Crate metatensor

Metatensor

Metatensor is a library providing a specialized data storage format for atomistic machine learning (think numpy or torch.Tensor, but also carrying extra metadata for atomistic systems).

The core of the format is implemented in metatensor-core, and exported as a C API. This C API is then re-exported to Rust in this crate. By default, metatensor-core is distributed as a shared library that you’ll need to install separately on end user machines.

Features

You can enable the static feature in Cargo.toml to use a static build of the C API, removing the need to carry around the metatensor-core shared library.

[dependencies]
metatensor = {version = "...", features = ["static"]}

Re-exports

pub use metatensor_sys as c_api;

Modules

errors

io

Input/Output facilities for storing crate::TensorMap and crate::Labels on disk

Structs

ArrayRef

Reference to a data array in metatensor-core

ArrayRefMut

Mutable reference to a data array in metatensor-core

EmptyArray

An implementation of the Array trait without any data.

Error

Error type used in metatensor

GradientsIter

Iterator over parameter/TensorBlockRef pairs for all gradients in a TensorBlockRef

GradientsMutIter

Iterator over parameter/TensorBlockRefMut pairs for all gradients in a TensorBlockRefMut

LabelValue

A single value inside a label.

Labels

A set of labels used to carry metadata associated with a tensor map.

LabelsBuilder

Builder for Labels

LabelsFixedSizeIter

Iterator over entries in a set of Labels as fixed size arrays

LabelsIter

iterator over Labels entries

LazyMetadata

Lazily accessed metadata inside a TensorBlock

TensorBlock

A single block, containing both values & optionally gradients of these values w.r.t. any relevant quantity.

TensorBlockData

All the basic data in a TensorBlockRef as a struct with separate fields.

TensorBlockDataMut

All the basic data in a TensorBlockRefMut as a struct with separate fields.

TensorBlockRef

Reference to a TensorBlock

TensorBlockRefMut

Mutable reference to a TensorBlock

TensorMap

TensorMap is the main user-facing struct of this library, and can store any kind of data used in atomistic machine learning.

TensorMapIter

Iterator over key/block pairs in a TensorMap

TensorMapIterMut

Iterator over key/block pairs in a TensorMap, with mutable access to the blocks

Traits

Array

The Array trait is used by metatensor to manage different kind of data array with a single API. Metatensor only knows about Box<dyn Array>, and manipulate the data through the functions on this trait.