Struct microtensor::Graph

pub struct Graph<T: Real + 'static> {
    pub inputs: Vec<Variable<T>>,
    pub outputs: Vec<Variable<T>>,
}

Snapshot of a computation graph with defined inputs and outputs.

Can be used for recomputing the entire graph or saving it to disc.

Fields

inputs: Vec<Variable<T>>

outputs: Vec<Variable<T>>

Implementations

impl<T: Real + Serialize + DeserializeOwned + 'static> Graph<T>

pub fn new(inputs: &[Variable<T>], outputs: &[Variable<T>]) -> Self

Examples found in repository

examples/graph.rs (line 31)

fn build_model(filename: &str) {
  // Have some inputs
  let x1 = Tensor::vec(&[1.0, 2.0]).tracked();
  let x2 = Tensor::ones(&[16]).tracked();
  let w = Tensor::randn(&[2, 16]).trained();

  // Do some computations
  let y = x1.mm(&w);
  let z = (&y * &x2).sum(0);

  // Pack the resulting graph into a Graph structure to make its inputs
  // and outputs explicit and arrange them in an order of your liking.
  let graph = Graph::new(&[x1, x2], &[y, z]);

  // Save entire computation graph to disc
  graph.save(filename).unwrap();
}

pub fn run(&self, inputs: &[&Variable<T>])

Examples found in repository

examples/graph.rs (lines 43-46)

fn load_model(filename: &str) {
  let graph = Graph::load(filename).unwrap();

  // Feed new data using #run.
  // Updating the entire graph in this way is more efficient
  // than calling #forward on each individual output.
  graph.run(&[
    &Tensor::vec(&[5.0, 6.0]).tracked(),
    &Tensor::randn(&[16]).tracked(),
  ]);

  // Get new output..
  let z = &graph.outputs[1];
  println!("z is now {}", z.item());

  // ..or train the model further
  let z = &graph.outputs[1];
  z.backward();
  for mut param in z.parameters() {
    param -= param.grad().unwrap() * 0.01
  }
}

pub fn parameters(&self) -> Vec<Variable<T>>

pub fn load(filename: &str) -> Result<Self>

Examples found in repository

examples/graph.rs (line 38)

fn load_model(filename: &str) {
  let graph = Graph::load(filename).unwrap();

  // Feed new data using #run.
  // Updating the entire graph in this way is more efficient
  // than calling #forward on each individual output.
  graph.run(&[
    &Tensor::vec(&[5.0, 6.0]).tracked(),
    &Tensor::randn(&[16]).tracked(),
  ]);

  // Get new output..
  let z = &graph.outputs[1];
  println!("z is now {}", z.item());

  // ..or train the model further
  let z = &graph.outputs[1];
  z.backward();
  for mut param in z.parameters() {
    param -= param.grad().unwrap() * 0.01
  }
}

pub fn save(&self, filename: &str) -> Result<()>

Examples found in repository

examples/graph.rs (line 34)

fn build_model(filename: &str) {
  // Have some inputs
  let x1 = Tensor::vec(&[1.0, 2.0]).tracked();
  let x2 = Tensor::ones(&[16]).tracked();
  let w = Tensor::randn(&[2, 16]).trained();

  // Do some computations
  let y = x1.mm(&w);
  let z = (&y * &x2).sum(0);

  // Pack the resulting graph into a Graph structure to make its inputs
  // and outputs explicit and arrange them in an order of your liking.
  let graph = Graph::new(&[x1, x2], &[y, z]);

  // Save entire computation graph to disc
  graph.save(filename).unwrap();
}

Trait Implementations

impl<T: Clone + Real + 'static> Clone for Graph<T>

fn clone(&self) -> Graph<T>

Returns a copy of the value.

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

Performs copy-assignment from source.

impl<T: Debug + Real + 'static> Debug for Graph<T>

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

Formats the value using the given formatter.