Using a custom TF op in Keras

[jia...@google.com]

Hi everyone,

I'm just getting started with Keras and considering porting a network written in vanilla Tensorflow. My network contains a custom TF op written in CUDA. What's the recommended way of wrapping it inside a Keras layer? I Googled around but wasn't able to find my guidance. I'm hoping to have the Keras wrapper simply call the corresponding tf function but am unsure how to get started.

Thanks!

[isaac.grisw...@gmail.com]

I didn't dig into it much. But I did see this,

https://stackoverflow.com/questions/45067108/tensorflow-op-in-keras-model

[Jiawen Chen]

It turned out to be pretty straightforward. In call(), the input is a tf.Tensor so I can just directly call my op on it.

If the op has multiple inputs, then the input is a list of Tensors.

--
You received this message because you are subscribed to a topic in the Google Groups "Keras-users" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/keras-users/5T6WonvYSrE/unsubscribe.
To unsubscribe from this group and all its topics, send an email to keras-users...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/ef634662-8d9c-4be9-8f05-e92dfdf8c809%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.