Merge pull request #20 from rdnfn/dev/general

Add support for sinergym and other fixes

.devcontainer/Dockerfile

@@ -14,14 +14,6 @@ RUN pip3 --disable-pip-version-check --no-cache-dir install -r /tmp/pip-tmp/requ
     -r /tmp/pip-tmp/boptest_requirements.txt \
     && rm -rf /tmp/pip-tmp
 
-# Install beobench problem set libraries
-ENV MAIN_PYTHONPATH = "${PYTHONPATH}"
-ENV PYTHONPATH "${PYTHONPATH}:/tmp"
-COPY beobench/installer.py beobench/constants.py /tmp/beobench/
-RUN python /tmp/beobench/installer.py \
-    && rm -rf /tmp/beobench
-ENV PYTHONPATH = "${MAIN_PYTHONPATH}"
-
 # Install nvidia container runtime
 RUN wget -O tmp/nvidia-container-toolkit.deb https://nvidia.github.io/libnvidia-container/stable/debian10/amd64/nvidia-container-toolkit_1.7.0-1_amd64.deb \
     && wget -O tmp/nvidia-container-tools.deb https://nvidia.github.io/libnvidia-container/stable/debian10/amd64/libnvidia-container-tools_1.7.0-1_amd64.deb \

.devcontainer/devcontainer.json

@@ -17,12 +17,13 @@
 	"workspaceFolder": "/workspace",
 	// "workspaceMount": "source=remote-workspace,target=/workspace,type=volume",
 	// ADAPT: the mount must be adapted to cloned repo location
-	"workspaceMount": "source=/home/rdnfn-docker/main/repos/github/beobench/,target=/workspace,type=bind,consistency=cached",
+	"workspaceMount": "source=/home/rdnfn-docker/main/repos/github/beobench/,target=/workspace/beobench/,type=bind,consistency=cached",
 	// ADAPT: the mount must be adapted to the gitconfig location on the remote machine
 	"mounts": [
-		"source=/home/rdnfn-docker/.gitconfig,target=/.gitconfig,type=bind,consistency=cached"
+		"source=/home/rdnfn-docker/.gitconfig,target=/root/.gitconfig,type=bind,consistency=cached",
+		"source=/home/rdnfn-docker/main/repos/github/beobench_contrib,target=/workspace/beobench_contrib/,type=bind,consistency=cached"
 	],
-	"postCreateCommand": "pwd && pip install -e .",
+	"postCreateCommand": "cd ./beobench && pip install -e .",
 	// Set *default* container specific settings.json values on container create.
 	"settings": {
 		"python.defaultInterpreterPath": "/usr/local/bin/python",

CONTRIBUTING.rst

@@ -117,12 +117,26 @@ To run a subset of tests::
 Deploying
 ---------
 
-A reminder for the maintainers on how to deploy.
-Make sure all your changes are committed (including an entry in HISTORY.rst).
-Then run::
+A reminder for the maintainers on how to deploy. Follow this checklist (inspired by `this checklist <https://gist.github.com/audreyfeldroy/5990987>`_ and `this packaging tutorial <https://packaging.python.org/en/latest/tutorials/packaging-projects/>`_):
 
-$ bump2version patch # possible: major / minor / patch
-$ git push
-$ git push --tags
+1. Update ``HISTORY.rst`` and commit with message like "aux: add changelog for upcoming release 0.1.0"
+2. Run
 
-Travis will then deploy to PyPI if tests pass.
+    .. code-block:: console
+
+        bump2version patch # possible: major / minor / patch
+
+3. Push commits
+4. Build the package
+
+    .. code-block:: console
+
+        python3 -m build
+
+5. Upload the package (this will ask for password/token from pypi.org)
+
+    .. code-block:: console
+
+        twine upload dist/*
+
+6. Add release tag on github

HISTORY.rst

@@ -2,6 +2,13 @@
 History
 =======
 
+0.2.1 (2022-02-03)
+------------------
+
+* Add integration with sinergym
+* Move gym integrations to separate beobench_contrib repo
+* Make usage of GPUs in containers optional
+
 0.2.0 (2022-01-18)
 ------------------
 

PYPI_README.rst

@@ -23,12 +23,12 @@ Features
 
 *Some of the features are work in progress*
 
-**Main features**
+Main features
 
 - *RL algorithm collection:* what's the best RL method for your BEO problem? Building on `Ray RLlib <https://github.com/ray-project/ray/tree/master/rllib>`_, beobench provides a large collection of pre-configured RL algorithm experiments that can be easily applied to your new BEO problem.
 - *Problem collection:* beobench provides ready-to-use docker containers for popular BEO gym-type problem libraries. By enforcing a strict OpenAI ``gym.Env`` it makes testing your method on different libraries easy.
 
