gr-libs 0.1.4__py3-none-any.whl → 0.1.6.post1__py3-none-any.whl

gr_libs/recognizer/graml/graml_recognizer.py

@@ -103,7 +103,6 @@ class Graml(LearningRecognizer):
 		self.plans_dict[f"{true_goal}_true"] = true_sequence
 
 		with open(embeddings_path + f'/{true_goal}_{percentage}_plans_dict.pkl', 'wb') as plans_file:
-			# TODO erase AGENT_BASED macros
 			to_dump = {}
 			for goal, obss in self.plans_dict.items():
 				if goal == f"{true_goal}_true":
@@ -243,7 +242,7 @@ class GCGraml(Graml, GaAdaptingRecognizer):
 		if num_timesteps != None: kwargs["num_timesteps"] = num_timesteps
 		gc_agent = self.rl_agent_type(**kwargs)
 		gc_agent.learn()
-		self.agents.append(ContextualAgent(problem_name=self.env_prop.name, problem_goal="general", agent=gc_agent)) # TODO change
+		self.agents.append(ContextualAgent(problem_name=self.env_prop.name, problem_goal="general", agent=gc_agent))
 
 	def generate_sequences_library(self, goal: str) -> List[List[Tuple[np.ndarray, np.ndarray]]]:
 		problem_name = self.env_prop.goal_to_problem_str(goal)

gr_libs/recognizer/recognizer.py

@@ -1,6 +1,5 @@
 from abc import ABC, abstractmethod
 from typing import List, Type
-
 from gr_libs.environment.environment import EnvProperty, SUPPORTED_DOMAINS
 from gr_libs.environment.utils.utils import domain_to_env_property
 from gr_libs.ml.base.rl_agent import RLAgent
@@ -18,7 +17,7 @@ class Recognizer(ABC):
 	def inference_phase(self, inf_sequence, true_goal, percentage) -> str:
 		pass
 
-class LearningRecognizer(Recognizer): # TODO add a class diagram with the inheritance of all calsses
+class LearningRecognizer(Recognizer):
 	def __init__(self, *args, **kwargs):
 		super().__init__(*args, **kwargs)
 
@@ -26,7 +25,7 @@ class LearningRecognizer(Recognizer): # TODO add a class diagram with the inheri
 		self.original_train_configs = train_configs
 
 # a recognizer that needs to train agents for every new goal as part of the goal adaptation phase (that's why it needs dynamic train configs)
-class GaAgentTrainerRecognizer(Recognizer): # TODO add a class diagram with the inheritance of all calsses
+class GaAgentTrainerRecognizer(Recognizer):
 	def __init__(self, *args, **kwargs):
 		super().__init__(*args, **kwargs)
 
@@ -37,7 +36,7 @@ class GaAgentTrainerRecognizer(Recognizer): # TODO add a class diagram with the
 	def domain_learning_phase(self, base_goals: List[str], train_configs: List):
 		super().domain_learning_phase(base_goals, train_configs)
 
-class GaAdaptingRecognizer(Recognizer): # TODO add a class diagram with the inheritance of all calsses
+class GaAdaptingRecognizer(Recognizer):
 	def __init__(self, *args, **kwargs):
 		super().__init__(*args, **kwargs)
 

{gr_libs-0.1.4.dist-info → gr_libs-0.1.6.post1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gr_libs
-Version: 0.1.4
+Version: 0.1.6.post1
 Summary: Package with goal recognition frameworks baselines
 Author: Ben Nageris
 Author-email: Matan Shamir <matan.shamir@live.biu.ac.il>, Osher Elhadad <osher.elhadad@live.biu.ac.il>
@@ -17,6 +17,7 @@ Requires-Dist: torchvision
 Requires-Dist: rl_zoo3
 Requires-Dist: stable_baselines3[extra]
 Requires-Dist: sb3_contrib
+Requires-Dist: pytest
 Provides-Extra: minigrid
 Requires-Dist: gr_envs[minigrid]; extra == "minigrid"
 Provides-Extra: highway
@@ -111,6 +112,25 @@ After installing GRLib, you will have access to custom Gym environments, allowin
 
 Tutorials demonstrating basic ODGR scenarios is available in the sub-package `tutorials`. These tutorials walk through the initialization and deployment process, showcasing how different GR algorithms adapt to emerging goals in various Gym environments.
 
