Known issues

Tensor construction helpers

Std.Tensor.Construct exports stack, squeeze, and unsqueeze, but a call with an ordinary concrete tensor fails chelis check. Their signatures place the axis between two rank spreads, tensor[..pre, ..post, p], and the checker cannot decide where pre ends and post begins in a concrete shape:

import Std.Tensor.Construct (unsqueeze)
y = unsqueeze(to_tensor([1.0, 2.0], f32), 0i32)

chelis check reports a type error on argument 1 of unsqueeze, saying that two adjacent rank spreads cannot be split against a concrete shape.

The built-in shape operations do the same jobs with an explicit axis or target shape:

module Recipes
row = insert(to_tensor([1.0, 2.0, 3.0], f32), 0, 1i64)
flat = reshape(to_tensor([[1.0, 2.0, 3.0]], f32), [3i64])
stacked = concat([insert(to_tensor([1.0, 2.0], f32), 0, 1i64), insert(to_tensor([3.0, 4.0], f32), 0, 1i64)], 0)

row = tensor(shape=[1, 3], data=[1.0, 2.0, 3.0])
flat = tensor(shape=[3], data=[1.0, 2.0, 3.0])
stacked = tensor(shape=[2, 2], data=[1.0, 2.0, 3.0, 4.0])

• In place of unsqueeze(x, axis), use insert(x, axis, 1i64), which adds an axis of extent one at a constant position.
• In place of squeeze(x, axis), use reshape(x, [...]) with the target extents written out as i64 values.
• In place of stack(xs, axis), insert the new axis into each tensor and concat along it. map does this for a List of any length, including one built at run time:

def stack_rows[n](xs: List[tensor[n, f32]]) -> tensor[*, n, f32] = concat(map(fn (x) -> insert(x, 0i32, 1i64), xs), 0i32)
rows = [to_tensor([1.0, 2.0], f32), to_tensor([3.0, 4.0], f32), to_tensor([5.0, 6.0], f32)]
stacked = stack_rows(rows)

rows = [tensor(shape=[2], data=[1.0, 2.0]), tensor(shape=[2], data=[3.0, 4.0]), tensor(shape=[2], data=[5.0, 6.0])]
stacked = tensor(shape=[3, 2], data=[1.0, 2.0, 3.0, 4.0, 5.0, 6.0])

The result's first extent is * because it depends on the list's length. The same program builds with chelis build and prints the same output.

Module-level dim declarations

A file with a top-level dim declaration passes chelis check, but chelis eval and chelis build stop with a lowered root count mismatch error. A named axis needs no declaration: write it in the signature, as in def reduce_seq[pre, post](x: &tensor[..pre, seq, ..post, f32]), and the checker treats seq as a named dimension.