completed lstm

[mei1127]

@huningxin @fdwr @shiyi9801 @BruceDai PTAL, thanks!

[BruceDai]

LGTM, thanks @mei1127.

[BruceDai]

Now squeeze op has been removed, would you please help also remove it from this WebNN Baseline.
You can refer to the given squeeze method in Spec, thanks.

function squeeze(builder, op) {
  return builder.reshape(op, op.shape().remove(0));
}

[fdwr]

The definition is:

function squeeze(input, axes) {
  if (!axes) axes = [];
  if (!axes.length)
    input.shape().forEach((item, i) => { axes.push(i); });
  shape = Array.from(input.shape());
  for (let axis in axes.sort().reverse())
    if (axis < shape.length && shape[axis] == 1)
      shape.splice(axis, 1);
  return builder.reshape(input, shape);
}

[mei1127]

OK， I will revise it next week:)

[BruceDai]

Please delete this blank line.

[BruceDai]

Suggested change

	activations = [sigmoid, tanh, tanh]}={}) {
	activations = [sigmoid, tanh, tanh]} = {}) {

[fdwr]

The error message will be split :( :

The shape of cellState
  [2, 3] is invalid.

[fdwr]

Could be easier for future readers to visual parse if the indentation was aligned:

Suggested change

	const starts = layout === 'iofg' ? {i: 0, o: hiddenSize, f: 2 * hiddenSize, g: 3 *hiddenSize} :
	{i: 0, f: hiddenSize, g: 2 * hiddenSize, o: 3 * hiddenSize};
	const starts = (layout === 'iofg') ? {i: 0, o: hiddenSize, f: 2 * hiddenSize, g: 3 * hiddenSize} :
	{i: 0, f: hiddenSize, g: 2 * hiddenSize, o: 3 * hiddenSize};

[fdwr]

It's cool how easy it is to compose these from existing operators :).

[fdwr]

Looks good after some small comments. This is a complex operator! Thanks for adding it. 🙂

[fdwr]

Just minor comments. Functionally looks correct. TY Mei.

[fdwr]

Suggested change

	const inpAxes = axes ?? new Array(input.rank).fill(0).map((_, i) => i);
	const inputAxes = axes ?? new Array(input.rank).fill(0).map((_, i) => i);

Whole word identifiers are more readable for others later (and it's consistent with input.rank rather than inp.rank).

[fdwr]

Suggested change

	const inpAxes = axes ?? new Array(input.rank).fill(0).map((_, i) => i);
	const inputAxes = axes ?? new Array.from({length: input.rank}, (_, i) => i);

Hmm, kinda confusing way of initializing a sequence by chaining multiple methods together (took me a bit to figure out), rather than just saying Array.from({length: input.rank}, (_, i) => i), which is longer character-wise, but it's clearer intention-wise (and probably more performant than a fill which is overwritten by a map). This one surprising case where C++ is shorter :b std::ranges::iota(inputAxes, 0).