Help Running Custom Python Functions in "Eager-like" Mode

[arshad171]

Hi,

I am new HeteroCL, but I am familiar with TensorFlow, PyTorch, and the graph execution paradigm (from earlier versions of TensorFlow).

I was trying to compute the backward reachability tube for one of the reinforcement learning applications I was working on. In the dynamics.opt_ctrl function, expression the new_state in terms of state and inptuts is quite straight forward if the relations can be broken down into addition, multiplication operations.

However, I would like to run a model prediction to generate the new state, i.e. new_state = my_ml_model.predict(state, inputs), which requires access to the values (numpy array) during the graph execution (analogous to eager execution in TensorFlow).

Is it possible to fetch the numpy array during the graph execution so that I run a ML prediction and return back a hcl.Tensor?

I would really appreciate your help!

Thanks,
Arshad