Save

To save a list of Tensors

Create saver

let saver = TensorSaver::new("path/to/save/file");

Push Tensors to the saver. Note: Tensors can be different type. If the Tensor is cuda, you must convert it to cpu first.

let a = Tensor::<f32>::randn([1, 100, 100, 100])?;
// push method will move the saver and will return saver back.
let saver = saver.push(
    /*name of the tensor to save*/ "a",
    /*tensor to save*/ a,
    hpt::save_load::CompressionAlgo::Gzip,
    /*compression level*/ 9,
);

Save

saver.save()?;

Final code will be

use hpt::{ops::Random, save_load::TensorSaver, Tensor};
fn main() -> anyhow::Result<()> {
    let a = Tensor::<f32>::randn([1, 100, 100, 100])?;
    let saver = TensorSaver::new("path/to/save/file");
    let saver = saver.push(
        "a",
        a,
        hpt::save_load::CompressionAlgo::Gzip,
        9,
    );
    saver.save()?;
    Ok(())
}

Load

To load a list of Tensors, you must make sure the Tensors you load are in same type. You can only load data into CPU.

Create a loader

let loader = TensorLoader::new("path/to/load/file");

Load Tensors by passing the tensor name and slice.

let loaded: std::collections::HashMap<String, Tensor<f32>> = loader
    .push("a", &select![:, 1:-1:2, 3::, :]) // you can use python slice syntax
    .push("b", &[]) // if you don't slice, simply use empty slice
    .load::<Tensor<f32>>()?;

If your file contains different type of Tensors, load it multiple times.

Save Struct

If you have a struct that contains a Tensor and you want to save a whole struct.

Use Save and Load Derive macro.

use hpt::{Load, Save};

Derive the Save and Load for your struct.

#[derive(Save, Load)]
pub struct Linear {
    weight: Tensor<f32>,
    bias: Tensor<f32>,
}

If you want to save a Tensor with specific compression

#[derive(Save, Load)]
pub struct Linear {
    #[compress(algo = "gzip", level = "9")]
    weight: Tensor<f32>,
    #[compress(algo = "zlib", level = "5")]
    bias: Tensor<f32>,
}

Save the struct

impl Linear {
    pub fn new(in_features: i64, out_features: i64) -> Result<Self, TensorError> {
        Ok(Self {
            weight: Tensor::<f32>::randn([in_features, out_features])?,
            bias: Tensor::<f32>::zeros([out_features])?,
        })
    }
}
let linear = Linear::new(1000, 1000)?;
linear.save("path/to/save/linear")?;

Load Struct

If you saved a struct in a file and you want to load the struct

Make sure your struct Derived Load.
Load the file

let linear = Linear::load("path/to/save/linear")?;

Backend Support

Backend	Supported
CPU	✅
Cuda	✅