Merge pull request #133 from bark-simulator/experiment_runner_package

Place ExperimentRunner in bark_ml and introduce max. velocity

bark_ml/BUILD

@@ -28,7 +28,7 @@ test_suite(
     "//bark_ml/tests:py_library_tfa_tests",
     "//bark_ml/tests:py_bark_behavior_model_tests",
     "//bark_ml/tests:py_tracer_tests",
-    "//experiments:py_experiment_tests"
+    "//bark_ml/experiment:py_experiment_tests"
   ]
 )
 

bark_ml/evaluators/reward_shaping.py

@@ -52,6 +52,10 @@ def __init__(self,
       self._params["ML"]["RewardShapingEvaluator"]["MaxSteps",
         "Maximum steps per episode.",
         60]
+    self._max_vel = \
+      self._params["ML"]["RewardShapingEvaluator"]["MaxVelocity",
+        "Maximum velocity in episode.",
+        10]
     self._active_shaping_functions = \
       self._params["ML"]["RewardShapingEvaluator"]["RewardShapingPotentials",
         "Reward shaping functions.", {
@@ -147,9 +151,10 @@ def _evaluate(self, observed_world, eval_results, action):
     step_count = eval_results["step_count"]
     collision = eval_results["collision"] or eval_results["drivable_area"]
 
-    # TERMINATE WITH NEGATIVE VEL.
+    # TERMINATE WITH NEGATIVE VEL. or if larger than max. vel
     ego_state = observed_world.ego_agent.state
-    ego_vel = ego_state[int(StateDefinition.VEL_POSITION)] < 0.
+    ego_vel = ego_state[int(StateDefinition.VEL_POSITION)] < 0 or \
+      ego_state[int(StateDefinition.VEL_POSITION)] > self._max_vel
 
     reward_shaping_signal = self.RewardShapingFunction(observed_world)
     if success or collision or step_count > self._max_steps or ego_vel:

bark_ml/experiment/BUILD

@@ -0,0 +1,40 @@
+filegroup(
+    name = "data",
+    srcs = glob(["*.json", "**/*.yaml", "**/*.json", "**/*.xodr", "**/*.py"]),
+    visibility = ["//visibility:public"],
+)
+
+py_library(
+    name = "experiment",
+    srcs = ["experiment.py"],
+    data = ["@bark_project//bark/python_wrapper:core.so"],
+    imports = ["../external/bark_project/bark/python_wrapper/"],
+    deps = ["//bark_ml/environments:environments",
+            "//bark_ml/library_wrappers/lib_tf_agents/agents:agents",
+            "//bark_ml/library_wrappers/lib_tf_agents/runners:runners",
+            "//bark_ml/evaluators:evaluators",
+            "//bark_ml/observers:observers"],
+    visibility = ["//visibility:public"],
+)
+
+py_binary(
+    name = "experiment_runner",
+    srcs = ["experiment_runner.py"],
+    data = ["@bark_project//bark/python_wrapper:core.so",
+            "//bark_ml:generate_core"],
+    imports = ["../external/bark_project/bark/python_wrapper/"],
+    deps = [":experiment"],
+    visibility = ["//visibility:public"],
+)
+
+py_test(
+    name = "py_experiment_tests",
+    srcs = ["py_experiment_tests.py"],
+    data = ["@bark_project//bark/python_wrapper:core.so",
+            ":data"],
+    imports = ["../external/bark_project/bark/python_wrapper/",
+               "../python_wrapper/"],
+    deps = [":experiment"],
+    visibility = ["//visibility:public"]
+)
+

