adeptRL
adeptRL copied to clipboard
heronsystems
Reinforcement learning framework to accelerate research
Adept uses global gradient norm clipping of 0.5 by default which could prevent learning (just slowing it down) if the gradients are high and the clipping occurs at every training...
Still need to run a test env with a reward implementation to verify it is learning correctly
Example: ``` python cpu_ops = [CustomOpFn(name_filters=[dict_key])] cpu_preprocessor = ObsPreprocessor( cpu_ops, Space.from_gym(observation_space), Space.dtypes_from_gym(observation_space), ) ``` I expect that my CustomOpFn should only be receiving the name_filters specified when calling `update_shape`, `update_dtype`...
Hi, as shown in the following full dependency graph of **_adeptrl_**, **_adeptrl_** requires _**cloudpickle**_ (>=0.5)， while the installed version of **_gym_**(0.17.1) requires _**cloudpickle>=1.2.0,
In particular, see here: https://github.com/heronsystems/adeptRL/blob/2d8b0d3be9d0530f0df55d7f4b919177f8767388/adept/network/modular_network.py#L216-L238 In the following if-statement: ``` if dim in head_submodules: continue ``` `type(dim) == int`; however, keys of `head_submodules` are guaranteed to be `str` due to...
For example, looking at `SubModule2D`: ``` def output_shape(self, dim=None): if dim == 1: f, l = self._output_shape return (f * l,) ... ... def _to_1d(self, submodule_output): """ :param submodule_output: torch.Tensor...
don't use modular network directory name when using a custom network
Metadata
Owner
Metadata
Reinforcement learning framework to accelerate research