Problem running examples (bad results / python errors) about nnom HOT 2 CLOSED

Hello,
I am using tensorflow 2.3.1, but I was indeed setting Keras to channels_first in keras.json from previous experiments with CMSIS-NN (forgot about that one!). After fixing this (and applying the cudNN fix), both scripts ran with the expected results.

For information, with multiple GPUs I have to use this cudNN fix (loop over GPUs) or I get the error below:

for device in physical_devices:
        tf.config.experimental.set_memory_growth(device, True)

ValueError: Memory growth cannot differ between GPU devices

Also for informaiton, the included weights.h failed at compile-time, from what I assume is an inconsistency with the recent interface changes. Here is the full compile log:
compile_log.txt

Thank you very much for the support, I will be moving on to trying it on my own model

from nnom.

Hi @Mjonir ,
May I know the TF version you are using? recommened TF2.2+. For the crash, are you setting Keras to 'CHW' or 'channel first'? The script currently only support the default HWC shape format.

I am checking both attachement you provided.

Would you use the example weights.h in the example folder other that the one you converted from you model to test? see if the prediction is correct? for the mnist-simple example.

Thanks,
Jianjia

from nnom.