Can't import any model using ImportGraphDef

[niskov]

Tried the following:

1. Train model (python) on Keras (TF 2.2) and saving the model as h5 - can't import [Exception: 'InvalidGraphDef']
2. Freeze the above model and migrate to .pb file (similar to what we've done @ tf 1.15) - can't import [Exception: 'Input 1 of node bn_Conv1/cond was passed float from bn_Conv1/gamma:0 incompatible with expected resource.']
3. Train model (python) on Keras (TF 2.2) and saving the model as pb ('savedmodel' format) - can't import [Exception: 'InvalidGraphDef']

Is there's another way to import models?

[emgucv]

The issues has been fixed in this commit:
476fe2a
TF 2.x can only saved model as a "SavedModel" format, which exports to a folder. The folder needs to be loaded with the added function above.

TensorFlow v1.x exported the models to a single .pb file, which is no longer supported in python Tensorflow v2.x. See:
tensorflow/tensorflow#29253

We also added a resnet demo to show how to import "SavedModel" folder from TF 2.0 here:
9390abe

This will be included as part of the up coming TF 2.3 release.

Let me know if you want to give it a try and I can send over a development build.

[vpenades]

Is there a roadmap for TF 2.3 release? we're having a similar issue and we suspect it's related to requiring TensorFlow 2.3

We don't need an exact date.... but knowing if it's going to be a matter of weeks... or months, is critical for us.

[emgucv]

We should be releasing Emgu TF v2.3 in about 2 weeks.

[vpenades]

@emgucv Any update on TF v2.3 release?

[emgucv]

It will be delayed. We run into issues with the TF v2.3 GPU build: We are not able to get TF v2.3.0 to compile with CUDA 10.2 and DNN 8.0. The 2.3.0 release is also not able to compile with CUDA 11, the issue has been fixed in the TF master branch. We may have a release based on the TF master to get the GPU version working.