+## Working with an initial dataset of trained agents
+gr_libs also includes a library of trained agents for the various supported environments within the package.
+To get the dataset of trained agents, you can run:
+```sh
+python download_dataset.py
+```
+
+An alternative is to use our docker image, which includes the dataset in it.
+You can:
+1. pull the image:
+```sh
+docker pull ghcr.io/MatanShamir1/gr_test_base:latest
+```
+2. run a container:
+```sh
+docker run -it ghcr.io/MatanShamir1/gr_test_base:latest bash
+```
+3. don't forget to install the package from within the container, go back to 'Setup' for that.
+
 ### Method 1: Writing a Custom Script
 
 1. **Create a recognizer**
@@ -118,6 +138,7 @@ Tutorials demonstrating basic ODGR scenarios is available in the sub-package `tu
    Specify the domain name and specific environment for the recognizer, effectively telling it the domain theory - the collection of states and actions in the environment.
 
    ```python
+   import gr_libs.environment # Triggers gym env registration - you must run it!
    recognizer = Graql(
        domain_name="minigrid",
        env_name="MiniGrid-SimpleCrossingS13N4"

{gr_libs-0.1.4.dist-info → gr_libs-0.1.6.post1.dist-info}/RECORD

@@ -1,3 +1,5 @@
+CI/README.md,sha256=CbWNAWrXFFwYq3sWAORhoQIE5busoNyYh_rFWVH1enw,800
+CI/docker_build_context/Dockerfile,sha256=Rk7LYTxOW7VVJcmNa8csZ4BwkunMYIiHX4WVSuMam50,311
 evaluation/analyze_results_cross_alg_cross_domain.py,sha256=s_DDh4rNfRnvQ0PDa2d5411jYOa7CaI1YeB8Dpup7QU,9803
 evaluation/create_minigrid_map_image.py,sha256=jaSW3n3tY222iFUeAMqedBP9cvD88GCzPrQ6_XHv5oQ,1242
 evaluation/file_system.py,sha256=SSYnj8QGFkq-8V_0s7x2MWbD88aFaoFY4Ogc_Pt8m6U,1601
@@ -8,19 +10,20 @@ evaluation/generate_task_specific_statistics_plots.py,sha256=rBsqaMe2irP_Cfo-icw
 evaluation/get_plans_images.py,sha256=BT-bGWuOPUAYpZVDwk7YMRBLdgKaDbNOBjMrtcl1Vjk,2346
 evaluation/increasing_and_decreasing_.py,sha256=fu1hkEjhOQC3jEsjiS7emW_UPRpVFCaae0d0E2MGZqI,2991
 gr_libs/__init__.py,sha256=-uKsQiHIL7yojbDwlTR-I8sj1WX9XT52PoFbPjtUTKo,145
-gr_libs/environment/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-gr_libs/environment/environment.py,sha256=0-3kZJCmLMF9o0NignZaMEwQb94NZQ2gmsOyfjPXKDI,6919
+gr_libs/_version.py,sha256=C8Me-BH17Mqlv65Ba3Tqc5gFEzabp8fxxyIA9C_XdDQ,517
+gr_libs/environment/__init__.py,sha256=HFVGBcufWf8-ahCo6h_s2pFEyvDy59cFg8z908RgdYo,1038
+gr_libs/environment/environment.py,sha256=d6ZbiAQ4H1aLrUFI8sm0BN9DVW3JtzpkodSi_70Z_PY,6780
 gr_libs/environment/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 gr_libs/environment/utils/utils.py,sha256=4yM3s30KjyuEmWR8UuICE5rR03zsLi3tzqNDvBkdPcU,537
 gr_libs/metrics/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-gr_libs/metrics/metrics.py,sha256=IYPL90tuxbTRaNLaFvgzd5SMUwS8gsSANuWZcSVuKkM,8737
+gr_libs/metrics/metrics.py,sha256=4bnvs5suv-QrK9i1NuOzkE_E8uIzS1nlEazNDRXvZGs,8700
 gr_libs/ml/__init__.py,sha256=jrjxYqvSRgWwFWw7XQP9DzOwvmprMZ2umwT7t-DYtDU,233
 gr_libs/ml/agent.py,sha256=DSnK8nRx9SS76fAOZZEEvA68_meLjzm9lfQpMUXmGQU,1957
 gr_libs/ml/consts.py,sha256=mrbZk8n6QoGzLGaKmaxq4QlAsBbk4fhkCgXLuO9jXKw,365
 gr_libs/ml/base/__init__.py,sha256=MfIYhl_UqH8H7YoTCih8wBFA_gpTOUFq8Ph0_Nq0XQk,68
 gr_libs/ml/base/rl_agent.py,sha256=u9rnb-ma9iDM5b_BlwjcTJGSFezIGrxXINw6b-Dbl8s,1598
 gr_libs/ml/neural/__init__.py,sha256=g-0D5oFX8W52To4OR8vO8kDoBLSxAupVqwcQw8XjT5E,180
-gr_libs/ml/neural/deep_rl_learner.py,sha256=_d6LdbMPqN4qJlOI_UqSD7o0yzIa7EjRONdFSVYO_Ag,20677
+gr_libs/ml/neural/deep_rl_learner.py,sha256=b41_b4GVlYqxhjrr1_YMcGdU9iwcMXsf3zH8D2kEucs,20659
