Struct candle_nn::var_builder::VarBuilderArgs

pub struct VarBuilderArgs<'a, B: Backend> { /* private fields */ }

A structure used to retrieve variables, these variables can either come from storage or be generated via some form of initialization.

The way to retrieve variables is defined in the backend embedded in the VarBuilder.

Implementations

impl<'a, B: Backend> VarBuilderArgs<'a, B>

pub fn new_with_args(backend: B, dtype: DType, dev: &Device) -> Self

pub fn prefix(&self) -> String

Returns the prefix of the VarBuilder.

pub fn root(&self) -> Self

Returns a new VarBuilder using the root path.

pub fn set_prefix(&self, prefix: impl ToString) -> Self

Returns a new VarBuilder with the prefix set to prefix.

pub fn push_prefix<S: ToString>(&self, s: S) -> Self

Return a new VarBuilder adding s to the current prefix. This can be think of as cd into a directory.

pub fn pp<S: ToString>(&self, s: S) -> Self

Short alias for push_prefix.

pub fn device(&self) -> &Device

The device used by default.

pub fn dtype(&self) -> DType

The dtype used by default.

pub fn contains_tensor(&self, tensor_name: &str) -> bool

This returns true only if a tensor with the passed in name is available. E.g. when passed a, true is returned if prefix.a exists but false is returned if only prefix.a.b exists.

pub fn get_with_hints<S: Into<Shape>>( &self, s: S, name: &str, hints: B::Hints ) -> Result<Tensor>

Retrieve the tensor associated with the given name at the current path.

pub fn get<S: Into<Shape>>(&self, s: S, name: &str) -> Result<Tensor>

Retrieve the tensor associated with the given name at the current path.

pub fn get_with_hints_dtype<S: Into<Shape>>( &self, s: S, name: &str, hints: B::Hints, dtype: DType ) -> Result<Tensor>

Retrieve the tensor associated with the given name & dtype at the current path.

impl<'a> VarBuilderArgs<'a, Box<dyn SimpleBackend + 'a>>

pub fn from_backend( backend: Box<dyn SimpleBackend + 'a>, dtype: DType, device: Device ) -> Self

Initializes a VarBuilder using a custom backend.

It is preferred to use one of the more specific constructors. This constructor is provided to allow downstream users to define their own backends.

pub fn zeros(dtype: DType, dev: &Device) -> Self

Initializes a VarBuilder that uses zeros for any tensor.

pub fn from_tensors( ts: HashMap<String, Tensor>, dtype: DType, dev: &Device ) -> Self

Initializes a VarBuilder that retrieves tensors stored in a hashtable. An error is returned if no tensor is available under the requested path or on shape mismatches.

pub fn from_varmap(varmap: &VarMap, dtype: DType, dev: &Device) -> Self

Initializes a VarBuilder using a VarMap. The requested tensors are created and initialized on new paths, the same tensor is used if the same path is requested multiple times. This is commonly used when initializing a model before training.

Note that it is possible to load the tensor values after model creation using the load method on varmap, this can be used to start model training from an existing checkpoint.

pub unsafe fn from_mmaped_safetensors<P: AsRef<Path>>( paths: &[P], dtype: DType, dev: &Device ) -> Result<Self>

Initializes a VarBuilder that retrieves tensors stored in a collection of safetensors files.

Safety

The unsafe is inherited from [memmap2::MmapOptions].

pub fn from_buffered_safetensors( data: Vec<u8>, dtype: DType, dev: &Device ) -> Result<Self>

Initializes a VarBuilder from a binary builder in the safetensor format.

pub fn from_npz<P: AsRef<Path>>( p: P, dtype: DType, dev: &Device ) -> Result<Self>

Initializes a VarBuilder that retrieves tensors stored in a numpy npz file.

pub fn from_pth<P: AsRef<Path>>( p: P, dtype: DType, dev: &Device ) -> Result<Self>

Initializes a VarBuilder that retrieves tensors stored in a pytorch pth file.

pub fn rename_f<F: Fn(&str) -> String + Sync + Send + 'static>( self, f: F ) -> Self

Gets a VarBuilder that applies some renaming function on tensor it gets queried for before passing the new names to the inner VarBuilder.

use candle::{Tensor, DType, Device};

let a = Tensor::arange(0f32, 6f32, &Device::Cpu)?.reshape((2, 3))?;
let tensors: std::collections::HashMap<_, _> = [
    ("foo".to_string(), a),
]
.into_iter()
.collect();
let vb = candle_nn::VarBuilder::from_tensors(tensors, DType::F32, &Device::Cpu);
assert!(vb.contains_tensor("foo"));
assert!(vb.get((2, 3), "foo").is_ok());
assert!(!vb.contains_tensor("bar"));
let vb = vb.rename_f(|f: &str| if f == "bar" { "foo".to_string() } else { f.to_string() });
assert!(vb.contains_tensor("bar"));
assert!(vb.contains_tensor("foo"));
assert!(vb.get((2, 3), "bar").is_ok());
assert!(vb.get((2, 3), "foo").is_ok());
assert!(!vb.contains_tensor("baz"));

pub fn rename<R: Renamer + Send + Sync + 'a>(self, renamer: R) -> Self

Trait Implementations

impl<'a, B: Backend> Clone for VarBuilderArgs<'a, B>

fn clone(&self) -> Self

Returns a copy of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.