Struct Var

pub struct Var { /* private fields */ }

Var can be thought as the value it holds plus a link to the computation graph. Majority of operators are methods on Var.

Implementations

impl Var

pub fn new(input: &[f64], dim: &[usize]) -> Var

pub fn new_f64(input: &[f64], dim: &[usize]) -> Var

pub fn new_f32(input: &[f32], dim: &[usize]) -> Var

pub fn ref_copy(self: &Var) -> Var

Where it needs a assign operator, we should use this ref_copy. If a hard copy is necessary, then call clone().

pub fn set(&self, o: &Var)

With a &Var, use set to copy a value.

pub fn size(&self) -> Vec<usize>

pub fn numel(&self) -> usize

pub fn get_f32(&self, o: &[usize]) -> Result<f32, AutoDiffError>

pub fn set_f32(&self, o: &[usize], v: f32) -> Result<(), AutoDiffError>

pub fn get_f64(&self, o: &[usize]) -> Result<f64, AutoDiffError>

pub fn set_f64(&self, o: &[usize], v: f64) -> Result<(), AutoDiffError>

pub fn fill(size: &[usize], fill_value: &Var) -> Var

pub fn fill_f32(size: &[usize], fill_value: f32) -> Var

pub fn fill_f64(size: &[usize], fill_value: f64) -> Var

pub fn zeros(dim: &[usize]) -> Var

pub fn ones(dim: &[usize]) -> Var

pub fn twos(dim: &[usize]) -> Var

pub fn eye(n: usize, m: usize) -> Var

Identity matrix

pub fn empty(dim: &[usize]) -> Var

pub fn from_record_f32(&self, row: usize, record: &[f32])

Fill row by row.

pub fn from_record_f64(&self, row: usize, record: &[f64])

pub fn rand_usize( rng: &mut StdRng, dim: &[usize], left: usize, right: usize, ) -> Var

pub fn normal_f64(rng: &mut StdRng, dim: &[usize], mean: f64, std: f64) -> Var

pub fn normal_f32(rng: &mut StdRng, dim: &[usize], mean: f32, std: f32) -> Var

pub fn normal(rng: &mut StdRng, dim: &[usize], mean: f64, std: f64) -> Var

pub fn uniform_f64(rng: &mut StdRng, dim: &[usize], from: f64, to: f64) -> Var

pub fn uniform_f32(rng: &mut StdRng, dim: &[usize], from: f32, to: f32) -> Var

pub fn uniform(rng: &mut StdRng, dim: &[usize], from: f64, to: f64) -> Var

pub fn _add(&self, other: &Var) -> Var

pub fn _sub(&self, other: &Var) -> Var

pub fn _mul(&self, other: &Var) -> Var

pub fn _div(&self, other: &Var) -> Var

pub fn matmul(&self, other: &Var) -> Result<Var, AutoDiffError>

Matrix/inner/dot product

use auto_diff::{Var, var_f64, AutoDiffError};

extern crate openblas_src;

fn test_matmul() -> Result<(), AutoDiffError> {

let v1 = var_f64!([[1., 2., 3.], [4., 5., 6.]]); let v2 = var_f64!([[11., 12., 13.], [14., 15., 16.], [17., 18., 19.]]); let v3 = v1.matmul(&v2)?; let em = var_f64!([[90.0, 96.0, 102.0], [216.0, 231.0, 246.0]]); assert_eq!(v3, em);

Ok(())

}

test_matmul();

pub fn outer(&self, other: &Var) -> Result<Var, AutoDiffError>

Outer product

let v1 = Var::new_f64(&[1., 2., 3.], &[3]);
let v2 = Var::new_f64(&[4., 5., 6.], &[3]);
let v3 = v1.outer(&v2)?;
let em = var_f64!([[4.,   5.,  6.],
                   [8.,  10., 12.],
                   [12., 15., 18.]]);
assert_eq!(v3, em);

pub fn elu(&self, alpha: Var) -> Result<Var, AutoDiffError>

pub fn relu(&self) -> Result<Var, AutoDiffError>

pub fn sigmoid(&self) -> Result<Var, AutoDiffError>

pub fn mse_loss(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn bce_with_logits_loss(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn cross_entropy_loss(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn abs(&self) -> Result<Var, AutoDiffError>