 gr_libs/ml/neural/utils/__init__.py,sha256=bJgPfRnmfDQxdnb0OyRGwzgebEc1PnlO7-GpqszPBcc,106
 gr_libs/ml/neural/utils/dictlist.py,sha256=WpHfdWpVZ_T3PcSnOQUC--ro_tsS0dvam2WG3LcsHDw,1039
 gr_libs/ml/neural/utils/penv.py,sha256=R1uW8sePQqvTlJjpAuMx16eDU6TuGAjQF3hTR1QasMo,1862
@@ -42,22 +45,26 @@ gr_libs/ml/utils/format.py,sha256=nu7RzVwn_raG_fqqmnqlJgUjtA0yzKztkB3a5QZnRYo,30
 gr_libs/ml/utils/math.py,sha256=n62zssVOLHnUb4dPofAoFhoLOKl5n_xBzaKQOUQBoNc,440
 gr_libs/ml/utils/other.py,sha256=HKUfeLBbd4DgJxSTs3ya9KQ85Acx4TjycRrtGD9WQ3s,505
 gr_libs/ml/utils/storage.py,sha256=oCdvL_ypCglnSJsyyXzNyV_UJASTfioa3yJhFlFso64,4277
+gr_libs/problems/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+gr_libs/problems/consts.py,sha256=ON7yfKTAKETg7i3okDYuOzEU7KWvynyubl0m7TlU6Hs,38808
 gr_libs/recognizer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-gr_libs/recognizer/recognizer.py,sha256=ysJYOGe5OlERMAeMwclKpwqw2tQvbSvGnLZrq4qP0xk,1895
+gr_libs/recognizer/recognizer.py,sha256=ZrApJVdBQxKRYhhDiWLCNGmlxgi674nwgb30BgVggC8,1705
 gr_libs/recognizer/recognizer_doc.md,sha256=RnTvbZhl2opvU7-QT4pULCV5HCdJTw2dsu8WQOOiR3E,2521
 gr_libs/recognizer/gr_as_rl/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 gr_libs/recognizer/gr_as_rl/gr_as_rl_recognizer.py,sha256=84GdfohC2dZoNH_QEo7GpSt8nZWdfqSRKCTY99X_iME,5215
 gr_libs/recognizer/graml/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 gr_libs/recognizer/graml/gr_dataset.py,sha256=lG6m3ulxFELpH1oURnlcmNDWOrxyuzvlAR28ZTqB7L8,7224
-gr_libs/recognizer/graml/graml_recognizer.py,sha256=OblODjvWSi8KZ5ByvGyL5Mm1GY3IZb8yspfk81Dbebg,15721
+gr_libs/recognizer/graml/graml_recognizer.py,sha256=SGs7rtkA73lbCv9HISa6dfjVUJUhlH54QriVsoGVRss,15672
 gr_libs/recognizer/utils/__init__.py,sha256=ewSroxL7aATvvm-Xzc1_-61mP2LU2U28YaOEqvVVDB0,41
 gr_libs/recognizer/utils/format.py,sha256=e0AnqtPeYoJsV9Z7cEBpgbzTM0hLNxFIjn07fQ3YbQw,492
-tutorials/graml_minigrid_tutorial.py,sha256=iLs7mOYNAZ5wtxtSwiE8tvbLMIueQ5VmVmDnBBjWG_4,1589
-tutorials/graml_panda_tutorial.py,sha256=DuHVDLe49qwgWouLxwalqdT1P4dlNOOMdgDc3ocNX5Y,1820
-tutorials/graml_parking_tutorial.py,sha256=sQ496DNuAo9GZ_0iUZ_6Hqe5zFxIYZ_pBIHQscQvR4o,2501
-tutorials/graml_point_maze_tutorial.py,sha256=TnLT9FdDj6AF8lm0lDIZum4ouPE5rye4RBH8z4Exj2Y,2713
-tutorials/graql_minigrid_tutorial.py,sha256=VoXbEgL_hjQLfau6WohXxPK8rrv1VLA874F8PZ7ZtPk,1421
-gr_libs-0.1.4.dist-info/METADATA,sha256=wH7aEvKh4kRTXBs75uPSBW87s23dgcsKRqgkVSwKDQc,8905
-gr_libs-0.1.4.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
-gr_libs-0.1.4.dist-info/top_level.txt,sha256=k7_l789QSJEr9JrtvsRMxNoTIDwNduq8mhIN-YoPJUM,29
-gr_libs-0.1.4.dist-info/RECORD,,
+tests/test_graml.py,sha256=ZJB2jqtf4Q2-KZredkJq90teqmHBIvigCAQpvR5G110,559
+tests/test_graql.py,sha256=-onMi13e2wStOmB5bYv2f3Ita3QFFiw416XMBkby0OI,141
+tutorials/graml_minigrid_tutorial.py,sha256=ONvxFi79R7d8dcd6gy083Z_yy9A2flhGTDIDRxurdx8,1782
+tutorials/graml_panda_tutorial.py,sha256=wtv_lsw0vsU7j45GKeWecTfE7jzfh4iVGEVnQyaWthM,2063
+tutorials/graml_parking_tutorial.py,sha256=46-sfxmYA9jLRSpqIF9z69MLSfOSTJarfjlQ_Igq294,1769
+tutorials/graml_point_maze_tutorial.py,sha256=mYq3IxYbf9jidq-4VdT3MdStV80Q5lytFv6Xzzn22Ys,1835
+tutorials/graql_minigrid_tutorial.py,sha256=Jb0TCUhiZQkFeafJWUTPnCISd4FKfPrqP-xfHiqCGKE,1635
+gr_libs-0.1.6.post1.dist-info/METADATA,sha256=UPwlwVlbGTpTsUhYwWH5hYr-hSBpcWjrFIA7sWg0Kj4,9620
+gr_libs-0.1.6.post1.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
+gr_libs-0.1.6.post1.dist-info/top_level.txt,sha256=rL-bbK-KnLzVbLIUCdN1riH58lup3jG0NJ3LTt_qSwo,38
+gr_libs-0.1.6.post1.dist-info/RECORD,,

