Does OpenVino Supports GluonCV pretrained SSD-VGG-16-300 model

[Lakshmi_U_Intel]

Hi,

I am not able to generate IR for the GluonCV pretrained SSD-VGG-16-300 model using mscoco dataset.

The command used for running the model-optimizer is:

python mo_mxnet.py --input_model ~/UBD_MXNET/Openvino_Gluon/ssd_300_vgg16_atrous_coco-0000.params --input_shape [1,3,300,300] --log_level=DEBUG

OpenVino Version: openvino_2019.1.133(Linux)

mxnet-mkl version: 1.5.0

gluon-cv version: 0.4.0

CPU Model: Intel(R) Xeon(R) Gold 6128 CPU

The error message is as follows:

[ WARNING ]  Node attributes: {'symbol_dict': {'op': 'broadcast_sub', 'name': '_defaultpreprocess0_broadcast_minus0', 'inputs': [[0, 0, 0], [1, 0, 0]]}, '_in_ports': set(), '_out_ports': set(), 'kind': 'op', 'name': '_defaultpreprocess0_broadcast_minus0', 'type': 'broadcast_sub', 'op': 'broadcast_sub', 'infer': None, 'precision': 'FP32'}
[ ERROR ]  Unexpected exception happened during extracting attributes for node _defaultpreprocess0_init_mean.
Original exception message: Operation 'broadcast_sub' not supported. Please register it as custom op.
 For more information please refer to Model Optimizer FAQ (<INSTALL_DIR>/deployment_tools/documentation/docs/MO_FAQ.html), question #86.

Does OpenVino supports GluonCV pretrained SSD-VGG-16-300 models?

Thank you.

[Shubha_R_Intel]

Dear U Menon, Lakshmi,

We just released OpenVino R2 today and the documentation is not quite updated yet but according to the R1 OpenVino MxNet documentation Gluon is not a supported MxNet model. If you study the documentation for mxnet models however, your model optimizer command doesn't look correct, i.e.

To convert an MXNet* model contained in a model-file-symbol.json and model-file-0000.params, run the Model Optimizer launch script mo.py, specifying a path to the input model file:

python3 mo_mxnet.py --input_model model-file-0000.params

Reading https://gluon-cv.mxnet.io/ it seems like you can create a deep learning project using OpenCV as the inference backend (in this case you will not need Model Optimizer or Inference Engine Core API).

Thanks,

Shubha

[Lakshmi_U_Intel]

Hi Shubha,

Thanks for your reply.

I had already tried the model optimizer with just specifying the input model as mentioned here(https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_MxNet.html) but it asks to specify --input_shape parameter.

Please find the error below.

 python mo_mxnet.py --input_model ~/UBD_MXNET/Openvino_Gluon/ssd_300_vgg16_atrous_coco-0000.params

[ ERROR ]  Input shape is required to convert MXNet model. Please provide it with --input_shape.

 For more information please refer to Model Optimizer FAQ (<INSTALL_DIR>/deployment_tools/documentation/docs/MO_FAQ.html), question #16.
 

Is there any option to use SSD-VGG-16-300 model trained on MS coco dataset in OpenVino R1 using MXNET?

Thanks,

Lakshmi

[Shubha_R_Intel]

Dear U Menon, Lakshmi,

I believe I have addressed your question in your other post . 

Thanks,

Shubha

[Lakshmi_U_Intel]

Hi Shubha,

I tried to convert GluonCV pretrained SSD-VGG-16-300 model using MSCOCO dataset on OpenVino2019R2.

As mentioned in this url (https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_MxNet.html), there is a parameter --enable_ssd_gluoncv in model -optimizer for enabling conversion of GluonCV SSD topologies, even though it is accepting that parameter I am getting the same custom op related error in the latest openvino version as well.

Please look into this issue.

Thanks,

Lakshmi.

[Shubha_R_Intel]

Dear U Menon, Lakshmi 

I reproduced the error using the following command on OpenVino 2019R2 :

"c:\Program Files (x86)\IntelSWTools\openvino_2019.2.242\deployment_tools\model_optimizer\mo_mxnet.py" --input_shape [1,3,300,300] --input_model ssd_300-0000.params --log_level DEBUG --enable_ssd_gluoncv > log.txt 2>&1

Sorry about the trouble. I have filed a high priority bug straightaway.

Thanks !

Shubha

[Lakshmi_U_Intel]

Thanks Shubha. It would be great if you could let us know once the issue is resolved so that I can try out mo_mxnet.py with

GluonCV vgg16_ssd_300

Thanks,

Lakshmi.