bark_ml/experiment/data/highway_gnn.json

@@ -0,0 +1,173 @@
+{
+    "Experiment": {
+        "Observer": {
+            "ModuleName": "GraphObserver",
+            "Config": {}
+        },
+        "Evaluator": {
+            "ModuleName": "RewardShapingEvaluator",
+            "Config": {}
+        },
+        "Runtime": {
+            "ModuleName": "SingleAgentRuntime",
+            "Config": {}
+        },
+        "Runner": {
+            "ModuleName": "SACRunner",
+            "Config": {}
+        },
+        "Agent": {
+            "ModuleName": "BehaviorGraphSACAgent",
+            "Config": {
+                "init_gnn": "init_interaction_network"
+            }
+        },
+        "Blueprint": {
+            "ModuleName": "ContinuousHighwayBlueprint",
+            "Config": {
+                "num_scenarios": 10000,
+                "viewer": true,
+                "mode": "dense"
+            }
+        },
+        "NumEvaluationEpisodes": 10,
+        "NumVisualizationEpisodes": 10
+    },
+    "Visualization": {
+        "Agents": {
+            "Color": {
+                "Other": {
+                    "Lines": [
+                        0.5,
+                        0.5,
+                        0.5
+                    ],
+                    "Face": [
+                        0.7,
+                        0.7,
+                        0.7
+                    ]
+                },
+                "Controlled": {
+                    "Lines": [
+                        0.0,
+                        0.27,
+                        0.58
+                    ],
+                    "Face": [
+                        0.49,
+                        0.63,
+                        0.83
+                    ]
+                },
+                "UseColormapForOtherAgents": false,
+                "IfColormapUseLineColorOthers": true
+            },
+            "Alpha": {
+                "Controlled": 1.0,
+                "Other": 1
+            },
+            "ColorRoute": [
+                0.2,
+                0.2,
+                0.2
+            ],
+            "DrawRoute": false,
+            "DrawAgentId": false,
+            "DrawEvalGoals": true,
+            "EvalGoalColor": [
+                0.49,
+                0.63,
+                0.83
+            ],
+            "DrawHistory": false,
+            "DrawHistoryDrawFace": true
+        }
+    },
+    "ML": {
+        "RewardShapingEvaluator": {
+            "RewardShapingPotentials": {
+                "VelocityPotential" : {
+                    "desired_vel": 20.0, "vel_dev_max": 20.0, "exponent": 0.4, "type": "positive"
+                  }
+            }
+        },
+        "GoalReachedEvaluator": {
+            "MaxSteps": 200
+        },
+        "InteractionNetwork": {
+            "NumMessagePassingLayers": 2,
+            "EmbeddingSize": 80
+        },
+        "BehaviorTFAAgents": {
+            "NumCheckpointsToKeep": 3
+        },
+        "TFARunner": {
+            "EvaluationSteps": 25,
+            "InitialCollectionEpisodes": 50,
+            "CollectionEpisodesPerStep": 1
+        },
+        "BehaviorContinuousML": {
+            "ActionsLowerBound": [
+                -5.0,
+                -0.2
+            ],
+            "ActionsUpperBound": [
+                4.0,
+                0.2
+            ]
+        },
+        "GraphObserver": {
+            "NormalizationEnabled": true,
+            "AgentLimit": 5,
+            "VisibilityRadius": 1500,
+            "SelfLoops": true,
+            "EnabledNodeFeatures": [
+                "x",
+                "y",
+                "theta",
+                "vel"
+            ],
+            "EnabledEdgeFeatures": [
+                "dx",
+                "dy"
+            ]
+        },
+        "BehaviorGraphSACAgent": {
+            "ActorFcLayerParams": [
+                512,
+                512
+            ],
+            "CriticObservationFcLayerParams": null,
+            "CriticActionFcLayerParams": null,
+            "CriticJointFcLayerParams": [
+                512,
+                512
+            ],
+            "ActorLearningRate": 0.0003,
+            "CriticLearningRate": 0.0003,
+            "AlphaLearningRate": 0.0003,
+            "TargetUpdateTau": 0.05,
+            "TargetUpdatePeriod": 3,
+            "Gamma": 0.995,
+            "RewardScaleFactor": 1.0,
+            "AgentName": "gnn_sac_agent",
+            "DebugSummaries": false,
+            "ReplayBufferCapacity": 10000,
+            "ParallelBufferCalls": 1,
+            "BatchSize": 512,
+            "BufferNumSteps": 2,
+            "BufferPrefetch": 3
+        },
+        "SACRunner": {
+            "NumberOfCollections": 250000,
+            "EvaluateEveryNSteps": 2000
+        }
+    },
+    "BehaviorDynamicModel": {
+        "IntegrationTimeDelta": 0.05000000074505806
+    },
+    "World": {
+        "remove_agents_out_of_map": true
+    }
+}

experiments/experiment.py → bark_ml/experiment/experiment.py

@@ -1,4 +1,3 @@
-from bark.runtime.commons.parameters import ParameterServer
 from bark_ml.environments.blueprints import *
 from bark_ml.environments import *
 from bark_ml.behaviors import *
@@ -8,6 +7,7 @@
 from bark_ml.library_wrappers.lib_tf_agents.runners import *
 from bark_ml.core.observers import *
 from bark_ml.core.evaluators import *
+from bark.runtime.commons.parameters import ParameterServer
 
 
 def LoadModule(module_name, dict_items):
@@ -25,6 +25,7 @@ def LoadModule(module_name, dict_items):
     return FrenetObserver(dict_items['params'])
   return eval("{}(**dict_items)".format(module_name))
 
