Tensorflow2 saved model workflow

[m-herrera]

Hi all,
I am trying to load a tensorflow v2 model with NNStreamer; however I get an error when loading the model.
The error I get comes from this line: L248

Does NNstreamer tensorflow support loading the SavedModel format?
I believe the error happens because I'm using the SavedModel from tf v2 instead of the GraphDef from tf v1 directly.

I am building from source with the latest in the main branch.

Thanks in advance!

[myungjoo]

What's the error, then?
And does it work with its original model?
What about other (simple) saved models? (a model file to be tested?)

[m-herrera]

The error comes from L248, the message shown is from L253

The specific error looks like the following:

** (gst-launch-1.0:11612): CRITICAL **: 07:32:35.162: Error deleting graph!! - [Code: 3] Invalid GraphDef

** (gst-launch-1.0:11612): CRITICAL **: 07:32:35.162: Failed to load model
Caught SIGSEGV

It can be reproduced using the example model found here and the following pipeline:

gst-launch-1.0 videotestsrc ! videoconvert ! video/x-raw,format=RGB,width=224,height=224 ! queue ! tensor_converter input-dim=1:224:224:3 ! tensor_filter input=1:224:224:3 inputtype=uint8 framework=tensorflow model=./saved_model.pb ! fakesink

I've been able to run the tensorflow examples without any issues, however these I believe don't use the SavedModel export format.

[myungjoo]

Could you please give me the direct link to the .pb file? not the instruction on how to make one?

[m-herrera]

I uploaded the saved model files here. Tensorflow v2 exports a folder not only the .pb file, you can run the pipeline with the .pb file in that folder and should get the same error.