[tmva][sofie] Add new ScatterND operator#21621
Open
lmoneta wants to merge 4 commits intoroot-project:masterfrom
Open
[tmva][sofie] Add new ScatterND operator#21621lmoneta wants to merge 4 commits intoroot-project:masterfrom
lmoneta wants to merge 4 commits intoroot-project:masterfrom
Conversation
Add an implementation of ScatterND operator which is needed to parse the MLPF model from CMS Include also 3 tests to probe the different type of scattering wich can be performed
guitargeek
reviewed
Mar 16, 2026
Contributor
guitargeek
left a comment
guitargeek
left a comment
There was a problem hiding this comment.
Great! I have also tested it with my AD refactoring PR, and it works well.
I have a few small suggestions though, and also one more general question: why not format new source file with clang-format? That would make the code more consistent in the long run, and also not confuse new contributors with the CI checks that are red when the code is not formatted.
Comment on lines
+15
to
+26
| if (!parser.IsRegisteredTensorType(nodeproto.input(0))){ | ||
| throw std::runtime_error("TMVA::SOFIE ONNX Parser ScatterND op has input tensor " + nodeproto.input(0) | ||
| + " but its type is not yet registered"); | ||
| } | ||
| if (!parser.IsRegisteredTensorType(nodeproto.input(1))){ | ||
| throw std::runtime_error("TMVA::SOFIE ONNX Parser ScatterND op has input tensor " + nodeproto.input(1) | ||
| + " but its type is not yet registered"); | ||
| } | ||
| if (!parser.IsRegisteredTensorType(nodeproto.input(2))){ | ||
| throw std::runtime_error("TMVA::SOFIE ONNX Parser ScatterND op has input tensor " + nodeproto.input(2) | ||
| + " but its type is not yet registered"); | ||
| } |
Contributor
There was a problem hiding this comment.
I guess that can also be done in a loop from zero to nodeproto.input_size().
Comment on lines
+41
to
+45
| std::unique_ptr<ROperator> op; | ||
| std::string output_name = nodeproto.output(0); | ||
|
|
||
| op.reset(new ROperator_ScatterND(nodeproto.input(0), nodeproto.input(1), nodeproto.input(2), | ||
| output_name, reduction)); |
Contributor
There was a problem hiding this comment.
Suggested change
| std::unique_ptr<ROperator> op; | |
| std::string output_name = nodeproto.output(0); | |
| op.reset(new ROperator_ScatterND(nodeproto.input(0), nodeproto.input(1), nodeproto.input(2), | |
| output_name, reduction)); | |
| auto op = std::make_unique<ROperator_ScatterND>(nodeproto.input(0), nodeproto.input(1), nodeproto.input(2), | |
| nodeproto.output(0), reduction)); |
We can use C++17 now, and it's better to have no naked new or delete, because these can be a red flag.
Comment on lines
+5
to
+7
| namespace TMVA { | ||
| namespace Experimental { | ||
| namespace SOFIE { |
Contributor
There was a problem hiding this comment.
Suggested change
| namespace TMVA { | |
| namespace Experimental { | |
| namespace SOFIE { | |
| namespace TMVA::Experimental::SOFIE { |
Personal preference maybe, but I think this is more readable.
Comment on lines
+12
to
+15
| namespace TMVA{ | ||
| namespace Experimental{ | ||
| namespace SOFIE{ | ||
|
|
Contributor
There was a problem hiding this comment.
Suggested change
| namespace TMVA{ | |
| namespace Experimental{ | |
| namespace SOFIE{ | |
| namespace TMVA::Experimental::SOFIE { | |
Same here maybe.
Test Results 22 files 22 suites 3d 5h 8m 41s ⏱️ For more details on these failures, see this check. Results for commit 9260cd5. ♻️ This comment has been updated with latest results. |
- Fix in operator Reduce to return a scalar and not a tensor of shape [1] - Fix handling of output boolean type in Cast. Do not convert type in a string, because a boolean is converted to a uint8_t which can be a native uint8_t or a bool. Avoid then calling function ConvertStrigToType if possible - Fix fusion of operators. Perform fusion not at first op encountered but at the last onem in order to parse before all operators which can provide an input to last fused one. This was the case in MLPF where there was a MatMul + Constant + Add, where COnstant is an input to Add. - remove check in Generate on empty shapes because scalars tensors have empty shapes
Fix also a bug when doing Gemm and applying the bias in case of stacked matrix multiplications. The bias was not correctly broadcasted in this case
lmoneta
force-pushed
the
tmva_sofie_new_operator_devs
branch
from
58a9ff4 to
9260cd5
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Add implementation of ScatterND operator which is needed to parse MLPF model from CMS
Include also 3 tests to probe the different type of scattering wich can be performed by the operator