CNNNetReader::ReadNetwork() throws on Deconvolution layer

[Cheah__Yee_Keng]

Hi,

An ONNX model, that has been successfully converted by Model Optimizer, generated the following runtime error when InferenceEngine tries to read the IR XML file:

Error reading network: Error of validate layer: 484 with type: Deconvolution. Cannot parse parameter pads_end  from IR for layer 484. Value -4,-4 cannot be casted to int.

With log_level = DEBUG, part of Model Optimizer's output contains:

[ DEBUG ] [ infer:150 ]  --------------------
[ DEBUG ] [ infer:151 ]  Partial infer for 484
[ DEBUG ] [ infer:152 ]  Op: Deconv2D
[ DEBUG ] [ infer:163 ]  Inputs:
[ DEBUG ] [ infer:40 ]  input[0]: shape = [ 1  3 40 80], value = <UNKNOWN>
[ DEBUG ] [ infer:40 ]  input[1]: shape = [ 3  3 16 16], value = [[[[ ... 

[ DEBUG ] [ infer:165 ]  Outputs:
[ DEBUG ] [ infer:40 ]  output[0]: shape = [  1   3 320 640], value = <UNKNOWN>
[ DEBUG ] [ infer:150 ]  --------------------
...
Model Optimizer arguments:
Common parameters:
    - Path to the Input Model:     <withheld>.onnx
    - Path for generated IR:     <withheld>
    - IR output name:     <withheld>
    - Log level:     DEBUG
    - Batch:     Not specified, inherited from the model
    - Input layers:     Not specified, inherited from the model
    - Output layers:     Not specified, inherited from the model
    - Input shapes:     Not specified, inherited from the model
    - Mean values:     Not specified
    - Scale values:     Not specified
    - Scale factor:     Not specified
    - Precision of IR:     FP32
    - Enable fusing:     True
    - Enable grouped convolutions fusing:     True
    - Move mean values to preprocess section:     False
    - Reverse input channels:     False
ONNX specific parameters:
Model Optimizer version:     1.5.12.49d067a0

And the generated IR XML file contains:

        <layer id="106" name="484" precision="FP32" type="Deconvolution">
            <data dilations="1,1" group="1" kernel="16,16" output="3" pads_begin="4,4" pads_end="-4,-4" strides="8,8"/>
            <input>
                <port id="0">
                    <dim>1</dim>
                    <dim>3</dim>
                    <dim>40</dim>
                    <dim>80</dim>
                </port>
            </input>
            <output>
                <port id="2">
                    <dim>1</dim>
                    <dim>3</dim>
                    <dim>320</dim>
                    <dim>640</dim>
                </port>
            </output>
            <blobs>
                <weights offset="<withheld>" size="9216"/>
            </blobs>
        </layer>

What could have caused Model Optimizer to generate pads_end = -4, -4 ?

[Shubha_R_Intel]

Dear Yee, this is a known bug which will be fixed for the next release.  This forum post is a related issue:

https://software.intel.com/en-us/forums/computer-vision/topic/804935

Thanks for using OpenVino !

Shubha

[Cheah__Yee_Keng]

Thanks Shubha, looking forward to the next release!

[Gleb_K_Intel]

Hi Yee, do you have any update on this issue? Is it resolved?

[Shubha_R_Intel]

Dearest Gleb and Yee,

Please download OpenVino 2019 R1 and give it a whirl. Lots of issues have been fixed !

Thanks,

Shubha

[Cheah__Yee_Keng]

Hi Gleb / Shubha,

The issue has been resolved with Model Optimizer 2019.1.0-341-gc9b66a2. The Deconvolution layer's XML now reads

<layer id="120" name="484" precision="FP32" type="Deconvolution">
    <data dilations="1,1" group="1" kernel="16,16" output="3" pads_begin="4,4" pads_end="4,4" strides="8,8"/>
    <input>
        <port id="0">
            <dim>1</dim>
            <dim>3</dim>
            <dim>40</dim>
            <dim>80</dim>
        </port>
    </input>
    <output>
        <port id="2">
            <dim>1</dim>
            <dim>3</dim>
            <dim>320</dim>
            <dim>640</dim>
        </port>
    </output>
    <blobs>
        <weights offset="<withheld>" size="9216"/>
    </blobs>
</layer>

Thanks for your help!

Rgds, Yee Keng

[Shubha_R_Intel]

Dear Yee, 

fantastic. Very happy to hear that this issue is resolved in 2019 R1.

Thanks for using OpenVino !

Shubha