Stdlib

std.torch

libtorch bindings: tensors, neural-network modules, optimizers, and device management. Available only when Eta is built with -DETA_BUILD_TORCH=ON.

(import std.torch)

Tensor construction

Symbol	Description
`(tensor data shape)`	Build a tensor from nested data.
`(ones shape)`	Tensor of ones.
`(zeros shape)`	Tensor of zeros.
`(randn shape)`	Standard-normal samples.
`(arange start end [step])`	1-D range tensor.
`(linspace start end n)`	Linearly spaced 1-D tensor.
`(from-list xs)`	Build a tensor from a flat list.
`(manual-seed n)`	Set the global RNG seed.
`(tensor? x)`	True when `x` is a tensor.

Symbol	Description
`(matmul a b)`	Matrix multiplication.
`(dot a b)`	Dot product.
`(cholesky a)`	Cholesky factor.
`(matrix-exp a)`	Matrix exponential.
`(mvnormal mean cov n)`	Sample multivariate normals.
`(column-l2-norm m)`	L2 norm per column.

torch:column-l2-norm is exported as a long-name alias.

Symbol	Description
`(shape t)`	Shape as a list.
`(reshape t shape)`	Reshape (no copy when possible).
`(transpose t dim0 dim1)`	Swap two dimensions.
`(squeeze t [dim])`	Drop unit dimensions.
`(unsqueeze t dim)`	Insert a unit dimension.
`(cat tensors dim)`	Concatenate along `dim`.
`(fact-table->tensor table cols)`	Build a tensor from selected columns.

Symbol	Description
`(tsum t [dim])`	Sum (full or per dim).
`(mean t [dim])`	Mean.
`(tmax t [dim])`	Max.
`(tmin t [dim])`	Min.
`(argmax t [dim])`	Index of maximum.
`(argmin t [dim])`	Index of minimum.
`(item t)`	Scalar value of a 0-d tensor.
`(to-list t)`	Convert tensor to nested lists.
`(numel t)`	Number of elements.

Symbol	Description
`(requires-grad! t flag)`	Toggle gradient tracking.
`(requires-grad? t)`	Read the flag.
`(detach t)`	Detach from the autograd graph.
`(backward loss)`	Backpropagate from a scalar loss.
`(grad t)`	Gradient tensor accumulated on `t`.
`(zero-grad! t)`	Clear gradient.

Symbol	Description
`(linear in-features out-features)`	Linear layer.
`(sequential layers)`	Chain modules.
`(relu-layer)`	ReLU as a module.
`(sigmoid-layer)`	Sigmoid as a module.
`(dropout p)`	Dropout layer.
`(forward module input)`	Run the forward pass.
`(parameters module)`	List of trainable parameters.
`(train! module)`	Set training mode.
`(eval! module)`	Set eval mode.
`(module? x)`	True when `x` is a module.

Symbol	Description
`(mse-loss pred target)`	Mean-squared error.
`(l1-loss pred target)`	Mean absolute error.
`(cross-entropy-loss logits targets)`	Cross-entropy.

Symbol	Description
`(sgd params lr . opts)`	Build an SGD optimizer.
`(adam params lr . opts)`	Build an Adam optimizer.
`(step! optimizer)`	Apply one optimizer step.
`(optim-zero-grad! optimizer)`	Zero parameter gradients.
`(optimizer? x)`	True when `x` is an optimizer.
`(train-step! optimizer loss-thunk)`	Convenience: zero, forward+backward, step.