+
 class Experiment:
   """The Experiment-class contains all entities in order to run,
   train, or evaluate an agent."""

experiments/experiment_runner.py → bark_ml/experiment/experiment_runner.py

@@ -3,28 +3,11 @@
 import hashlib
 import logging
 from pathlib import Path
-from absl import app
-from absl import flags
 import tensorflow as tf
 import numpy as np
 from bark.runtime.commons.parameters import ParameterServer
-from experiments.experiment import Experiment
+from bark_ml.experiment.experiment import Experiment
 
-FLAGS = flags.FLAGS
-flags.DEFINE_enum("mode",
-                  "visualize",
-                  ["train", "visualize", "evaluate", "print", "save"],
-                  "Mode the configuration should be executed in.")
-# NOTE: absolute paths are required
-flags.DEFINE_string("exp_json",
-                    "/Users/hart/Development/bark-ml/experiments/configs/highway_gat_small.json",
-                    "Path to the experiment json.")
-flags.DEFINE_string("save_path",
-                    "/Users/hart/Development/bark-ml/experiments/configs/new_exp.json",
-                    "Path to the experiment json.")
-flags.DEFINE_integer("random_seed",
-                     0,
-                     "Random seed to be used.")
 
 class ExperimentRunner:
   """The ExperimentRunner-Class provides an easy-to-use interface to
@@ -133,14 +116,4 @@ def PrintExperiment(self):
     pprint.pprint(self._experiment.params.ConvertToDict())
 
   def SaveExperiment(self, file_path):
-    self._params.Save(file_path)
-
-# run experiment
-def run_experiment(argv):
-  exp_runner = ExperimentRunner(
-    json_file=FLAGS.exp_json,
-    mode=FLAGS.mode,
-    random_seed=FLAGS.random_seed)
-
-if __name__ == '__main__':
-  app.run(run_experiment)
+    self._params.Save(file_path)

bark_ml/experiment/py_experiment_tests.py

@@ -0,0 +1,29 @@
+# Copyright (c) 2020 fortiss GmbH
+#
+# Authors: Patrick Hart, Julian Bernhard, Klemens Esterle, and
+# Tobias Kessler
+#
+# This software is released under the MIT License.
+# https://opensource.org/licenses/MIT
+
+
+import unittest
+import os
+from bark_ml.experiment.experiment import Experiment
+from bark_ml.experiment.experiment_runner import ExperimentRunner
+
+
+class PyExperimentTests(unittest.TestCase):
+  def test_experiment(self):
+    experiment = Experiment(
+      os.path.join(os.path.dirname(__file__),
+      "data/highway_gnn.json"))
+
+  def test_experiment_runner(self):
+    exp_runner = ExperimentRunner(
+      json_file=os.path.join(os.path.dirname(__file__),
+      "data/highway_gnn.json"), mode="evaluate", random_seed=0)
+
+
+if __name__ == '__main__':
+  unittest.main()

experiments/BUILD

@@ -4,39 +4,13 @@ filegroup(
     visibility = ["//visibility:public"],
 )
 
-
-py_library(
-    name = "experiment",
-    srcs = ["experiment.py"],
-    data = ["@bark_project//bark/python_wrapper:core.so"],
-    imports = ["../external/bark_project/bark/python_wrapper/"],
-    deps = ["//bark_ml/environments:environments",
-            "//bark_ml/library_wrappers/lib_tf_agents/agents:agents",
-            "//bark_ml/library_wrappers/lib_tf_agents/runners:runners",
-            "//bark_ml/evaluators:evaluators",
-            "//bark_ml/observers:observers"],
-    visibility = ["//visibility:public"],
-)
-
 py_binary(
-    name = "experiment_runner",
-    srcs = ["experiment_runner.py"],
+    name = "run_experiment",
+    srcs = ["run_experiment.py"],
     data = ["@bark_project//bark/python_wrapper:core.so",
             "//bark_ml:generate_core",
             ":data"],
     imports = ["../external/bark_project/bark/python_wrapper/"],
-    deps = [":experiment"],
+    deps = ["//bark_ml/experiment:experiment_runner"],
     visibility = ["//visibility:public"],
-)
-
-py_test(
-    name = "py_experiment_tests",
-    srcs = ["py_experiment_tests.py"],
-    data = ["@bark_project//bark/python_wrapper:core.so",
-            ":data"],
-    imports = ["../external/bark_project/bark/python_wrapper/",
-               "../python_wrapper/"],
-    deps = [":experiment"],
-    visibility = ["//visibility:public"]
-)
-
+)