use std::path::Path;
use onnx_extractor::{DataType, Model, Tensor};
use super::error;
use super::types::{ElementType, ModelIo, TensorInfo};
pub(super) fn load(path: &Path) -> syn::Result<ModelIo> {
let model = Model::load_from_file(path)
.map_err(|parse_error| error(path, format!("failed to parse ONNX model: {parse_error}")))?;
let graph = model.graph();
Ok(ModelIo {
input: parse_tensor(&graph.tensors()[&graph.inputs()[0]], path, "input")?,
output: parse_tensor(&graph.tensors()[&graph.outputs()[0]], path, "output")?,
})
}
fn parse_tensor(tensor: &Tensor, path: &Path, label: &str) -> syn::Result<TensorInfo> {
let element_type = element_type(tensor.data_type()).ok_or_else(|| {
error(
path,
format!(
"unsupported ONNX graph {label} tensor element type '{:?}' for '{}'",
tensor.data_type(),
tensor.name()
),
)
})?;
let dimensions = tensor.shape();
if dimensions.len() > 4 {
return Err(error(
path,
format!(
"ONNX graph {label} tensor '{}' rank {} exceeds Tensor's four dimensions",
tensor.name(),
dimensions.len()
),
));
}
let mut shape = [1; 4];
let offset = 4 - dimensions.len();
for (index, dimension) in dimensions.iter().enumerate() {
shape[offset + index] = usize::try_from(*dimension).map_err(|_| {
error(
path,
format!(
"ONNX graph {label} tensor '{}' has a dynamic dimension",
tensor.name()
),
)
})?;
}
Ok(TensorInfo {
element_type,
shape,
})
}
fn element_type(value: DataType) -> Option<ElementType> {
match value {
DataType::Int8 => Some(ElementType::I8),
DataType::Int16 => Some(ElementType::I16),
DataType::Int32 => Some(ElementType::I32),
DataType::Int64 => Some(ElementType::I64),
DataType::Uint8 => Some(ElementType::U8),
DataType::Uint16 => Some(ElementType::U16),
DataType::Uint32 => Some(ElementType::U32),
DataType::Uint64 => Some(ElementType::U64),
DataType::Float => Some(ElementType::F32),
DataType::Double => Some(ElementType::F64),
_ => None,
}
}