Trouble generating MyriadX blob

dave

Hi, I'm wondering if I could get some help. I'm trying to export a model to MyriadX blob using the script at the end of this post but I am getting the following error message:

{
    "exit_code": -11,

    "message": "Command failed with exit code -11, command: /opt/intel/openvino2022_1/tools/compile_tool/compile_tool -m /tmp/blobconverter/12373da38d034f99b4b4cbb60d0158b1/modelSimplified/FP16/modelSimplified.xml -o /tmp/blobconverter/12373da38d034f99b4b4cbb60d0158b1/modelSimplified/FP16/modelSimplified.blob -c /tmp/blobconverter/12373da38d034f99b4b4cbb60d0158b1/myriad_compile_config.txt -d MYRIAD ",        

    "stderr": "",

    "stdout": "OpenVINO Runtime version ......... 2022.1.0
    Build ........... 2022.1.0-7019-cdb9bec7210-releases/2022/1
    Network inputs:
        A : f16 / [...]
        B : f16 / [...]
    Network outputs:
        out/sink_port_0 : f16 / [...]"
}

I have gone over the operations used with Netron against ONNX Support Layers and I think they are supported. Is there something i've missed?

The script to reproduce the error is below:

from torch import nn
import torch, onnx, blobconverter
from onnxsim import simplify

class model(nn.Module):
    def forward(self, A, B):
        cross = torch.cross(A,B,dim=1)
        cross_mean = torch.mean(cross,dim=0)
        cross_norm = torch.norm(cross_mean)
        cross_mean /= cross_norm
        return cross

X1 = torch.ones((400,3),dtype=torch.float32)
X2 = torch.ones((400,3),dtype=torch.float32)

torch.onnx.export(
    model(),
    (X1,X2),
    "./model.onnx",
    opset_version=12,
    do_constant_folding=True,
    input_names=['A', 'B'],
    output_names=['out'],
    dynamic_axes={'A': {0: 'n'}, 'B': {0: 'n'}}
)

onnx_model = onnx.load("./model.onnx")
model_simplified, check = simplify(onnx_model)
onnx.save(model_simplified, "./modelSimplified.onnx")

blobconverter.from_onnx(
    model="./modelSimplified.onnx",
    output_dir="./model",
    data_type="FP16",
    shaves=1,
    use_cache=False,
    compile_params=[],
    optimizer_params=[]
)

Thanks

jakaskerl

Hi dave
Omitting the

    dynamic_axes={'A': {0: 'n'}, 'B': {0: 'n'}}

solved the issue for me. I'm not sure what this is used for; could you check if there is a workaround?

Thanks,
Jaka

dave

Hi, thanks for your reply. The length of the first dimension of A and B will be the same but their length is variable (together). I'm not sure at the moment if there is another way to have the length of A and B variable.

Do you know if "dynamic_axes" are generally not supported for generating a blob?

Thanks

jakaskerl

Adding @Matija, he might know more.

Matija

Yes, in general the graphs need to be static and that usually holds for most of the embedded chips. One option to support different input shapes would be to use multiple models if you can afford that compute-wise and then use script node to pipe the inputs into the relevant models.

dave

Thanks both for your insight and suggestion on a workaround. I'll see if I can make that work.