Error when loading in Python an .onnx neural net exported via Matlab

Hello,

I can't use in Python an .onnx neural net exported with Matlab. Let say I want to use the googlenet model, the code for exporting it is the following:

net = googlenet;
filename = 'googleNet.onnx';
exportONNXNetwork(net,filename);

In Python, commands for loading the .onnx file are the following (according to https://microsoft.github.io/onnxruntime/)

import onnxruntime 
sess = onnxruntime.InferenceSession('googlenet.onnx') 

But an error message occurs at this stage:

RuntimeError: [ONNXRuntimeError] : 1 : GENERAL ERROR : Load model from googlenet.onnx failed:

Node:prob Output:prob [ShapeInferenceError] Mismatch between number of source and target dimensions. Source=2 Target=4

I tried different net (alexnet, squeezenet, personal nets...) and the same error always appears.

Here is my config:

-----------------------------------------------------------------------------------------------------

MATLAB Version: 9.6.0.1072779 (R2019a)

Operating System: Microsoft Windows 10 Pro Version 10.0 (Build 17763)

Java Version: Java 1.8.0_181-b13 with Oracle Corporation Java HotSpot(TM) 64-Bit Server VM mixed mode

-----------------------------------------------------------------------------------------------------

Deep Learning Toolbox Version 12.1 (R2019a)

Any help is welcomed !

Answer 1

[Don Mathis]

I could not reproduce your error. The following works for me:

In MATLAB:

>> net = googlenet;
>> filename = 'googleNet.onnx';
>> exportONNXNetwork(net,filename,'OpsetVersion',8)

In python:

import numpy
import onnxruntime as rt
sess = rt.InferenceSession("googleNet.onnx")
input_name = sess.get_inputs()[0].name
n = 1
c = 3
h = 224
w = 224
X = numpy.random.random((n,c,h,w)).astype(numpy.float32)
pred_onnx = sess.run(None, {input_name: X})
print(pred_onnx)

It outputs:

[array([[3.29882569e-05, 3.58083460e-04, 3.37624690e-04, 1.43901940e-04, 5.39901492e-04, 4.93929256e-04, 1.84278106e-04, 1.47032852e-05, 3.41630061e-06, 7.50037043e-06, 2.41960952e-05, 4.77660433e-06, 8.67359086e-06, 8.24564086e-06, 2.09670925e-05, 2.51299825e-05, 2.65392214e-06, 3.01301202e-06, 1.45755412e-05, 6.66411279e-06, 2.57993106e-05, 1.68685292e-05, 4.03514641e-05, 3.40506740e-05, 6.18301056e-05, 1.30592525e-05, 7.45224024e-05, 5.93718396e-05, 2.10106184e-04, 2.63419988e-05, 5.05311709e-06, 1.60537282e-04, 6.04824818e-05, 1.52395834e-04, 9.41899605e-04, 1.93663309e-05, 1.47942395e-04, 1.34101238e-05, 4.75002344e-05, 1.01176765e-05, 8.80616863e-05, 1.62361575e-05, 2.06871373e-05, 1.32702444e-05,

...

My MATLAB config:

>> ver
-----------------------------------------------------------------------------------------------------
MATLAB Version: 9.6.0.1092380 (R2019a) Update 1
MATLAB License Number: unknown
Operating System: Microsoft Windows 10 Enterprise Version 10.0 (Build 17134)
Java Version: Java 1.8.0_181-b13 with Oracle Corporation Java HotSpot(TM) 64-Bit Server VM mixed mode
-----------------------------------------------------------------------------------------------------

Deep Learning Toolbox                                 Version 12.1        (R2019a)

[Patrick Marmaroli]

Thank you for the quick reply.

Upgrading the OpsetVersion from 6 to 8, as in your example, still returned me the same error but I observed it disappears after:

1) retrograding the onnxruntime version (current version is 0.4.0)

pip uninstall onnxruntime 
pip install onnxruntime==0.3.0

2) and using an OpsetVersion for exportONNXNetwork superior to 7, e.g.

net = googlenet;
filename = 'googleNet.onnx';
exportONNXNetwork(net,filename,'OpsetVersion',9)

However, when I run your example code, a new error occurs (I tried all the OpsetVersion from 6 to 9 and onnxruntime ver 0.3.0 and ver 0.2.1)

import numpy
import onnxruntime as rt
sess = rt.InferenceSession("googleNet.onnx")
input_name = sess.get_inputs()[0].name
n = 1
c = 3
h = 224
w = 224
X = numpy.random.random((n,c,h,w)).astype(numpy.float32)
pred_onnx = sess.run(None, {input_name: X})
print(pred_onnx)

returns

RuntimeError: Method run failed due to: [ONNXRuntimeError] : 1 : GENERAL ERROR : d:\3\s\onnxruntime\core\providers\cpu\math\gemm_helper.h:13 onnxruntime::GemmHelper::GemmHelper left.NumDimensions() == 2 || left.NumDimensions() == 1 was false.

Does it speak to you ?

[Don Mathis]

You may need to download the latest version of MATLAB's onnx support package. There was an April update.

[Martijn]

Thanks Don, I spoke to Patrick and after helping him update to the latest version, things are working correctly.

[Patrick Marmaroli]

Indeed, pb solved with "Deep Learning Toolbox Converter for ONNX Model Format" version 19.1.2 (I used the version 19.1.0). Thank you guys.