Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Could you please share a longer description of this change? Including why this change is made, description of previous
behavior and newly introduced differences etc., See guideline: https://chromium.googlesource.com/chromium/src/+/main/docs/contributing.md#uploading-a-change-for-review
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Could you please share a longer description of this change? Including why this change is made, description of previous
behavior and newly introduced differences etc., See guideline: https://chromium.googlesource.com/chromium/src/+/main/docs/contributing.md#uploading-a-change-for-review
+1, very confused why a change which claims to be fixing tolerances for the QDQ subgraph tests is updating the tolerances for a bunch of other tests, but not the QDQ subgraph one.
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
👀
tolerance = sizes.length ?
Just for clarity (so others understand what's going on), I'd name it `elementCount`. e.g.
```
const elementCount = sizes.reduce(
(accumulator, currentValue) => accumulator * currentValue, 1);
tolerance = elementCount;
```
reduceL2: tolerance * 2 + 1,
I think I wrote +2 last time for the wrapping sqrt.
reduceSum: tolerance * 2,
Addition should incur no more loss than multiplication, and reduceProduct takes the elementCount directly. Did you mean to put the `* 2` onto `reduceSumSquare` instead?
Updated. Please take another look, thanks!
Reilly GrantCould you please share a longer description of this change? Including why this change is made, description of previous
behavior and newly introduced differences etc., See guideline: https://chromium.googlesource.com/chromium/src/+/main/docs/contributing.md#uploading-a-change-for-review
+1, very confused why a change which claims to be fixing tolerances for the QDQ subgraph tests is updating the tolerances for a bunch of other tests, but not the QDQ subgraph one.
Could you please share a longer description of this change?
@ningx...@intel.com Added this description, please take another look, thanks!
...is updating the tolerances for a bunch of other tests, but not the QDQ subgraph one
@rei...@chromium.org, The tolerance for the QDQ subgraph tests is defined in `getPrecisionTolerance()` method of utils.js.
And the updated bunch of other tests are for these operators fused in QDQ subgraph, they are using `getPrecisionTolerance()` instead which also provides tolerance for the graph of single operator.
Just for clarity (so others understand what's going on), I'd name it `elementCount`. e.g.
```
const elementCount = sizes.reduce(
(accumulator, currentValue) => accumulator * currentValue, 1);
tolerance = elementCount;
```
👍
I think I wrote +2 last time for the wrapping sqrt.
👍
Addition should incur no more loss than multiplication, and reduceProduct takes the elementCount directly. Did you mean to put the `* 2` onto `reduceSumSquare` instead?
Good catch! 👍
Sorry, I missed put these two tolerance values for `reduceSum` and `reduceSumSquare`, they need swap each other.
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Code-Review | +1 |
// others reduction operators
```suggestion
// other reduction operators
```
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Code-Review | +1 |
lgtm with nits
the cumulative tolerance of each operators of subgraph by modifying
```suggestion
the cumulative tolerance of each operator of a subgraph by modifying
```
And this CL is also update a bunch of other tests of these operators
```suggestion
And this CL also updates a bunch of other tests of operators that are
```
operator, with this CL the updated tolerance definitions is equal to the
```suggestion
operator. With this CL, the updated tolerance definitions is equal to the
```
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Code-Review | +1 |
lgtm
the cumulative tolerance of each operators of subgraph by modifying
```suggestion
the cumulative tolerance of each operator of a subgraph by modifying
```
Acknowledged
And this CL is also update a bunch of other tests of these operators
```suggestion
And this CL also updates a bunch of other tests of operators that are
```
Acknowledged
operator, with this CL the updated tolerance definitions is equal to the
```suggestion
operator. With this CL, the updated tolerance definitions is equal to the
```
Acknowledged
```suggestion
// other reduction operators
```
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Thanks Ningxin for helping review latest changes!👍
Current I fail to rebasline expected results after fixing conflict of rebase latest code, I meet "Access blocked: Authorization Error" when executing rebasline command.
Any suggestions? Thanks!
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Thanks Ningxin for helping review latest changes!👍
Current I fail to rebasline expected results after fixing conflict of rebase latest code, I meet "Access blocked: Authorization Error" when executing rebasline command.
Any suggestions? Thanks!
Try running `git cl creds-check` to make sure you have all the necessary credentials set up.
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Code-Review | +1 |
LGTM (once the rebaseline gets added)
Reilly GrantThanks Ningxin for helping review latest changes!👍
Current I fail to rebasline expected results after fixing conflict of rebase latest code, I meet "Access blocked: Authorization Error" when executing rebasline command.
Any suggestions? Thanks!
Try running `git cl creds-check` to make sure you have all the necessary credentials set up.
Thanks Reilly! 👍
I tried this command, and I also switched to another develop machine, the problem still blocks me.
I'm ongoing to find how to fix it.
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
I've updated expected test results. Now all trybots run Pass.
@rei...@chromium.org and @ningx...@intel.com, please take another look, thanks!
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Code-Review | +1 |
LGTM
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Code-Review | +1 |
lgtm
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Auto-Submit | +1 |
Commit-Queue | +2 |
Thanks all for your review! I'm going to submit this CL.
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
webnn: update tolerance definitions for QDQ subgraph tests
This CL is to update tolerance definitions of QDQ subgraph tests using
the cumulative tolerance of each operator of a subgraph by modifying
`getPrecisionTolerance()` method of `utils.js` help file [1].
And this CL also updates a bunch of other tests of operators that are
fused in QDQ subgraph using `getPrecisionTolerance()` or
`getZeroULPTolerance()` methods instead.
For example, the previous tolerance definitions of this "quantized tanh"
QDQ subgraph test [2] is equal to tolerance definitions of `tanh`
operator. With this CL, the updated tolerance definitions is equal to the
accumulated sum of tolerance of `quantizeLinear` on layer1, tolerance of
`dequantizeLinear` on layer2, tolerance of `tanh` on layer3, tolerance
of `quantizeLinear` on layer4, and tolerance of `dequantizeLinear` on
layer5.
```
input
|
[quantizeLinear] // layer1
|
[dequantizeLinear] // layer2
|
[tanh] // layer3
|
[quantizeLinear] // layer4
|
[dequantizeLinear] // layer5
|
output
```
[1]https://source.chromium.org/chromium/chromium/src/+/main:third_party/blink/web_tests/external/wpt/webnn/resources/utils.js;l=39
[2]https://source.chromium.org/chromium/chromium/src/+/main:third_party/blink/web_tests/external/wpt/webnn/conformance_tests/qdq_subgraph.https.any.js;l=1761
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
The WPT PR for this CL has been merged upstream! https://github.com/web-platform-tests/wpt/pull/53535
Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |