System.NotImplementedException : 'Model load of h5 format has not been supported

[Cilouche]

Hello,
I am trying to load a .h5 model with the load_model function (Tensorflow.Keras.Models) but I get this error
issue:
System.NotImplementedException : 'Model load of h5 format has not been supported. Please submit an issue to https://github.com/SciSharp/TensorFlow.NET/issues if it's needed.'

any suggestions please,
Thank you

[AsakusaRinne]

Hello, the loading of "h5" format keras model has not been supported. We has implemented the loading of "SavedModel" format first because it's recommended by tensorflow. The loading of "h5" format is also on our schedule but is not very urgent now. May I ask why you choose "h5" format instead of "SavedModel"?

[Cilouche]

Thank you for your answer, I chose the .h because it is a network that I have already trained in Python with tensorflow-keras, and since I work in c# I wanted to reuse it

[AsakusaRinne]

I see. The loading of "h5" format is a big feature and support of the loading complex models (such as bert) with "SavedModel" format has higher priority in tf.net. We will implement the loading of of "h5" format model in the future but it costs time. A possible alternative is loading your model in python and save it again to "SavedModel" format, then load it in C#. If your model contains self-defined keras module, please try after the merge of #1022, which added support for it.

[Cilouche]

hello,
I followed your suggestions and loaded the .h model then saved it in savedModel (format.pb), so even I retrained the model and saved it in .pb format, but when I tried to test it I got following error:

the model attached

saved_model.zip

thanks

[Cilouche]

in python it loads normal (.h and .pb format) but in c# no

import tensorflow as tf

#import tensorflow.python.framework.convert_variables_to_constants_v2
from tensorflow.python.keras.models import load_model
from tensorflow.keras.models import Model
model = tf.keras.models.load_model('mod/InceptionResNet/')
print(model.summary())

[AsakusaRinne]

Hi, the file seems to be SavedModel format saved prior to TF 2.5. The loading itself is supported by using tf.saved_model.load. However it seems that your model has other parts such as restoring optimizer so it will still fail. I'll finish the lacked features in the next 2 or 3 days.

[Cilouche]

Hi, I used another savedModel conversion script, it solved my problem, attached the link so that others can benefit
https://ksingh7.medium.com/part-iii-convert-keras-model-to-tensorflow-frozen-graph-model-a6aa6b1aaeee

Thanks

[AsakusaRinne]

Hey, is the model converted by this script can be directly loaded by tf.net?

[Cilouche]

Yes

[AsakusaRinne]

Yes

Thanks for your sharing. That's a very good work around. I'll still fix the bug of loading your uploaded model in this issue.

[AsakusaRinne]

hello, I followed your suggestions and loaded the .h model then saved it in savedModel (format.pb), so even I retrained the model and saved it in .pb format, but when I tried to test it I got following error:

the model attached

saved_model.zip

thanks

#1032 supports this way by adding an API tf.saved_model.load. However both tensorflow and tf.net will throw an exception that no variable file is found. Did you remove this part in the uploaded file?

[Cilouche]

No, I didn't delete anything, as indicated above, I opted for a format.h5 conversion to frozen model.pb with a well-detailed tutorial.

[AsakusaRinne]

Did you succeed in tensorflow (python)? I tried loading the file with tensorflow python but failed, which tells me it cannot find variable files.