add minimum data types and rank range for operations #910
Conversation
BruceDai
changed the title
| <tr> | ||
| <td>{{input}}</td> | ||
| <td>specified as part of operation steps</td> | ||
| <td></td> |
There was a problem hiding this comment.
According to @fdwr's comments on #278 (comment), l2Pool2d is still not implemented in TFLite backend yet, so here I just followed @huningxin's comments on #853.
"l2Pool2d": {
"input": {
"dataTypes": [],
"rankRange": {}
},
"output": {
"dataTypes": []
}
},
There was a problem hiding this comment.
That TFLite l2Pool2d issue has been pending since 2022-06-30 😳. The Chromium backend that calls TFLite could just emulate it for now to unblock us. e.g.:
powResult = pow(input, exponent)
poolSumResult = poolSum(powResult, axes, windowDimensions, padding, strides, dilations)
// where poolSum can be emulated via poolAverage(input * product(windowDimensions)...)
// or by convolution with a filter of 1's.
result = root(poolSumResult, exponent)
(no action expected in this PR - resolve me)
BruceDai
changed the title
@BruceDai : Remerge with main. My PR adds an extra blank line that pacifies the build error (and yes, that line was already correct, and nothing changed there for months, but bikeshed changed). |
reillyeon
left a comment
reillyeon
left a comment
There was a problem hiding this comment.
General wording looks good. I have not validated the specific operator limits specified.
fdwr
left a comment
fdwr
left a comment
There was a problem hiding this comment.
👀 Thanks for the updates Bruce - I'm sure they were tedious :b.
| <td>[=/any data type|any=]</td> | ||
| <td>{{MLOperandDataType/"float32"}}, {{MLOperandDataType/"float16"}}, {{MLOperandDataType/"int32"}}</td> | ||
| <td>[=/any rank|N=]</td> | ||
| <td>0 to 4</td> |
There was a problem hiding this comment.
Which ML API is the outlier for pow being 0 to 4 instead of 0 to 5? It will be problematic for callers if mul/div/pow have inconsistent ranks from each other, as certain substitution operations might expect that they can safely chain these or switch from one to another (e.g. mul(a,a) == pow(a,2)) without surprise. I propose the outlier API fix this in their Chromium backend with flanking reshapes (flatten to 1D if > 4D, pow, then back to the original shape). Then pow can be remerged back with the rest above.
There was a problem hiding this comment.
Which ML API is the outlier for pow being 0 to 4 instead of 0 to 5?
It's limited by TFLite backend.
Please refer to https://source.chromium.org/chromium/chromium/src/+/main:services/webnn/tflite/graph_builder_tflite.cc;l=568
// Limited to 4D when broadcasting is required:
// https://source.chromium.org/chromium/chromium/src/+/main:third_party/tflite/src/tensorflow/lite/kernels/pow.cc
/*pow_input=*/{kFloat16To32AndInt32, SupportedRanks::UpTo(4)},There was a problem hiding this comment.
Can we open a TFLite issue? Alternately we open a Chromium issue to {reshape, expand if needed, pow} until TFLite can address it.
There was a problem hiding this comment.
Filed a Chromium issue - https://issues.chromium.org/issues/471017637.
| <tr> | ||
| <td>{{input}}</td> | ||
| <td>specified as part of operation steps</td> | ||
| <td></td> |
There was a problem hiding this comment.
That TFLite l2Pool2d issue has been pending since 2022-06-30 😳. The Chromium backend that calls TFLite could just emulate it for now to unblock us. e.g.:
powResult = pow(input, exponent)
poolSumResult = poolSum(powResult, axes, windowDimensions, padding, strides, dilations)
// where poolSum can be emulated via poolAverage(input * product(windowDimensions)...)
// or by convolution with a filter of 1's.
result = root(poolSumResult, exponent)
(no action expected in this PR - resolve me)
| <tr> | ||
| <td>{{input}}</td> | ||
| <td>[=/any data type|any=]</td> | ||
| <td>{{MLOperandDataType/"float32"}}, {{MLOperandDataType/"float16"}}, {{MLOperandDataType/"int32"}}, {{MLOperandDataType/"int8"}}, {{MLOperandDataType/"uint8"}}</td> |
There was a problem hiding this comment.
gatherElements includes int8/uint8, but its symmetric counterpart scatterElements does not 😳 I don't know which API has this surprising inconsistency, but I'd rather restrict gatherElements's required data types to exclude int8/uint8 than introduce this asymmetry.
| </thead> | ||
| <tr> | ||
| <td>{{input}}</td> | ||
| <td>specified as part of operation steps</td> |
There was a problem hiding this comment.
Should we reflect allowed data types in table?
There was a problem hiding this comment.
@huningxin In operation steps, the input of averagePool2d supports "float32", "float16" two data types, and the input of maxPool2d supports any data types, so I keep pervious descriptions here.
index.bs
Outdated
| <tr> | ||
| <td>{{input}}</td> | ||
| <td>[=/any data type|any=]</td> | ||
| <td>{{MLOperandDataType/"float32"}}, {{MLOperandDataType/"float16"}}, {{MLOperandDataType/"int32"}}</td> |
There was a problem hiding this comment.
ORT doesn't support int32
There was a problem hiding this comment.
🤔 An ONNX BitCast operator could help with cases of pure data movement (since no actual math happens) like this one when the operator supports a different data type of the same bit size (sizeof(int32) == sizeof(uint32) == sizeof(float32)). For example, we could call BitCast(int32input, type=float32) -> Triangular -> BitCast(input, type=int32). I'll comment on that ONNX PR.
|
Thanks @huningxin and @fdwr for review and identifications. 👍 |
4 participants
This PR is to fix #896.
@huningxin @fdwr @reillyeon PTAL, thanks!
Preview | Diff