[Config] New default config and test config

Signed-off-by: Matteo Bettini <[email protected]>

benchmarl/conf/experiment/base_experiment.yaml

@@ -4,30 +4,32 @@ defaults:
 
 sampling_device: "cpu"
 train_device: "cpu"
+
 gamma: 0.99
 polyak_tau: 0.005
 lr: 0.00005
 n_optimizer_steps: 45
-collected_frames_per_batch: 60_000
-n_envs_per_worker: 600
-n_iters: 500
-prefer_continuous_actions: True
 clip_grad_norm: True
-clip_grad_val: 40
+clip_grad_val: 5
+prefer_continuous_actions: True
+
+collected_frames_per_batch: 6000
+n_envs_per_worker: 10
+n_iters: 500
 
-on_policy_minibatch_size: 4096
+on_policy_minibatch_size: 400
 
 off_policy_memory_size: 100_000
 off_policy_train_batch_size: 10_000
 off_policy_prioritised_alpha:  0.7
 off_policy_prioritised_beta: 0.5
 
 evaluation: True
-evaluation_interval: 30
-evaluation_episodes: 200
+evaluation_interval: 20
+evaluation_episodes: 10
 
 loggers: [wandb]
 create_json: True
 
 restore_file: null
-checkpoint_interval: 100
+checkpoint_interval: 50

test/conftest.py

@@ -0,0 +1,20 @@
+import pytest
+
+from benchmarl.experiment import ExperimentConfig
+
+
+@pytest.fixture
+def experiment_config() -> ExperimentConfig:
+    experiment_config: ExperimentConfig = ExperimentConfig.get_from_yaml()
+    experiment_config.n_iters = 3
+    experiment_config.n_optimizer_steps = 2
+    experiment_config.n_envs_per_worker = 2
+    experiment_config.collected_frames_per_batch = 100
+    experiment_config.on_policy_minibatch_size = 10
+    experiment_config.off_policy_memory_size = 200
+    experiment_config.off_policy_train_batch_size = 100
+    experiment_config.evaluation = False
+    experiment_config.loggers = []
+    experiment_config.create_json = False
+    experiment_config.checkpoint_interval = 0
+    return experiment_config

test/test_algorithms.py

@@ -1,18 +1,22 @@
-import pathlib
+import importlib
 
 import pytest
-
 from benchmarl.algorithms import algorithm_config_registry
+
 from benchmarl.environments import VmasTask
-from benchmarl.experiment import Experiment, ExperimentConfig
+from benchmarl.experiment import Experiment
 from benchmarl.models.common import SequenceModelConfig
 from benchmarl.models.mlp import MlpConfig
 from torch import nn
 
 
+_has_vmas = importlib.util.find_spec("vmas") is not None
+
+
+@pytest.mark.skipif(not _has_vmas, reason="VMAS not found")
 @pytest.mark.parametrize("algo_config", algorithm_config_registry.values())
 @pytest.mark.parametrize("continuous", [True, False])
-def test_all_algos_balance(algo_config, continuous):
+def test_all_algos_vmas(algo_config, continuous, experiment_config):
     task = VmasTask.BALANCE.get_from_yaml()
     model_config = SequenceModelConfig(
         model_configs=[
@@ -21,14 +25,6 @@ def test_all_algos_balance(algo_config, continuous):
         ],
         intermediate_sizes=[5],
     )
-    experiment_config: ExperimentConfig = ExperimentConfig.get_from_yaml(
-        str(
-            pathlib.Path(__file__).parent
-            / "conf"
-            / "experiment"
-            / "base_experiment.yaml"
-        )
-    )
     experiment_config.prefer_continuous_actions = continuous
 
     experiment = Experiment(