-**Additional features**
+Additional features
 
 - *Experiment logging:* log experiment results in a reproducible and shareable manner via `Weights and Biases`_.
 - *Hyperparameter tuning:* easily tune hyperparameters using the extensive `Ray Tune Search API <https://docs.ray.io/en/master/tune/index.html>`_.

README.rst

@@ -31,12 +31,12 @@ Features
 
 *Some of the features are work in progress*
 
-**Main features**
+Main features
 
 - *RL algorithm collection:* what's the best RL method for your BEO problem? Building on `Ray RLlib <https://github.com/ray-project/ray/tree/master/rllib>`_, beobench provides a large collection of pre-configured RL algorithm experiments that can be easily applied to your new BEO problem.
 - *Problem collection:* beobench provides ready-to-use docker containers for popular BEO gym-type problem libraries. By enforcing a strict OpenAI ``gym.Env`` it makes testing your method on different libraries easy.
 
-**Additional features**
+Additional features
 
 - *Experiment logging:* log experiment results in a reproducible and shareable manner via `Weights and Biases`_.
 - *Hyperparameter tuning:* easily tune hyperparameters using the extensive `Ray Tune Search API <https://docs.ray.io/en/master/tune/index.html>`_.
@@ -61,7 +61,7 @@ Run your first beobench experiment in three steps:
         .. code-block:: console
 
                 pip install beobench
-                
+
 3. Finally, start your first experiment using:
 
         .. code-block:: console

beobench/__init__.py

@@ -2,6 +2,6 @@
 
 __author__ = """rdnfn"""
 __email__ = "-"
-__version__ = "0.2.0"
+__version__ = "0.2.1"
 
 from beobench.utils import restart

beobench/experiment/containers.py

@@ -21,14 +21,14 @@ def build_experiment_container(
     if use_no_cache:
         flags.append("--no-cache")
 
-    AVAILABLE_INTEGRATIONS = ["boptest"]  # pylint: disable=invalid-name
+    AVAILABLE_INTEGRATIONS = ["boptest","sinergym"]  # pylint: disable=invalid-name
 
     if build_context in AVAILABLE_INTEGRATIONS:
         docker_tag = f"beobench_{build_context}:latest"
         integration_name = build_context
         build_context = (
             f"https://github.com/rdnfn/"
-            f"beobench.git#master:beobench/integrations/{build_context}"
+            f"beobench_contrib.git#main:gyms/{build_context}"
         )
         print(
             (

beobench/experiment/definitions/default.py

@@ -56,7 +56,6 @@
             "framework": "torch",
             "log_level": "WARNING",
             "num_workers": 8,  # 1 for silent mode, can at least be 6
-            "num_gpus": 1,
             "seed": ray.tune.randint(0, 10000000),
         },
         "log_to_file": True,

beobench/experiment/definitions/experiment_003_sinergym.py

@@ -0,0 +1,42 @@
+"""An experiment to test sinergym integration."""
+
+import ray.tune
+from beobench.experiment.definitions.default import problem, method, rllib_setup
+
+
+problem = {
+    "name": "sinergym_test_problem",
+    "description": ("Control problem corresponding to " "created by sinergym env ''."),
+    "problem_library": "sinergym",
+    "rllib_experiment_config": {
+        "config": {
+            "env": "sinergym-Eplus-5Zone-hot-continuous-v1",
+            "env_config": {
+                "name": "Eplus-5Zone-hot-continuous-v1",
+                "normalize": True,
+            },
+            "gamma": 0.999,
+            "output": "/root/ray_results/debug/",
+            "output_compress_columns": [],
+            "soft_horizon": False,
+            "no_done_at_end": False,
+            "horizon": 10000,
+        },
+        "stop": {"timesteps_total": 400000},
+    },
+}
+
+rllib_setup = {
+    "rllib_experiment_config": {
+        "config": {
+            # Utilities settings
+            "framework": "torch",
+            "log_level": "WARNING",
+            "num_workers": 8,  # 1 for silent mode, can at least be 6
+            "num_gpus": 1,
+            "seed": ray.tune.randint(0, 10000000),
+        },
+        "log_to_file": True,
+        "checkpoint_freq": 10000,
+    },
+}

