ScatterND's doc gives wrong result for 3d indice case

[haowang5128]

Bug Report

Describe the bug

In ScatterND doc , it shows scatternd computation as:

output = np.copy(data)
update_indices = indices.shape[:-1]
for idx in np.ndindex(update_indices):
    output[indices[idx]] = f(output[indices[idx]], updates[idx])

For a 3d indices case like this:

input = [[[0., 1.],  [2., 3.]], [[4., 5.], [6., 7.]]]
indices = [[[0, 0]]]
updates = [[[100., 101.]]]
reduction  = None

It provides results different from ONNXRuntime.

• The doc's method: [[[100., 101.], [100., 101.]], [[ 4., 5.], [ 6., 7.]]]
• ONNXRuntime: [[[100., 101.], [ 2., 3.]], [[ 4., 5.], [ 6., 7.]]]

In the doc's method, indices[idx] gives a numpy array, output[indices[idx]] will not select the correct element. Convert indices[idx] to tuple will fix this.

output = np.copy(data)
update_indices = indices.shape[:-1]
for idx in np.ndindex(update_indices):
    output[tuple(indices[idx])] = f(output[tuple(indices[idx])], updates[idx])

[liqunfu]

@haowang5128 , thank you for pointing out this. Would you like to create a PR for the fix. If so, please also correct sample implementation at:

onnx/onnx/backend/test/case/node/scatternd.py

Line 26 in 9cc907f

output[tuple(indices[i])] = np.maximum(output[indices[i]], updates[i])

. The code does produce correct test data. However, to have extra tuple at the right hand side makes more sense.

[xadupre]

Is it on CPU or GPU? This PR addresses one issue on GPU: microsoft/onnxruntime#19540.