Struct Tape 

pub struct Tape<'d> { /* private fields */ }

Autograd tape. Records forward operations, runs backward to compute gradients.

Implementations

impl<'d> Tape<'d>

pub fn new(dev: &'d GpuDevice) -> Self

pub fn leaf(&mut self, data: &[f32]) -> TensorId

Register a leaf tensor (parameter or input data). No backward through this.

pub fn read(&self, id: TensorId) -> Result<Vec<f32>>

Read tensor data back to CPU.

pub fn read_grad(&self, id: TensorId) -> Result<Option<Vec<f32>>>

Read gradient data back to CPU. Returns None if no gradient computed.

pub fn add(&mut self, a: TensorId, b: TensorId) -> Result<TensorId>

pub fn sub(&mut self, a: TensorId, b: TensorId) -> Result<TensorId>

pub fn mul(&mut self, a: TensorId, b: TensorId) -> Result<TensorId>

pub fn scale(&mut self, a: TensorId, s: f32) -> Result<TensorId>

pub fn relu(&mut self, a: TensorId) -> Result<TensorId>

pub fn sigmoid(&mut self, a: TensorId) -> Result<TensorId>

pub fn swish(&mut self, a: TensorId) -> Result<TensorId>

pub fn tanh_act(&mut self, a: TensorId) -> Result<TensorId>

pub fn matmul( &mut self, a: TensorId, b: TensorId, m: u32, n: u32, k: u32, ) -> Result<TensorId>

pub fn mse_loss(&mut self, pred: TensorId, target: TensorId) -> Result<TensorId>

pub fn conv2d( &mut self, input: TensorId, weight: TensorId, bias: Option<TensorId>, batch: u32, in_c: u32, in_h: u32, in_w: u32, out_c: u32, kh: u32, kw: u32, stride: (u32, u32), padding: (u32, u32), dilation: (u32, u32), groups: u32, ) -> Result<TensorId>

pub fn backward(&mut self, loss: TensorId) -> Result<()>

Run backward pass from a loss tensor. Computes gradients for all tensors on the tape.