Loading PyTorch Faster RCNN with ONNX fails

Specific Domains Machine Learning
1

Hi,
I’m trying to load a trained Faster R-CNN model that has been exported with ONNX:

It has been saved like this:

torch.onnx.export(model, 
                  x[None, :, :, :],
                  "detector.onnx",
                  export_params=True,
                  input_names=["input"],
                  output_names=["output"],
                  dynamic_axes={'input' : {0 : 'batch_size'}, 'output' : {0 : 'batch_size'}})

Loading it like this gives an error:

julia> path = "detector.onnx"
julia> data = Float32.(randn((1, 3, 145, 501)))
julia> out = ONNX.load(path, data)

┌ Error: Error while loading node ("/Split_output_0",) = Split("input", "/Constant_1_output_0")
│     axis => 0
└ @ ONNX ~/.julia/packages/ONNX/weSi3/src/load.jl:66
ERROR: UndefVarError: `args` not defined
Stacktrace:
  [1] load_node!(tape::Umlaut.Tape{ONNX.ONNXCtx}, #unused#::ONNX.OpConfig{:ONNX, :Split}, inputs::Vector{Umlaut.Variable}, attrs::Dict{Symbol, Any})
    @ ONNX ~/.julia/packages/ONNX/weSi3/src/load.jl:229
  [2] load_node!(tape::Umlaut.Tape{ONNX.ONNXCtx}, nd::NodeProto, backend::Symbol)
    @ ONNX ~/.julia/packages/ONNX/weSi3/src/load.jl:56
  [3] load(io::IOStream, args::Array{Float32, 4}; backends::Vector{Symbol}, exec::Bool)
    @ ONNX ~/.julia/packages/ONNX/weSi3/src/load.jl:310
  [4] load
    @ ~/.julia/packages/ONNX/weSi3/src/load.jl:267 [inlined]
  [5] #98
    @ ~/.julia/packages/ONNX/weSi3/src/load.jl:341 [inlined]
  [6] open(f::ONNX.var"#98#99"{Vector{Symbol}, Bool, Tuple{Array{Float32, 4}}}, args::String; kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ Base ./io.jl:395
  [7] open
    @ ./io.jl:392 [inlined]
  [8] #load#97
    @ ~/.julia/packages/ONNX/weSi3/src/load.jl:340 [inlined]
  [9] load(filename::String, args::Array{Float32, 4})
    @ ONNX ~/.julia/packages/ONNX/weSi3/src/load.jl:339
 [10] top-level scope

Does anyone know how to properly load the model?

2

that looks like an ONNX.jl bug: UndefVarError: `args` not defined · Issue #97 · FluxML/ONNX.jl · GitHub. You could try GitHub - jw3126/ONNXRunTime.jl: Julia bindings for onnxruntime instead

1 Like
3

Thanks, I’ll look into the onnx runtime.