{gr_libs-0.1.4.dist-info → gr_libs-0.1.6.post1.dist-info}/top_level.txt

@@ -1,3 +1,5 @@
+CI
 evaluation
 gr_libs
+tests
 tutorials

tests/test_graml.py

@@ -0,0 +1,16 @@
+from tutorials.graml_minigrid_tutorial import run_graml_minigrid_tutorial
+from tutorials.graml_panda_tutorial import run_graml_panda_tutorial
+from tutorials.graml_parking_tutorial import run_graml_parking_tutorial
+from tutorials.graml_point_maze_tutorial import run_graml_point_maze_tutorial
+
+def test_graml_minigrid_tutorial():
+	run_graml_minigrid_tutorial()
+
+def test_graml_panda_tutorial():
+	run_graml_panda_tutorial()
+
+def test_graml_parking_tutorial():
+	run_graml_parking_tutorial()
+
+def test_graml_point_maze_tutorial():
+	run_graml_point_maze_tutorial()

tests/test_graql.py

@@ -0,0 +1,4 @@
+from tutorials.graql_minigrid_tutorial import run_graql_minigrid_tutorial
+
+def test_graql_minigrid_tutorial():
+	run_graql_minigrid_tutorial()

tutorials/graml_minigrid_tutorial.py

@@ -1,30 +1,34 @@
-from gr_libs.environment.environment import QLEARNING
+from gr_libs.environment.environment import MINIGRID, QLEARNING
 from gr_libs.metrics.metrics import stochastic_amplified_selection
 from gr_libs.ml.tabular.tabular_q_learner import TabularQLearner
 from gr_libs.ml.utils.format import random_subset_with_order
 from gr_libs import ExpertBasedGraml
 
-recognizer = ExpertBasedGraml(
-	domain_name="minigrid",
-	env_name="MiniGrid-SimpleCrossingS13N4"
-)
+def run_graml_minigrid_tutorial():
+    recognizer = ExpertBasedGraml(
+        domain_name=MINIGRID,
+        env_name="MiniGrid-SimpleCrossingS13N4"
+    )
 
