unsupported layer type "Sign" when compiling onnx to blob

[mrvladimir]

I'm trying to compile an onnx model to blob format using the online compiler tool, and I'm getting the following error:

[ GENERAL_ERROR ] /home/jenkins/agent/workspace/private-ci/ie/build-linux-ubuntu20/b/repos/openvino/inference-engine/src/vpu/graph_transformer/src/frontend/frontend.cpp:441 Failed to compile layer "sign": unsupported layer type "Sign"

Inference Engine:

IE version ......... 2021.4.2

Build ........... 2021.4.2-3974-e2a469a3450-releases/2021/4

[Warning][VPU][Config] Deprecated option was used : VPU_MYRIAD_PLATFORM

Looking at the docs here the "Sign" layer should be supported: https://docs.openvino.ai/latest/openvino_docs_MO_DG_prepare_model_Supported_Frameworks_Layers.html#onnx-supported-operators

Is this a bug or am I doing something wrong?

[Peh_Intel]

Hi Mr Vladimir,

Have you try compiling the ONNX model using Compile Tool that included in the Intel® Distribution of OpenVINO™ toolkit?

Meanwhile, you can share your ONNX model with us for further investigation.

Regards,

Peh

[mrvladimir]

I still get the same error. I compiled compile_tools from source.

Here's my model:

import numpy as np

import onnx.helper as h

import onnxruntime as rt

from onnx import numpy_helper

from onnx import TensorProto

import onnx

import blobconverter

import math

nodes = [

h.make_node(

'Constant',

inputs=[],

outputs=['floor_mask'],

value=numpy_helper.from_array(np.array([1,2,3], dtype=np.float32))

),

h.make_node('Sub', inputs=['x', 'floor_mask'], outputs=['sub']),

h.make_node('Sign', inputs=['sub'], outputs=['sub_signed']),

]

g = h.make_graph(nodes, 'filtering', [h.make_tensor_value_info('x', TensorProto.FLOAT, [3])],

[h.make_tensor_value_info('sub_signed', TensorProto.FLOAT, [3])])

model = h.make_model(g, producer_name='vladimir')

model.opset_import[0].version = 16

onnx.checker.check_model(model)

onnx.save_model(model, 'test.onnx', )

I compile like so:

$ mo --input_model test.onnx

$ /usr/local/tools/compile_tool/compile_tool -m ~/experiments/floor/test.xml -ip U8 -op FP32 -d MYRIAD -o ~/experiments/floor/test.blob

OpenVINO Runtime version ......... 2022.2.0
Build ........... custom_master_ab8c2f6fd8ac9088a90f1009f8842498c7bda197
Network inputs:
x : u8 / [...]
Network outputs:
sub_signed/sink_port_0 : f32 / [...]
[ GENERAL_ERROR ]
/home/vlad/openvino/src/plugins/intel_myriad/graph_transformer/src/frontend/frontend.cpp:594 Failed to compile layer "sub_signed": unsupported layer type "Sign"

[Peh_Intel]

Hi Mr. Vladimir,

Thanks for sharing your ONNX model with us.

I obtained the same error (unsupported layer type "Sign") when compiling the IR files to blob file. Furthermore, I also ran Benchmark App with the ONNX model on Intel® Neural Compute Stick 2 (NCS2). It showed up the same error which NCS2 was unable to read ONNX model due to the unsupported layer type "Sign".

By the way, the ONNX model can be read and load to CPU.

Regards,

Peh

[Peh_Intel]

Hi Mr. Vladimir,

Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored. 

Regards,

Peh

[mrvladimir]

This sounds like a bug. Is there a work around?

[Peh_Intel]

Hi Mr Vladimir,

In this case, I would like to channel this issue to our development team for the rectification on the support layer type “Sign”.

Any fix or workaround is expected to be available in the future release.   

Sincerely,

Peh

[Hari_B_Intel]

Hi Vladimir,

Thank you for your patience, we received information from the development team that the "Sign" operation is not supported by Myriad/HDDL plugins.

Apology for the inconvenience and we will update the documentation accordingly.

[Peh_Intel]

Hi Mr Vladimir,

Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored.

Regards,

Peh