ONNX Inferencing Error

[mohamadmansourX]

Hello,
Thank you for this great project to properly convert SparseInst to Onnx/TensorRT.

I started training through your updated implementation using a customized version of configs/sparse_inst_r50_giam_aug.yaml and 10 classes

I converted to onnx successfully

python3 convert_onnx.py --config-file configs/sparse_inst_r50_giam_aug.yaml --output on2/myonnx2.onnx --image assets/figures/t1.jpg --opts MODEL.WEIGHTS output/checkpoints/model_0001999.pth

Then when trying to inference using

python3 eval_tensorrt_onnx.py  -c 0.2 --width_resized 640 --height_resized 640 --input assets/figures/* --use_onnx --onnx_engine on2/myonnx2.onnx --output_onnx on2/ --save_image

I'm receiving the below error!

onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Load model from on2/myonnx2.onnx failed:Type 
Error: Type parameter (T) of Optype (Add) bound to different types (tensor(float) and tensor(double) in node (Add_398).

[leandro-svg]

Dear @mohamadmansourX, thank you for saying that 😉
Would you mind sharing your ONNX file with me? Such that I can check it out on netron with the verbose output.

On my ONNX, the last Add node I have a the Add_319 which is link to the "rescoring_mask" definition in the sparseinst.py, called line 229. If it is the case, you should be able to change line 21 which define mask_pred_ as a float that could possibily be converted to double.
Let's see from where your Add_398 comes from ...

[mohamadmansourX]

Hello, thank you for your reply.
Yeah as you mentioned, after debugging with netron, the issue is in this line:

SparseInst_TensorRT/sparseinst/sparseinst.py

Line 22 in 1bbff3c

return scores * ((masks * mask_pred_).sum([1, 2]) / (mask_pred_.sum([1, 2]) + 1e-6))

Adding the terms mask_pred_.sum([1, 2]) and 1e-6 is causing that issue.
Tried changing 1e-6 to torch.tensor(1e-6).float() but for some reason it's still being interpreted as a tensor(double)!

[mohamadmansourX]

Done!!
I solved it by converting the array to double then float rather than 1e-6. Not sure why 1e-6 was mapped to double no matter what I do.
My solution:

@torch.jit.script
def rescoring_mask(scores, mask_pred, masks):
    mask_pred_ = mask_pred.float()
    factrr = mask_pred_.sum([1, 2]).double()
    factrr = factrr + 1e-6
    return scores * ((masks * mask_pred_).sum([1, 2]) / factrr.float())