-recognizer.domain_learning_phase(base_goals=[(11,1), (11,11), (1,11), (7,11), (8,1), (10,6), (6,9), (11,3), (11,5)],
-                                 train_configs=[(QLEARNING, 100000) for _ in range(9)])
+    recognizer.domain_learning_phase(base_goals=[(11,1), (11,11), (1,11), (7,11), (8,1), (10,6), (6,9), (11,3), (11,5)],
+                                    train_configs=[(QLEARNING, 100000) for _ in range(9)])
 
-recognizer.goals_adaptation_phase(
-    dynamic_goals = [(11,1), (11,11), (1,11)],
-    dynamic_train_configs=[(QLEARNING, 100000) for _ in range(3)] # for expert sequence generation.
-)
-# TD3 is different from recognizer and expert algorithms, which are SAC #
-actor = TabularQLearner(domain_name="minigrid", problem_name="MiniGrid-SimpleCrossingS13N4-DynamicGoal-11x1-v0", algorithm=QLEARNING, num_timesteps=100000)
-actor.learn()
-# sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
-full_sequence = actor.generate_observation(
-    action_selection_method=stochastic_amplified_selection,
-    random_optimalism=True, # the noise that's added to the actions
-)
+    recognizer.goals_adaptation_phase(
+        dynamic_goals = [(11,1), (11,11), (1,11)],
+        dynamic_train_configs=[(QLEARNING, 100000) for _ in range(3)] # for expert sequence generation.
+    )
+    # TD3 is different from recognizer and expert algorithms, which are SAC #
+    actor = TabularQLearner(domain_name="minigrid", problem_name="MiniGrid-SimpleCrossingS13N4-DynamicGoal-11x1-v0", algorithm=QLEARNING, num_timesteps=100000)
+    actor.learn()
+    # sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
+    full_sequence = actor.generate_observation(
+        action_selection_method=stochastic_amplified_selection,
+        random_optimalism=True, # the noise that's added to the actions
+    )
 
-partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)), is_consecutive=False)
-closest_goal = recognizer.inference_phase(partial_sequence, (11,1), 0.5)
-print(f"closest_goal returned by GRAML: {closest_goal}\nactual goal actor aimed towards: (11, 1)")
+    partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)), is_consecutive=False)
+    closest_goal = recognizer.inference_phase(partial_sequence, (11,1), 0.5)
+    print(f"closest_goal returned by GRAML: {closest_goal}\nactual goal actor aimed towards: (11, 1)")
+
+if __name__ == "__main__":
+    run_graml_minigrid_tutorial()

tutorials/graml_panda_tutorial.py

@@ -1,32 +1,41 @@
 
 import numpy as np
 from stable_baselines3 import PPO, SAC
-from gr_libs.environment.environment import PANDA, GCEnvProperty, PandaProperty
+import gr_libs.environment.environment
+from gr_libs.environment.environment import PANDA, EnvProperty, GCEnvProperty, PandaProperty
 from gr_libs.environment.utils.utils import domain_to_env_property
 from gr_libs.metrics.metrics import stochastic_amplified_selection
 from gr_libs.ml.neural.deep_rl_learner import DeepRLAgent, GCDeepRLAgent
 from gr_libs.ml.utils.format import random_subset_with_order
 from gr_libs import GCGraml
 
-recognizer = GCGraml( # TODO make these tutorials into pytests
-    domain_name=PANDA,
-	env_name="PandaMyReachDense"
-)
-recognizer.domain_learning_phase(base_goals=[np.array([PandaProperty.sample_goal()]) for _ in range(1,30)],
-					             train_configs=[(SAC, 800000)])
-recognizer.goals_adaptation_phase(dynamic_goals=[np.array([[-0.1, -0.1, 0.1]]), np.array([[-0.1, 0.1, 0.1]]), np.array([[0.2, 0.2, 0.1]])])
-# TD3 is different from recognizer and expert algorithms, which are SAC #
-property_type = domain_to_env_property(PANDA)
-env_property = property_type("PandaMyReachDense")
-problem_name = env_property.goal_to_problem_str(np.array([[-0.1, -0.1, 0.1]]))
-actor = DeepRLAgent(domain_name=PANDA, problem_name=problem_name, algorithm=PPO, num_timesteps=400000)
-actor.learn()
-# sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
-full_sequence = actor.generate_observation(
-    action_selection_method=stochastic_amplified_selection,
-    random_optimalism=True, # the noise that's added to the actions
-)
+def run_graml_panda_tutorial():
+    recognizer = GCGraml( # TODO make these tutorials into pytests
+        domain_name=PANDA,
+        env_name="PandaMyReachDense"
+    )
+    recognizer.domain_learning_phase(
+        base_goals=[np.array([PandaProperty.sample_goal()]) for _ in range(1,30)],
+        train_configs=[(SAC, 800000)]
+    )
+    recognizer.goals_adaptation_phase(
+        dynamic_goals=[np.array([[-0.1, -0.1, 0.1]]), np.array([[-0.1, 0.1, 0.1]]), np.array([[0.2, 0.2, 0.1]])]
+    )
+    # TD3 is different from recognizer and expert algorithms, which are SAC #
+    property_type = domain_to_env_property(PANDA)
+    env_property = property_type("PandaMyReachDense")
+    problem_name = env_property.goal_to_problem_str(np.array([[-0.1, -0.1, 0.1]]))
+    actor = DeepRLAgent(domain_name=PANDA, problem_name=problem_name, algorithm=PPO, num_timesteps=400000)
+    actor.learn()
+    # sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
+    full_sequence = actor.generate_observation(
+        action_selection_method=stochastic_amplified_selection,
+        random_optimalism=True, # the noise that's added to the actions
+    )
 
-partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)), is_consecutive=False)
-closest_goal = recognizer.inference_phase(partial_sequence, np.array([[-0.1, -0.1, 0.1]]), 0.5)
-print(f"closest_goal returned by GRAML: {closest_goal}\nactual goal actor aimed towards: [-0.1, -0.1, 0.1]")
+    partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)), is_consecutive=False)
+    closest_goal = recognizer.inference_phase(partial_sequence, np.array([[-0.1, -0.1, 0.1]]), 0.5)
+    print(f"closest_goal returned by GRAML: {closest_goal}\nactual goal actor aimed towards: [-0.1, -0.1, 0.1]")
+
+if __name__ == "__main__":
+    run_graml_panda_tutorial()

tutorials/graml_parking_tutorial.py

@@ -1,38 +1,38 @@
 
 from stable_baselines3 import PPO, SAC, TD3
-from gr_libs.environment.environment import EnvProperty, GCEnvProperty, ParkingProperty
+from gr_libs.environment.environment import PARKING, EnvProperty, GCEnvProperty, ParkingProperty
 from gr_libs.metrics.metrics import stochastic_amplified_selection
 from gr_libs.ml.neural.deep_rl_learner import DeepRLAgent, GCDeepRLAgent
 from gr_libs.ml.utils.format import random_subset_with_order
 from gr_libs.recognizer.graml.graml_recognizer import ExpertBasedGraml, GCGraml
 
-# Consider extracting all these to "default point_maze (or every other domain) variables" module which would simplify things like the problem_list_to_str_tuple function, sizes of inputs, etc.
-recognizer = GCGraml(
-    env_name="parking", # TODO change to macros which are importable from some info or env module of enums.
-    problems=[ParkingProperty("parking-v0")],
-    train_configs=[(PPO, 400000)],
-	gc_goal_set=[f"Parking-S-14-PC--GI-{i}-v0" for i in range(1,21)]
-)
-recognizer.domain_learning_phase()
-recognizer.goals_adaptation_phase(
-    dynamic_goals_problems = [ParkingProperty(p) for p in ["Parking-S-14-PC--GI-1-v0",
-                              "Parking-S-14-PC--GI-4-v0",
-                              "Parking-S-14-PC--GI-8-v0",
-                              "Parking-S-14-PC--GI-11-v0",
-                              "Parking-S-14-PC--GI-14-v0",
-                              "Parking-S-14-PC--GI-18-v0",
-                              "Parking-S-14-PC--GI-21-v0"]] # TODO detach the goal from the environment instance in every gym env, add the ability to alter it from outside.
-    #dynamic_train_configs=[(SAC, 400000) for _ in range(7)] # for expert sequence generation. TODO change to require this only if sequence generation method is EXPERT.
-)
-# TD3 is different from recognizer and expert algorithms, which are SAC #
-actor = DeepRLAgent(env_name="parking", problem_name="Parking-S-14-PC--GI-8-v0", algorithm=TD3, num_timesteps=400000)
-actor.learn()
-# sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
-full_sequence = actor.generate_observation(
-    action_selection_method=stochastic_amplified_selection,
-    random_optimalism=True, # the noise that's added to the actions
-)
+def run_graml_parking_tutorial():
+    recognizer = GCGraml(
+        domain_name=PARKING,
+        env_name="Parking-S-14-PC-"
+    )
+                        
+    recognizer.domain_learning_phase(
+        [i for i in range(1,21)],
+        [(PPO, 200000)]
+    )
+    recognizer.goals_adaptation_phase(
+        dynamic_goals = ["1", "11", "21"]
+        # no need for expert sequence generation since GCRL is used
+    )
 
-partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)), is_consecutive=False)
-closest_goal = recognizer.inference_phase(partial_sequence, ParkingProperty("Parking-S-14-PC--GI-8-v0").str_to_goal(), 0.5)
-print(f"closest_goal returned by GRAML: {closest_goal}\nactual goal actor aimed towards: 8")
+    # TD3 is different from recognizer and expert algorithms, which are SAC #
+    actor = DeepRLAgent(domain_name="parking", problem_name="Parking-S-14-PC--GI-11-v0", algorithm=TD3, num_timesteps=400000)
+    actor.learn()
+    # sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
+    full_sequence = actor.generate_observation(
+        action_selection_method=stochastic_amplified_selection,
+        random_optimalism=True, # the noise that's added to the actions
+    )
+
+    partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)), is_consecutive=False)
+    closest_goal = recognizer.inference_phase(partial_sequence, ParkingProperty("Parking-S-14-PC--GI-11-v0").str_to_goal(), 0.5)
+    print(f"closest_goal returned by GRAML: {closest_goal}\nactual goal actor aimed towards: 11")
+
+if __name__ == "__main__":
+    run_graml_parking_tutorial()

tutorials/graml_point_maze_tutorial.py

@@ -1,43 +1,39 @@
 
 from stable_baselines3 import SAC, TD3
-from gr_libs.environment.utils.format import maze_str_to_goal
+from gr_libs.environment.environment import POINT_MAZE, PointMazeProperty
 from gr_libs.metrics.metrics import stochastic_amplified_selection
 from gr_libs.ml.neural.deep_rl_learner import DeepRLAgent
 from gr_libs.ml.utils.format import random_subset_with_order
 from gr_libs.recognizer.graml.graml_recognizer import ExpertBasedGraml
 
-# Consider extracting all these to "default point_maze (or every other domain) variables" module which would simplify things like the problem_list_to_str_tuple function, sizes of inputs, etc.
-recognizer = ExpertBasedGraml(
-    env_name="point_maze", # TODO change to macros which are importable from some info or env module of enums.
-    problems=[("PointMaze-FourRoomsEnvDense-11x11-Goal-9x1"),
-              ("PointMaze-FourRoomsEnv-11x11-Goal-9x9"), # this one doesn't work with dense rewards because of encountering local minima
-              ("PointMaze-FourRoomsEnvDense-11x11-Goal-1x9"),
-              ("PointMaze-FourRoomsEnvDense-11x11-Goal-3x3"),
-              ("PointMaze-FourRoomsEnvDense-11x11-Goal-3x4"),
-              ("PointMaze-FourRoomsEnvDense-11x11-Goal-8x2"),
-              ("PointMaze-FourRoomsEnvDense-11x11-Goal-3x7"),
-              ("PointMaze-FourRoomsEnvDense-11x11-Goal-2x8")],
-    task_str_to_goal=maze_str_to_goal,
-    method=DeepRLAgent,
-    collect_statistics=False,
-    train_configs=[(SAC, 200000) for _ in range(8)],
-)
-recognizer.domain_learning_phase()
-recognizer.goals_adaptation_phase(
-    dynamic_goals_problems = ["PointMaze-FourRoomsEnvDense-11x11-Goal-4x4",
-                              "PointMaze-FourRoomsEnvDense-11x11-Goal-7x3",
-                              "PointMaze-FourRoomsEnvDense-11x11-Goal-3x7"],
-    dynamic_train_configs=[(SAC, 200000) for _ in range(3)] # for expert sequence generation. TODO change to require this only if sequence generation method is EXPERT.
-)
-# TD3 is different from recognizer and expert algorithms, which are SAC #
-actor = DeepRLAgent(env_name="point_maze", problem_name="PointMaze-FourRoomsEnvDense-11x11-Goal-4x4", algorithm=TD3, num_timesteps=200000)
-actor.learn()
-# sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
-full_sequence = actor.generate_observation(
-    action_selection_method=stochastic_amplified_selection,
-    random_optimalism=True, # the noise that's added to the actions
-)
+def run_graml_point_maze_tutorial():
+    recognizer = ExpertBasedGraml(
+        domain_name=POINT_MAZE,
+        env_name="PointMaze-FourRoomsEnvDense-11x11"
+    )
 
-partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)))
-closest_goal = recognizer.inference_phase(partial_sequence, maze_str_to_goal("PointMaze-FourRoomsEnvDense-11x11-Goal-4x4"), 0.5)
-print(f"closest_goal returned by GRAML: {closest_goal}\nactual goal actor aimed towards: (4, 4)")
+    recognizer.domain_learning_phase(
+        [(9,1), (9,9), (1,9), (3,3), (3,4), (8,2), (3,7), (2,8)],
+        [(SAC, 200000) for _ in range(8)]
+    )
+
+    recognizer.goals_adaptation_phase(
+        dynamic_goals = [(4,4), (7,3), (3,7)],
+        dynamic_train_configs=[(SAC, 200000) for _ in range(3)] # for expert sequence generation.
+    )
+
+    # TD3 is different from recognizer and expert algorithms, which are SAC #
+    actor = DeepRLAgent(domain_name="point_maze", problem_name="PointMaze-FourRoomsEnvDense-11x11-Goal-4x4", algorithm=TD3, num_timesteps=200000)
+    actor.learn()
+    # sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
+    full_sequence = actor.generate_observation(
+        action_selection_method=stochastic_amplified_selection,
+        random_optimalism=True, # the noise that's added to the actions
+    )
+
+    partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)))
+    closest_goal = recognizer.inference_phase(partial_sequence, PointMazeProperty("PointMaze-FourRoomsEnvDense-11x11-Goal-4x4").str_to_goal(), 0.5)
+    print(f"closest_goal returned by GRAML: {closest_goal}\nactual goal actor aimed towards: (4, 4)")
+
+if __name__ == "__main__":
+    run_graml_point_maze_tutorial()

tutorials/graql_minigrid_tutorial.py

@@ -4,26 +4,31 @@ from gr_libs.ml.tabular.tabular_q_learner import TabularQLearner
 from gr_libs.ml.utils.format import random_subset_with_order
 from gr_libs import Graql
 
-recognizer = Graql(
-	domain_name="minigrid",
-	env_name="MiniGrid-SimpleCrossingS13N4"
-)
+def run_graql_minigrid_tutorial():
+    recognizer = Graql(
+        domain_name="minigrid",
+        env_name="MiniGrid-SimpleCrossingS13N4"
+    )
 
-#Graql doesn't have a domain learning phase, so we skip it
+    #Graql doesn't have a domain learning phase, so we skip it
 
-recognizer.goals_adaptation_phase(
-    dynamic_goals = [(11,1), (11,11), (1,11)],
-    dynamic_train_configs=[(QLEARNING, 100000) for _ in range(3)] # for expert sequence generation.
-)
-# TD3 is different from recognizer and expert algorithms, which are SAC #
-actor = TabularQLearner(domain_name="minigrid", problem_name="MiniGrid-SimpleCrossingS13N4-DynamicGoal-11x1-v0", algorithm=QLEARNING, num_timesteps=100000)
-actor.learn()
-# sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
-full_sequence = actor.generate_observation(
-    action_selection_method=stochastic_amplified_selection,
-    random_optimalism=True, # the noise that's added to the actions
-)
+    recognizer.goals_adaptation_phase(
+        dynamic_goals = [(11,1), (11,11), (1,11)],
+        dynamic_train_configs=[(QLEARNING, 100000) for _ in range(3)] # for expert sequence generation.
+    )
+    # TD3 is different from recognizer and expert algorithms, which are SAC #
+    actor = TabularQLearner(domain_name="minigrid", problem_name="MiniGrid-SimpleCrossingS13N4-DynamicGoal-11x1-v0", algorithm=QLEARNING, num_timesteps=100000)
+    actor.learn()
+    # sample is generated stochastically to simulate suboptimal behavior, noise is added to the actions values #
+    full_sequence = actor.generate_observation(
+        action_selection_method=stochastic_amplified_selection,
+        random_optimalism=True, # the noise that's added to the actions
+    )
 
-partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)), is_consecutive=False)
-closest_goal = recognizer.inference_phase(partial_sequence, (11,1), 0.5)
-print(f"closest_goal returned by Graql: {closest_goal}\nactual goal actor aimed towards: (11, 1)")
+    partial_sequence = random_subset_with_order(full_sequence, (int)(0.5 * len(full_sequence)), is_consecutive=False)
+    closest_goal = recognizer.inference_phase(partial_sequence, (11,1), 0.5)
+    print(f"closest_goal returned by Graql: {closest_goal}\nactual goal actor aimed towards: (11, 1)")
+    return closest_goal, (11,1)
+
+if __name__ == "__main__":
+    run_graql_minigrid_tutorial()