Skip to content

Convert onnx Roialign to linalg ir error #4181

Open
@wso4133560

Description

@wso4133560

env ->
OS: ubuntu 22.04

onnx model:
git clone https://github.com/onnx/onnx.git

in onnx project, the model path is
onnx/onnx/backend/test/data/node/test_roialign_mode_max/model.onnx

torch-mlir version: newest in main branch

run command:
python -m torch_mlir.tools.import_onnx ./model.onnx -o onnx.mlir
torch-mlir-opt onnx.mlir --mlir-print-debuginfo --torch-onnx-to-torch-backend-pipeline -o torch.mlir
torch-mlir-opt torch.mlir --mlir-print-debuginfo --torch-backend-to-linalg-on-tensors-backend-pipeline -o linalg.mlir

onnx ir:
module {
func.func @test_roialign_aligned_false(%arg0: !torch.vtensor<[1,1,10,10],f32>, %arg1: !torch.vtensor<[3,4],f32>, %arg2: !torch.vtensor<[3],si64>) -> !torch.vtensor<[3,1,5,5],f32> attributes {torch.onnx_meta.ir_version = 10 : si64, torch.onnx_meta.opset_version = 22 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
%none = torch.constant.none
%0 = torch.operator "onnx.RoiAlign"(%arg0, %arg1, %arg2) {torch.onnx.coordinate_transformation_mode = "output_half_pixel", torch.onnx.output_height = 5 : si64, torch.onnx.output_width = 5 : si64, torch.onnx.sampling_ratio = 2 : si64, torch.onnx.spatial_scale = 1.000000e+00 : f32} : (!torch.vtensor<[1,1,10,10],f32>, !torch.vtensor<[3,4],f32>, !torch.vtensor<[3],si64>) -> !torch.vtensor<[3,1,5,5],f32>
return %0 : !torch.vtensor<[3,1,5,5],f32>
}
}

torch ir:
module {
func.func @test_roialign_aligned_false(%arg0: !torch.vtensor<[1,1,10,10],f32>, %arg1: !torch.vtensor<[3,4],f32>, %arg2: !torch.vtensor<[3],si64>) -> !torch.vtensor<[3,1,5,5],f32> attributes {torch.onnx_meta.ir_version = 10 : si64, torch.onnx_meta.opset_version = 22 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
%int2 = torch.constant.int 2
%int5 = torch.constant.int 5
%float1.000000e00 = torch.constant.float 1.000000e+00
%false = torch.constant.bool false
%int6 = torch.constant.int 6
%none = torch.constant.none
%int1 = torch.constant.int 1
%0 = torch.aten.unsqueeze %arg2, %int1 : !torch.vtensor<[3],si64>, !torch.int -> !torch.vtensor<[3,1],si64>
%1 = torch.aten.to.dtype %0, %int6, %false, %false, %none : !torch.vtensor<[3,1],si64>, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[3,1],f32>
%2 = torch.prim.ListConstruct %1, %arg1 : (!torch.vtensor<[3,1],f32>, !torch.vtensor<[3,4],f32>) -> !torch.list
%3 = torch.aten.cat %2, %int1 : !torch.list, !torch.int -> !torch.vtensor<[3,5],f32>
%4 = torch.torchvision.roi_align %arg0, %3, %float1.000000e00, %int5, %int5, %int2, %false : !torch.vtensor<[1,1,10,10],f32>, !torch.vtensor<[3,5],f32>, !torch.float, !torch.int, !torch.int, !torch.int, !torch.bool -> !torch.vtensor<[3,1,5,5],f32>
return %4 : !torch.vtensor<[3,1,5,5],f32>
}
}

error:
unknown :0: error: failed to legalize operation 'torch.constant.int'
unknown :0: note: see current operation: %0 = "torch.constant.int"() <{value = 2 : i64}> : () -> !torch.int loc(unknown)

Is this an error or not supported? All ONNX models that include the RoiAlign operator produce errors when converting Torch IR to Linalg IR.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions