Unsupported Activation layer type: exp" ?

[taka]

hi

Unable to execute models containing exp layer.

For example, mobilenet model saved with tr_keras.
Model creation method
------------------------------------------
import tensorflow as tf
from tensorflow import keras

pretrained_model = tf.keras.applications.MobileNet()

# Export the model to a SavedModel
keras.experimental.export_saved_model(pretrained_model, './tmp/mobileNet/keras
------------------------------------------

Model_Optimaizer command
"python mo.py --saved_model_dir .\tmp\mobileNet\keras --output_dir movilenet -b 1"

Conversion of model optimizer can be done without problems.

The following error occurs when running with benchmark_ap

"[ ERROR ] Error reading network: Unsupported Activation layer type: exp"

Is "exp" not supported?
However, the release notes for R2019 have the following descriptions
---------------------
New and Changed in the Release 1
CPU plugin
 Added new Activation types: 
 Exp
---------------------

Are there plans to support "exp layer" on VPU and GPU?

Best regards,

[Shubha_R_Intel]

Dear Taka,

There is a known bug which causes the Unsupported Activation layer type: exp.  In the next release of OpenVino it is fixed, however. Thanks for your patience !

Shubha

[taka]

Thank you !

Does VPU and GPU already support "exp"?
If not, do you have any plans?

Best regards,

[Shubha_R_Intel]

Dear taka, according the IR Layers Catalog exp is in the category of Activation function. Yes, it probably should be supported in VPU and GPU but the Supported Devices document doesn't mention it. I will file a bug on this omission. Activation-Exp should be mentioned and it's not. 

Thanks,

Shubha

[Shubha_R_Intel]

Dear taka,

Please download the latest OpenVino 2019 R1.1.  This bug should be fixed.

Thanks !

Shubha

[taka]

Dear Shubha,

In Open VIno 2019R1.1, I think that the bug is not fixed yet.
(I tried with "w_openvino_toolkit_p_2019.1.148.exe".)

Best regards,

[Shubha_R_Intel]

Dear taka,

Please kindly attach your generated IR here. Let me take a look...

Thanks !

Shubha

[taka]

Best regards,

[Shubha_R_Intel]

Dear taka,

Try a new entry node in your model optimizer command. So to create IR, add the following  --input 0:conv1/Conv2D 

More info about the technique here:

http://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_Cutting_Model.html

Just add that as another switch to your model optimizer command, generate a new IR and re-run inference. Let me know how it works.

Thanks for using OpenVino !

Shubha

[taka]

Dear Shubha,

The result has not changed.

[ ERROR ] Error reading network: Unsupported Activation layer type: exp

Best regards,

[Shubha_R_Intel]

Dearest taka, did you try what I suggested above, i.e. add the following  --input 0:conv1/Conv2D to your model optimizer command ? Please give me your entire model optimizer command. 

Also kindly in a zip file attach your Keras model.

Thanks !

Shubha

[taka]

Dear Shubha,

Attach keras model.

I tried the following two types, but it did not work.

python mo.py --saved_model_dir .\mobileNet --output_dir movilenet -b 1

python mo.py --saved_model_dir .\mobileNet --output_dir movilenet -b 1　--input 0:conv1/Conv2D 

Best regards,

[Shubha_R_Intel]

Dear taka, you have found a bug and I reproduced it. Really sorry for the inconvenience !

Shubha