pub fn acos(&self) -> Result<Var, AutoDiffError>

pub fn asin(&self) -> Result<Var, AutoDiffError>

pub fn atan(&self) -> Result<Var, AutoDiffError>

pub fn ceil(&self) -> Result<Var, AutoDiffError>

pub fn cos(&self) -> Result<Var, AutoDiffError>

pub fn cosh(&self) -> Result<Var, AutoDiffError>

pub fn exp(&self) -> Result<Var, AutoDiffError>

pub fn expm1(&self) -> Result<Var, AutoDiffError>

pub fn floor(&self) -> Result<Var, AutoDiffError>

pub fn frac(&self) -> Result<Var, AutoDiffError>

pub fn log(&self) -> Result<Var, AutoDiffError>

pub fn log10(&self) -> Result<Var, AutoDiffError>

pub fn log1p(&self) -> Result<Var, AutoDiffError>

pub fn log1pexp(&self) -> Result<Var, AutoDiffError>

pub fn log2(&self) -> Result<Var, AutoDiffError>

pub fn neg(&self) -> Result<Var, AutoDiffError>

pub fn _neg(&self) -> Var

pub fn reciprocal(&self) -> Result<Var, AutoDiffError>

pub fn round(&self) -> Result<Var, AutoDiffError>

pub fn rsqrt(&self) -> Result<Var, AutoDiffError>

pub fn sign(&self) -> Result<Var, AutoDiffError>

pub fn sin(&self) -> Result<Var, AutoDiffError>

pub fn sinh(&self) -> Result<Var, AutoDiffError>

pub fn sqrt(&self) -> Result<Var, AutoDiffError>

pub fn tan(&self) -> Result<Var, AutoDiffError>

pub fn tanh(&self) -> Result<Var, AutoDiffError>

pub fn trunc(&self) -> Result<Var, AutoDiffError>

pub fn max_pair(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn min_pair(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn arg_sort( &self, dim: usize, descending: bool, ) -> Result<Var, AutoDiffError>

pub fn eq_elem(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn equal(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn ge(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn gt(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn le(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn lt(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn ne(&self, other: &Var) -> Result<Var, AutoDiffError>

pub fn cat(&self, other: &[Var], dim: usize) -> Result<Var, AutoDiffError>

Concatenates the given sequence of seq tensors in the given dimension. The input tensor should all have the same size except on the given dimension. The output tensor will have all the same size as the input except the given dimension, which will be the sum of the inputs on the given dimension. Apply cat on [tensor(5, 3, 2), tensor(5, 7, 2), ] will get a tensor(5, 10, 2).

use auto_diff::{Var, var_f64, AutoDiffError};

extern crate openblas_src;

fn test_cat() -> Result<(), AutoDiffError> {

let m1 = Var::empty(&[3, 1]); let m2 = Var::empty(&[3, 1]); let m3 = Var::empty(&[3, 1]); let m4 = m1.cat(&[m2, m3], 1)?; assert_eq!(m4.size(), [3, 3]);

Ok(())

}

test_cat();

pub fn chunk( &self, chunks: usize, dim: usize, ) -> Result<Vec<Var>, AutoDiffError>

pub fn conditional_select(&self, x: &Var, y: &Var) -> Result<Var, AutoDiffError>

pub fn gather(&self, dim: usize, index: Var) -> Result<Var, AutoDiffError>

pub fn index_select(&self, dim: usize, index: Var) -> Result<Var, AutoDiffError>

pub fn index_exclude( &self, dim: usize, index: Var, ) -> Result<Var, AutoDiffError>

pub fn permute(&self, dim: &[usize]) -> Result<Var, AutoDiffError>

pub fn repeat(&self, dim: &[usize]) -> Result<Var, AutoDiffError>

pub fn reshape(&self, new_shape: &[usize]) -> Result<Var, AutoDiffError>

pub fn split( &self, sections: &[usize], dim: usize, ) -> Result<Vec<Var>, AutoDiffError>

pub fn squeeze(&self, dim: Option<usize>) -> Result<Var, AutoDiffError>

pub fn t(&self) -> Result<Var, AutoDiffError>

pub fn take(&self, index: &[usize]) -> Result<Var, AutoDiffError>

pub fn unsqueeze(&self, dim: usize) -> Result<Var, AutoDiffError>

pub fn stack(&self, other: &[Var], dim: usize) -> Result<Var, AutoDiffError>

Stack tensor with the same size along a new dimension specified by dim. The difference from cat is that cat don’t create new dimension.

