[webnn] Add float32 tests for WebNN API instanceNormalization op #43891
Conversation
|
|
||
| // https://webmachinelearning.github.io/webnn/#api-mlgraphbuilder-instancenorm | ||
|
|
||
| testWebNNOperation('instanceNormalization', buildLayerNorm, 'gpu'); |
There was a problem hiding this comment.
Here buildLayerNorm function of #43684 can be reused.
webnn/resources/utils.js
Outdated
| @@ -328,6 +328,7 @@ const PrecisionMetrics = { | |||
| elu: {ULP: {float32: 18, float16: 18}}, | |||
| expand: {ULP: {float32: 0, float16: 0}}, | |||
| gemm: {ULP: {float32: getGemmPrecisionTolerance, float16: getGemmPrecisionTolerance}}, | |||
| instanceNormalization: {ATOL: {float32: 1/1024, float16: 1/512}}, | |||
There was a problem hiding this comment.
This's a placeholder. I tried to test by ULP with input [2, 3, 30, 30] and got the maximums ULP data is about 388. By this ATOL tolerance, tests could run pass. @fdwr Please help provide this tolerance criteria when you're available, thanks.
There was a problem hiding this comment.
The division by square root really complicates giving a straight forward answer ((a - mean) * scale / sqrt(variance + epsilon) + bias). Just based on the number of lossy operations, you'd expect a worst case of about...
6 + mean cost + variance cost
mean cost = ElementsToReduce + 1
variance cost = (Input.Sizes.W * Input.Sizes.H) * 3 + 1
...but because the subtraction can produce nearly equal numbers, and the division magnifies differences, we get a larger delta than that. Comfortable tolerant values for many GPU's would be ~840 ULP for float32 and 8'400 ULP for float16.
There was a problem hiding this comment.
@fdwr Thanks!
I've updated tolerance as your recommended, please take another look, ☕
| } | ||
| }, | ||
| { | ||
| "name": "instanceNormalization float32 4D tensor options.scale", |
There was a problem hiding this comment.
Here's similar issue "NotSupportedError: DirectML: instanceNormalization: The scale and bias must be both given or not given." like layerNormalization .
BruceDai
force-pushed
the
add_webnn_instanceNorm
branch
from
50b69d7
to
5ce1b51
Compare
|
I've rebased to the latest master code, PTAL, Thanks. |
@fdwr @shiyi9801 @lisa0314 @Honry PTAL, thanks.