thiagopbueno / rddl2tf Goto Github PK

Hi

I am interested in using RDDLGym in PyTorch. It seems this codebase is tightly coupled with TF 1. Where do you recommend I start in order to build a framework independent RDDLGym such that it can be used like any other RDDLGym environment and analogous to any DL library?

Thanks in advance!

In my fork, I have successfully adjusted both rddl2tf and rddgym for TensorFlow 2, which is more easily available and maintained these days. Unfortunate is that we don't have a backward compatibility to TF1.

What would be the best offering we can make?
a) merge into the master, increasing a major/minor version.
b) proving a separate fork with a different name, say, rddl2tf_tf2 and rddlgym_tf2.
c) just leave it in a branch.

What do you say?

Extend base Compiler class to handle determinization schemes when compiling random variables. Refactor utility functions.

Define all required compilation methods and properties in a base class. Allow method overriding and reuse via **kwargs.

Extend base Compiler class to handle re-parameterized distributions. Refactor utility functions.

After freshly installing rddl2tf in a new conda environment, I am unable to run the Programmatic mode example in the README.

I get the following error

Traceback (most recent call last): File "testtf.py", line 8, in <module> compiler = Compiler(model) TypeError: Can't instantiate abstract class Compiler with abstract methods _compile_aggregation_expression, _compile_arithmetic_expression, _compile_boolean_expression, _compile_constant_expression, _compile_control_flow_expression, _compile_function_expression, _compile_pvariable_expression, _compile_random_variable_expression, _compile_relational_expression

Extend base Compiler class to handle log-likelihood of distributions. Refactor utility functions.

Define object-oriented low-level rddl2tf API.