let m1 = var_f64!([[1., 2., ],
               [3., 4., ]]);
let m2 = var_f64!([[5., 6., ],
               [7., 8., ]]);
let m3 = m1.stack(&[m2], 1)?;

pub fn det(&self) -> Result<Var, AutoDiffError>

pub fn inv(&self) -> Result<Var, AutoDiffError>

pub fn normalize_unit(&self) -> Result<Var, AutoDiffError>

pub fn tr(&self) -> Result<Var, AutoDiffError>

pub fn argmax( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn argmin( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn logsumexp( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn mean( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn prod( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn std( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn sum( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn var( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn max( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn min( &self, dim: Option<&[usize]>, keepdim: bool, ) -> Result<Var, AutoDiffError>

pub fn get_patch( &self, range: &[(usize, usize)], step: Option<&[usize]>, ) -> Result<Var, AutoDiffError>

Get a portion of the tensor and return it.

let m1 = var_f64!([[1., 2., 3.],
                   [4., 5., 6.],
                   [7., 8., 9.]]);
let m2 = var_f64!([[4., 5.],
                   [7., 8.]]);
assert_eq!(m1.get_patch(&[(1, 3), (0, 2)], None)?, m2);

pub fn set_patch( &self, other: &Var, range: &[(usize, usize)], step: Option<&[usize]>, ) -> Result<Var, AutoDiffError>

Set a portion of the tensor.

let m1 = var_f64!([[1., 2., 3.],
                   [4., 5., 6.],
                   [7., 8., 9.]]);
let m2 = var_f64!([[10., 11.],
                   [12., 13.]]);
let m3 = var_f64!([[1.,   2., 3.],
                   [10., 11., 6.],
                   [12., 13., 9.]]);
assert_eq!(m1.set_patch(&m2, &[(1, 3), (0, 2)], None)?, m3);

pub fn view(&self, new_shape: &[usize]) -> Result<Var, AutoDiffError>

pub fn val(&self) -> Tensor

pub fn set_grad(&self, use_gradient: bool)

Use gradient or not, default is to use.

pub fn reset_net(&self)

Reset net in the background.

pub fn grad(&self) -> Result<Var, AutoDiffError>

The current gradient for the Var.

pub fn bp(&self) -> Result<(), AutoDiffError>

Apply back propagation to get numerical gradient.

pub fn step(&self, opt: &mut dyn Optimizer) -> Result<(), AutoDiffError>

pub fn rerun(&self) -> Result<(), AutoDiffError>

Run the computation graph again.

pub fn get_io_var(&self) -> Result<(Vec<Var>, Vec<Var>), AutoDiffError>

Extract input and output from the hidden net.

pub fn get_var_by_label(&self, label: &str) -> Result<Var, AutoDiffError>

Get var by string label

pub fn set_label(&self, label: &str) -> Result<(), AutoDiffError>

pub fn set_predict(&self) -> Result<(), AutoDiffError>

pub fn predict(&self) -> Result<Var, AutoDiffError>

pub fn dump_net(&self) -> Rc<RefCell<Net>>

For debug.

Trait Implementations

impl Add for Var

type Output = Var

The resulting type after applying the + operator.

fn add(self, other: Self) -> Self

Performs the + operation.

impl Clone for Var

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Var

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Var

fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>

where D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Display for Var

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Div for Var

type Output = Var

The resulting type after applying the / operator.

fn div(self, other: Self) -> Self

Performs the / operation.

impl Mul for Var

type Output = Var

The resulting type after applying the * operator.

fn mul(self, other: Self) -> Self

Performs the * operation.

impl Neg for Var

type Output = Var

The resulting type after applying the - operator.

fn neg(self) -> Self

Performs the unary - operation.

impl PartialEq for Var

fn eq(&self, other: &Self) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Serialize for Var

fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>

where S: Serializer,

Serialize this value into the given Serde serializer.

impl Sub for Var

type Output = Var

The resulting type after applying the - operator.

fn sub(self, other: Self) -> Self

Performs the - operation.

impl TryFrom<Var> for Vec<usize>

type Error = AutoDiffError

The type returned in the event of a conversion error.

fn try_from(value: Var) -> Result<Self, Self::Error>

Performs the conversion.

impl TryFrom<Var> for f32

type Error = AutoDiffError

The type returned in the event of a conversion error.

fn try_from(value: Var) -> Result<Self, Self::Error>

Performs the conversion.

impl TryFrom<Var> for f64

type Error = AutoDiffError

The type returned in the event of a conversion error.

fn try_from(value: Var) -> Result<Self, Self::Error>

Performs the conversion.

impl Eq for Var