more example stuff

.github/workflows/pytest.yaml

@@ -16,7 +16,7 @@ on:
       - development
 
 env:
-  package-name: "ARLBench"
+  package-name: "arlbench"
   test-dir: tests
   extra-requires: "[dev,envpool]"  # "" for no extra_requires
 

examples/configs/base.yaml

@@ -1,10 +1,8 @@
 defaults:
   - _self_
-  - /cluster: local
   - /algorithm: dqn
   - /environment: cc_cartpole
   - search_space: dqn
-  - /experiments: cc_cartpole_dqn
 
 hydra:
   run:

examples/configs/random_search.yaml

@@ -7,11 +7,10 @@ defaults:
 
 hydra:
   sweeper:
-    n_trials: 256
+    n_trials: 16
     search_space: ${search_space}
     sweeper_kwargs:
-      max_parallelization: 0.1 #  hence, only 25 jobs per batch
-      job_array_size_limit: 100
+      max_parallelization: 1 #  run all of it at once
   run:
     dir: results/sobol/${algorithm}_${autorl.env_name}/${autorl.seed}
   sweep:

examples/configs/smac.yaml

@@ -7,7 +7,7 @@ defaults:
 
 hydra:
   sweeper:
-    n_trials: 50
+    n_trials: 16
     budget_variable: autorl.n_total_timesteps
     search_space: ${search_space}
     sweeper_kwargs:

examples/run_arlbench.py

@@ -2,6 +2,8 @@
 
 from __future__ import annotations
 
+import warnings
+warnings.filterwarnings("ignore")
 import csv
 import logging
 import sys
@@ -10,23 +12,17 @@
 import hydra
 import jax
 from arlbench.arlbench import run_arlbench
-from codecarbon import track_emissions
-from omegaconf import DictConfig, OmegaConf
+from omegaconf import DictConfig
 
 
 @hydra.main(version_base=None, config_path="configs", config_name="base")
-@track_emissions(offline=True, country_iso_code="DEU")
 def execute(cfg: DictConfig):
     """Helper function for nice logging and error handling."""
     logging.basicConfig(
         filename="job.log", format="%(asctime)s %(message)s", filemode="w"
     )
     logger = logging.getLogger()
     logger.setLevel(logging.INFO)
-    logger.info("Logging configured")
-    logger.info(f"JAX devices: {jax.devices()}")
-    logger.info(f"JAX device count: {jax.local_device_count()}")
-    logger.info(f"JAX default backend: {jax.default_backend()}")
 
     if cfg.jax_enable_x64:
         logger.info("Enabling x64 support for JAX.")
@@ -40,8 +36,6 @@ def execute(cfg: DictConfig):
 
 def run(cfg: DictConfig, logger: logging.Logger):
     """Console script for arlbench."""
-    logger.info("Starting run with config:")
-    logger.info(str(OmegaConf.to_yaml(cfg)))
 
     # check if file done exists and if so, return
     try:
@@ -52,7 +46,6 @@ def run(cfg: DictConfig, logger: logging.Logger):
             csvreader = csv.reader(pf)
             performance = next(csvreader)
             performance = float(performance[0])
-            logger.info(f"Returning performance {performance}.")
             return performance
     except FileNotFoundError:
         pass

examples/run_reactive_schedule.py

@@ -35,6 +35,9 @@ def run(cfg: DictConfig, logger: logging.Logger):
         grad_norm, _ = statistics["grad_info"]
 
         # If grad norm doesn't change much, spike the learning rate
+        if last_grad_norm is not None:
+            print(i)
+            print(abs(grad_norm - last_grad_norm))
         if last_grad_norm is not None and abs(grad_norm - last_grad_norm) < tolerance:
             last_lr = cfg.hp_config.learning_rate
             cfg.hp_config.learning_rate *= 10
@@ -46,6 +49,7 @@ def run(cfg: DictConfig, logger: logging.Logger):
             cfg.hp_config.learning_rate = last_lr
             spiked = False
             logger.info(f"Resetting learning rate to {cfg.hp_config.learning_rate}")
+        last_grad_norm = grad_norm
     logger.info(f"Training finished with a total reward of {objectives['reward_mean']}")