Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.
6392 Discussions

Cannot convert ssd_mobilenet_v2_coco using Openvino 2019 R2

pham__hai
Beginner
741 Views

Hi everyone,

I have retrained ssd_mobilenet_v2_coco model using tensorflow detection API (only train with one class).

Then I use Openvino 2019 R2 to convert this model to IR, as below command:

python mo.py --input_model frozen_inference_graph.pb --tensorflow_use_custom_operations_config extensions\front\tf\ssd_v2_support.json --tensorflow_object_detection_api_pipeline_config .\pipeline.config  --reverse_input_channels --input image_tensor --log_level DEBUG

This lead me to an error:

I0731 15:49:48.692595  8868 infer.py:32] input[0]: shape = [], value = 300.0
I0731 15:49:48.692595  8868 infer.py:32] input[1]: shape = [], value = -1.0
I0731 15:49:48.692595  8868 infer.py:144] Outputs:
I0731 15:49:48.692595  8868 infer.py:32] output[0]: shape = [], value = 0.0033333334
I0731 15:49:48.692595  8868 infer.py:129] --------------------
I0731 15:49:48.692595  8868 infer.py:130] Partial infer for Postprocessor/Cast_1
I0731 15:49:48.692595  8868 infer.py:131] Op: Cast
E0731 15:49:48.692595  8868 infer.py:180] Cannot infer shapes or values for node "Postprocessor/Cast_1".
E0731 15:49:48.692595  8868 infer.py:181] 0
E0731 15:49:48.692595  8868 infer.py:182]
E0731 15:49:48.692595  8868 infer.py:183] It can happen due to bug in custom shape infer function <function Cast.infer at 0x0000014C5B89B488>.
E0731 15:49:48.692595  8868 infer.py:184] Or because the node inputs have incorrect values/shapes.
E0731 15:49:48.692595  8868 infer.py:185] Or because input shapes are incorrect (embedded to the model or passed via --input_shape).
I0731 15:49:48.692595  8868 infer.py:196] Node "Postprocessor/Cast_1" attributes: {'pb': name: "Postprocessor/Cast_1"

 

I have search many related topics, and tried many commands, but it always leads to error relate to shape.

What shoud I do?

Thanks for any help.

0 Kudos
5 Replies
Shubha_R_Intel
Employee
741 Views

Dear pham, hai

Please refer to a similar post which I recently replied to:

https://software.intel.com/en-us/forums/computer-vision/topic/815126

You need to use ssd_support_api_v1.14.json

 ssd_support_api_v1.14.json is definitely there (Windows 10) n C:\Program Files (x86)\IntelSWTools\openvino_2019.2.242\deployment_tools\model_optimizer\extensions\front\tf  but you must install OpenVIno 2019R2 to get it.

Hope it helps,

Thanks,

Shubha

0 Kudos
pham__hai
Beginner
741 Views

Hi Shubha,

Thank you for your reply,

I have tried to use ssd_support_api_v1.14.json but got same error.

I also tried to add --input_shape=1,300,300,3 when freeze the model (using export_inference_graph.py) and when convert, the error is:

I0801 09:41:38.673687  1232 infer.py:130] Partial infer for Postprocessor/BatchMultiClassNonMaxSuppression/MultiClassNonMaxSuppression/Select
I0801 09:41:38.673687  1232 infer.py:131] Op: Select
E0801 09:41:38.673687  1232 infer.py:180] Cannot infer shapes or values for node "Postprocessor/BatchMultiClassNonMaxSuppression/MultiClassNonMaxSuppression/Select".
E0801 09:41:38.673687  1232 infer.py:181]
E0801 09:41:38.673687  1232 infer.py:182]
E0801 09:41:38.673687  1232 infer.py:183] It can happen due to bug in custom shape infer function <function Select.infer at 0x000001F3241550D0>.
E0801 09:41:38.673687  1232 infer.py:184] Or because the node inputs have incorrect values/shapes.
E0801 09:41:38.673687  1232 infer.py:185] Or because input shapes are incorrect (embedded to the model or passed via --input_shape).
I0801 09:41:38.673687  1232 infer.py:196] Node "Postprocessor/BatchMultiClassNonMaxSuppression/MultiClassNonMaxSuppression/Select" attributes: {'pb': name: "Postprocessor/BatchMultiClassNonMaxSuppression/MultiClassNonMaxSuppression/Select"
op: "Select"
input: "Postprocessor/BatchMultiClassNonMaxSuppression/MultiClassNonMaxSuppression/Less"
input: "Postprocessor/BatchMultiClassNonMaxSuppression/MultiClassNonMaxSuppression/Gather/GatherV2_1"
input: "Postprocessor/BatchMultiClassNonMaxSuppression/MultiClassNonMaxSuppression/mul"
attr {
  key: "T"
  value {
    type: DT_FLOAT
  }
}

I seem that some nodes in model have no shape. I tried to convert pre-trained model ssd_mobilenet_v2_coco (in model zoo) and successful.

Any suggestion about this?

Thank you!

 

0 Kudos
Shubha_R_Intel
Employee
741 Views

Dear pham, hai,

what version of Tensorflow are you using ? Can you update to at least 1.14 ? Please use Tensorflow freeze graph to freeze your model. Some nodes being shapeless is definitely a strange and abnormal phenomenon. Please retry, but this time do not use --input_shape.  Did you select an ssd from The MO Supported Tensorflow List ? And if so which one did you choose to custom-train ? I am assuming SSD MobileNet V2 COCO ?

If that one, notice that the pipeline.config has :

image_resizer {
      fixed_shape_resizer {
        height: 300
        width: 300
      }
    }

Please peruse this document :

http://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_Object_Detection_API_Models.html#fixed_shape_resizer_replacement

Read the part which says this:

If the --input_shape [1, H, W, 3] command line parameter is specified, the Model Optimizer sets the input layer height to H and width to W and convert the model. However, the conversion may fail because of the following reasons:

Thanks,

Shubha

 

0 Kudos
pham__hai
Beginner
741 Views

Hi Shubha,

I have found the root cause.

I train and freeze the model on Colab, which is using python 3.6, and download the freezed model to my computer, which is using  python3.7.

Then I convert model using Openvino and it false.

I have retried with python 3.6 and successful.

Thank you for you helping!

0 Kudos
Shubha_R_Intel
Employee
741 Views

Dear pham, hai,

I'm thrilled to hear of your success ! And most of all, I'm so glad that you shared your learnings with the OpenVino community. We are all in this together !

Shubha

 

0 Kudos
Reply