Add decomposition for ONNXSoftmaxCrossEntropyLossOp

[srcarroll]

No description provided.

[jenkins-droid]

Can one of the admins verify this patch?

[srcarroll]

i'm not really familiar with them so will have to look into it, but should i add to test/backend/inference_backend.py or do there already exists relevant tests for this op (i only see shape inference tests)? would such a test belong somewhere else?

[srcarroll]

also i'm not considering the ignore_index attribute here and will leave as a todo. but i don't know of a good way to match fail this case. any suggestions?

Edit: actually just realized it is an optional, not defaulted, attribute. so i know how to check. but if anyone has ideas on how to extend this to support it, i'm all ears. the simplest thing i can think of is to just subtract the slice at ignore_index from the result of ReduceSum.

[jenkins-droid]

Can one of the admins verify this patch?

[AlexandreEichenberger]

@jenkins-droid test this please

[jenkins-droid]

Can one of the admins verify this patch?

[srcarroll]

i think i misunderstood the mean reduction with weights case. looking into it now (my recent "fix" commit is wrong)

[srcarroll]

i think i misunderstood the mean reduction with weights case. looking into it now (my recent "fix" commit is wrong)

so i was just summing over the original weights tensor, but i should have been summing over W[n][d1][d2]...[dk] = weights[labels[i][d1][d2]...[dk]]. The simplest way i can think of producing W is doing onnx.Einsum(one_hot_labels, weights) {equation "ij...,j->i...} (or just a matmul if input is 2D). Any objection to this or any other suggestions? It probably wouldn't be very performant as one_hot_labels is a sparse matrix, but other solutions i can think of involve nasty indexing that i would like to avoid unless insisted upon.

[AlexandreEichenberger]

@jenkins-droid test this please

[jenkins-droid]

Can one of the admins verify this patch?

[srcarroll]

@jenkins-droid test this please

@AlexandreEichenberger you might want to cancel and rerun since i just pushed. also is there a rule of thumb for running tests? since this is still in draft i might push commits that shouldn't be tested yet. maybe i should just not have this drafted yet? sorry if i'm not following some etiquette for this

[jenkins-droid]

Can one of the admins verify this patch?

[AlexandreEichenberger]

@jenkins-droid test this please

[AlexandreEichenberger]

@jenkins-droid test this please

[srcarroll]

@AlexandreEichenberger should i undraft to get some feedback on the testing i asked about? I figured i should get feedback first but i don't know of a good person to ping for this.