Does HLS4ML support Tensorflow (without Keras) in Conv2D/Conv1D??

[IanCheng18]

Does HLS4ML support Tensorflow (without Keras) in Conv2D/Conv1D? I didn't see any converters for TensorFlow instead, but there is a lot for Keras.

I gave up using Keras since I am using lambda layer, or is there any good workaround?

Many thanksssssss

[thesps]

Hi, yes there is a TensorFlow converter. However you will find that a lambda operation is not supported - hls4ml cannot generate HLS from arbitrary functions. I'd recommend you try to implement what you need from only the supported layers (try print(hls4ml.converters.get_supported_keras_layers())). If you still need to do something very custom, the process to support a custom layer through hls4ml conversion is to:

• Write a custom layer in Keras
• Implement the HLS function that performs the same operation. Take inspiration from the existing nnet_utils implementations. Isolate things that can be configured into a struct of configuration settings.
• Implement a class to represent its attributes that inherits from hls4ml Layer, add it to the layer map
• Register it along with its HLS template (for example here)
• Add a conversion handler for it to extract the relevant attributes (for example here)