{ lib , fetchPypi , buildPythonPackage , frozendict , termcolor , matplotlib , gym , pandas , scipy , absl-py , dm-env , plotnine , scikitimage , dm-tree , patsy , tensorflow-probability , dm-haiku , statsmodels , mizani , trfl , optax , pytestCheckHook , dm-sonnet , rlax , distrax }: let bsuite = buildPythonPackage rec { pname = "bsuite"; version = "0.3.5"; src = fetchPypi { inherit pname version; hash = "sha256-ak9McvXl7Nz5toUaPaRaJek9lurxiQiIW209GnZEjX0="; }; buildInputs = [ absl-py dm-env dm-tree frozendict gym matplotlib mizani pandas patsy plotnine scikitimage scipy statsmodels termcolor ]; nativeCheckInputs = [ distrax dm-haiku dm-sonnet optax pytestCheckHook rlax tensorflow-probability trfl ]; pythonImportsCheck = [ "bsuite" ]; disabledTests = [ # Tests require network connection "test_run9" "test_longer_action_sequence" "test_reset" "test_step_after_reset" "test_step_on_fresh_environment" "test_longer_action_sequence" "test_reset" "test_step_after_reset" "test_step_on_fresh_environment" "test_longer_action_sequence" "test_reset" "test_step_after_reset" "test_step_on_fresh_environment" "test_logger" "test_episode_truncation" ]; # escape infinite recursion with rlax doCheck = false; passthru.tests = { check = bsuite.overridePythonAttrs (_: { doCheck = true; }); }; meta = with lib; { description = '' Core RL Behaviour Suite. A collection of reinforcement learning experiments ''; homepage = "https://github.com/deepmind/bsuite"; license = licenses.asl20; maintainers = with maintainers; [ onny ]; }; }; in bsuite