search

majianjia / nnom

A higher-level Neural Network library for microcontrollers.
Apache License 2.0
909 stars 245 forks source link

'KerasTensor' object has no attribute 'input' #137

Open alimynn opened 3 years ago

alimynn commented 3 years ago

I have a dense network with a InputLayer as the first layer followed by Dense layers and ReLU layers. At the end, there is a Activations layer with sigmoid activation function as can be seen in my screenshot.

image

I have tensorflow version 2.5.0 installed and when executing generate_model, i get the error "'KerasTensor' object has no attribute 'input'" in file nnom.py in line 827. Is this an error with the tensorflow version, a bug in the library or did i define anything wrong?

Thanks

P.S.: This is the stack trace and the model definition. image

image

majianjia commented 3 years ago

Hi @alimynn Please try to use layers.Input instead of layers.InputLayer and see if it helps. Other seems fine from me. If it doesn't help, please could you let me know your TF version? I recommend using the version above 2.1. I normally test with 2.2 and 2.3 Thanks,

alimynn commented 3 years ago

Hi and thank you for your fast answer. I have TF version 2.5 installed first and tested now also with version 2.2. With 2.2 the error comes in line 825 as you can see in the screenshot below.

image

Also, changing from layers.InputLayer to layers.Input makes no difference. What i found now is that when i change the model definition like in the next screenshot, everything works fine.

image

majianjia commented 3 years ago

Thanks, It looks like there is an unsupported layer with the sequential model.