[webnn] Add float32 tests for WebNN resample2d op #41811
Conversation
Both are reasonable and not really "wrong" - we just have to pick one. The only truly wrong ones are those libraries that always floor pixels (broken libraries) which result in shifted images, but I believe most image resampling API's round input coordinate on halves toward negative infinity, and ONNX's Resize defaults This is easy to control via |
|
@fdwr May we first to land PASS tests and create another pr of these failure 3 tests on "nearest-neighbor" mode which would help for resolving https://issues.chromium.org/issues/327337526, thanks. |
Yeah, go ahead 👍. |
BruceDai
force-pushed
the
add_webnn_resample2d_tests
branch
from
9b9a710
to
1750e02
Compare
| * Get ULP tolerance of resample2d operations. | ||
| * @param {Object} resources - Resources used for building a graph | ||
| * @param {String} operationName - An operation name | ||
| * @returns {Number} A tolerance number |
There was a problem hiding this comment.
| * @returns {Number} A tolerance number | |
| * @returns {Number} A tolerance number in ULP |
(to clarify it's not absolute difference, because 84 would be a huuuge tolerance 😉)
| tolerance = 84; | ||
| } else if (precisionType === 'float16') { | ||
| tolerance = 10; | ||
| } |
There was a problem hiding this comment.
else {
tolerance = 1;
}
tolerance remains uninitialized in the else code path path. For integer types (e.g. uint8), even if many backends don't currently support them, a reasonable default for linear is 1 ULP (due to rounding up or down).
There was a problem hiding this comment.
Note my comment above too (in case you missed it), but approved either way.
| tolerance = 84; | ||
| } else if (precisionType === 'float16') { | ||
| tolerance = 10; | ||
| } |
…#41811) * [webnn] Add float32 tests for WebNN resample2d op * [webnn] Removed some downsample tests for resample2d * [webnn] Rebased code * [webnn] Add default ULP tolerance for resample2d on linear mode.
@fdwr This PR is to add 13 float32 resample2d tests. I ran these tests by internal Chromium build using WebNN DirectML backend, got 10 PASS and 3 FAIL.
These 3 FAIL tests are all for resample(down sampling) on "nearest-neighbor" mode with below 4D input of shape [1, 1, 4, 6]
by
options.sizes = [2, 3]oroptions.scales = [0.5, 0.5],output is of shape [1, 1, 2, 3] and its actual value as below
while our expected data were generated by https://github.com/webmachinelearning/webnn-baseline/blob/main/src/resample2d.js with same input data and options:
I'm not sure which is wrong on
Do you have any ideas? Thanks.