Context

This arose during prototyping additional models and during Chromium code review and ORT review.

Overview

[1,2,3] - 3D tensor shape
[1,2] - 2D tensor shape
[1] - 1D tensor shape
[] - 0D tensor shape (scalar)

A 0D tensor shape vs a 1D tensor with a single element are semantically and functionally distinct. For simple operators (e.g. Add/Sin), treating a scalar as if it was a 1D array with a single element yields the same functional result, but for more complex operators where the rank matters (e.g. ReduceSum with all axes and keepDimensions = false, or Gather where the output rank is computed as input.rank + indices.rank - 1...), treating the two as synonymous is problematic and has been causing a multitude of problems with mapping certain ONNX models to WebNN.

References

Every known ML library represents 0D scalars via the shape [] (or ()):

Numpy

import numpy

x = numpy.array(42)
y = numpy.add(x, x)
print("value:", y)
print("shape:", y.shape)

# Prints:
# value: 84
# shape: ()

TensorFlow

import tensorflow as tf

x = tf.constant(42, dtype=tf.float32)
y = tf.add(x, x);
print("value:", y)
print("shape:", y.shape)

# Prints:
# value: tf.Tensor(84.0, shape=(), dtype=float32)
# shape: ()

PyTorch

import torch

x = torch.tensor(42, dtype=torch.float)
y = torch.add(x, x)
print("value:", y)
print("shape:", y.shape)
# Prints:
# value: tensor(84.)
# shape: torch.Size([])

ONNX

import onnx

# Scalar via [].
x = onnx.helper.make_tensor(
    "value", onnx.TensorProto.FLOAT, [], [42]
)

SafeTensorss

The SafeTensors file format (commonly used with Stable Diffusion models for custom weights) explicitly allows 0D scalars and 0-size tensors.

XNNPack

Bin Miao says below that XNNPack works.

Implementation experience

In my Chromium fork, I just needed to change 2 lines, deleting 1 errant validation statement in the graph builder, and adding 1 in the DML backend (TensorDesc constructors sets a minimum rank of 1) for the rest of it to work. Then scalar tests passed:

    // 1. Create a computational graph
    const inputTensorDesc = {type: 'float32', dimensions: []};
    const constantDesc = { type: 'float32', dimensions: []};

    const A = graphBuilder.input('A', inputTensorDesc);
    const B = graphBuilder.constant(constantDesc, new Float32Array([42.0]));
    const C = graphBuilder.add(A, B);
    LogTensorShapes(graphBuilder, {"A": A, "B": B, "C": C});

    // 2. Build the graph into an executable.
    const graph = await graphBuilder.build({'C': C});

    // 3. Bind inputs to the graph and execute for the result.
    const bufferA = new Float32Array([42]);
    const bufferC = new Float32Array(1);
    const inputs = {'A': bufferA};
    const outputs = {'C': bufferC};
    await mlContext.compute(graph, inputs, outputs);

    console.log(C);
    console.log(C.shape()); // Private extension needed for debugging

Specification Update

The problematic statement in the spec is If dimensions.length is 0, return false.,
https://www.w3.org/TR/webnn/#api-mloperand-create. Delete that line, and the rest should work.

The computation for the element count and byte length is even already correct (it would return 1 element for empty dimensions) https://www.w3.org/TR/webnn/#api-mloperanddescriptor:

The byte length of an MLOperandDescriptor desc is the value returned by the following steps:
Let elementLength be 1.

For each dimension of desc.dimensions:

Set elementLength to elementLength × dimension.

Let elementSize be the element size of one of the ArrayBufferView types that matches desc.type according to this table.

Return elementLength × elementSize.

Code locations to update

• https://chromium-review.googlesource.com/c/chromium/src/+/5004005 : ValidateReduceAndInferOutput