|e| ModelError::InvalidMetadata(e.to_string()))?;
    let model_file = meta_json
        .get("model_file")
        .and_then(|v| v.as_str())
        .unwrap_or("model.onnx");

    std::fs::write(temp_path.join("metadata.json"), meta_bytes)
        .expect("Failed to write metadata.json");
    std::fs::write(temp_path.join(model_file), model_bytes).expect("Failed to write model file");

    // Keep temp_dir alive until load_from_path completes.
    let result = ModelModule::<B>::load_from_path(temp_path)
        .map_err(|e| ModelError::InvalidMetadata(format!("{e}")))?;
    drop(temp_dir);
    Ok(result)
}

#[cfg(all(
    test,
    feature = "ndarray-backend",
    any(feature = "tch-model", feature = "onnx-model")
))]
mod unit_tests {
    use std::collections::HashMap;
    use std::marker::PhantomData;
    use std::sync::Arc;

    use burn_ndarray::NdArray;

    use super::{convert_generic_dict, validate_module};
    use crate::data::action::RelayRLData;
    use crate::data::tensor::{DType, DeviceType, NdArrayDType};
    use crate::model::{
        InferenceModel, Model, ModelError, ModelFileType, ModelMetadata, ModelModule,
    };

    fn stub_module(rank: usize) -> ModelModule<NdArray> {
        let dims = vec![2; rank];

        ModelModule {
            model: Model {
                file_type: ModelFileType::Onnx,
                raw_bytes: Arc::<[u8]>::from(Vec::<u8>::new()),
                inference: InferenceModel::Unsupported,
                _phantom: PhantomData,
            },
            metadata: ModelMetadata {
                model_file: "test.onnx".to_string(),
                model_type: ModelFileType::Onnx,
                input_dtype: DType::NdArray(NdArrayDType::F32),
                output_dtype: DType::NdArray(NdArrayDType::F32),
                input_shape: dims.clone(),
                output_shape: dims,
                default_device: Some(DeviceType::Cpu),
            },
        }
    }

    #[test]
    fn convert_generic_dict_clones_auxiliary_data() {
        let mut dict = HashMap::new();
        dict.insert("reward".to_string(), RelayRLData::F32(1.25));

        let cloned = convert_generic_dict(&dict).expect("the helper should always return data");
        dict.insert("done".to_string(), RelayRLData::Bool(true));

        assert_eq!(cloned.len(), 1);
        assert!(matches!(
            cloned.get("reward"),
            Some(RelayRLData::F32(value)) if (*value - 1.25).abs() < f32::EPSILON
        ));
        assert!(!cloned.contains_key("done"));
    }

    #[test]
    fn validate_module_accepts_rank_one() {
        let module = stub_module(1);

        assert!(validate_module::<NdArray>(&module).is_ok());
    }

    #[test]
    fn validate_module_accepts_rank_nine() {
        let module = stub_module(9);

        assert!(validate_module::<NdArray>(&module).is_ok());
    }

    #[test]
    fn validate_module_rejects_rank_zero() {
        let module = stub_module(0);

        let err = validate_module::<NdArray>(&module)
            .expect_err("rank 0 should remain outside the supported range");

        assert!(matches!(
            err,
            ModelError::UnsupportedRank(message) if message.contains("input 0 output 0")
        ));
    }

    #[test]
    fn validate_module_rejects_rank_ten() {
        let module = stub_module(10);

        let err = validate_module::<NdArray>(&module)
            .expect_err("rank 10 should remain outside the supported range");

        assert!(matches!(
            err,
            ModelError::UnsupportedRank(message) if message.contains("input 10 output 10")
        ));
    }
}