beobench/experiment/scheduler.py

@@ -21,6 +21,8 @@ def run(
     wandb_project: str = "",
     wandb_entity: str = "",
     wandb_api_key: str = "",
+    use_gpu: bool = False,
+    docker_shm_size: str = "2gb",
     no_additional_container: bool = False,
     use_no_cache: bool = False,
 ) -> None:
@@ -39,6 +41,10 @@ def run(
             "initial_experiments".
         wandb_entity (str, optional): Name of wandb entity. Defaults to "beobench".
         wandb_api_key (str, optional): wandb API key. Defaults to None.
+        use_gpu (bool, optional): whether to use GPU from the host system. Defaults to
+            False.
+        docker_shm_size(str, optional): size of the shared memory available to the
+            container. Defaults to '2gb'."
         no_additional_container (bool, optional): wether not to start another container
             to run experiments in. Defaults to False, which means that another container
             is started to run experiments in.
@@ -59,6 +65,16 @@ def run(
         experiment_def = importlib.util.module_from_spec(spec)
         spec.loader.exec_module(experiment_def)
 
+        # check if all attributes present and replace with default if not
+        if not hasattr(experiment_def, "problem"):
+            experiment_def.problem = beobench.experiment.definitions.default.problem
+        if not hasattr(experiment_def, "method"):
+            experiment_def.method = beobench.experiment.definitions.default.method
+        if not hasattr(experiment_def, "rllib_setup"):
+            experiment_def.rllib_setup = (
+                beobench.experiment.definitions.default.rllib_setup
+            )
+
     if no_additional_container:
 
         # Add wandb callback if sufficient information
@@ -71,11 +87,17 @@ def run(
             )
         else:
             callbacks = []
+
+        # change RLlib setup if GPU used
+        rllib_setup = experiment_def.rllib_setup
+        if use_gpu:
+            rllib_setup["rllib_experiment_config"]["config"]["num_gpus"] = 1
+
         # run experiment in ray tune
         run_in_tune(
             problem_def=experiment_def.problem,
             method_def=experiment_def.method,
-            rllib_setup=experiment_def.rllib_setup,
+            rllib_setup=rllib_setup,
             rllib_callbacks=callbacks,
         )
     else:
@@ -91,7 +113,6 @@ def run(
             build_context=experiment_def.problem["problem_library"],
             use_no_cache=use_no_cache,
         )
-        beobench.experiment.containers.create_docker_network("beobench-net")
 
         # define docker arguments/options/flags
         unique_id = uuid.uuid4().hex[:6]
@@ -106,6 +127,28 @@ def run(
                 f"{exp_file_abs}:{exp_file_on_docker}:ro",
             ]
 
+        # enable docker-from-docker access only for built-in boptest integration.
+        if experiment_def.problem["problem_library"] == "boptest":
+
+            # Create docker network (only useful if starting other containers)
+            beobench.experiment.containers.create_docker_network("beobench-net")
+
+            docker_flags += [
+                # enable access to docker-from-docker
+                "-v",
+                "/var/run/docker.sock:/var/run/docker.sock",
+                # network allows access to BOPTEST API in other containers
+                "--network",
+                "beobench-net",
+            ]
+
+        # enabling GPU access in docker container
+        if use_gpu:
+            docker_flags += [
+                # add all available GPUs
+                "--gpus=all",
+            ]
+
         # define flags for beobench scheduler call inside experiment container
         beobench_flags = []
         if experiment_file:
@@ -114,6 +157,8 @@ def run(
             beobench_flags.append(f"--wandb-project={wandb_project}")
         if wandb_entity:
             beobench_flags.append(f"--wandb-entity={wandb_entity}")
+        if use_gpu:
+            beobench_flags.append("--use-gpu")
         beobench_flag_str = " ".join(beobench_flags)
 
         # if no wandb API key is given try to get it from env
@@ -124,21 +169,13 @@ def run(
         args = [
             "docker",
             "run",
-            # the mounted socket allows access to docker
-            "-v",
-            "/var/run/docker.sock:/var/run/docker.sock",
             # mount experiment data dir
             "-v",
             f"{local_dir_abs}:/root/ray_results",
             # automatically remove container when stopped/exited
             "--rm",
-            # network allows access to BOPTEST API in other containers
-            "--network",
-            "beobench-net",
             # add more memory
-            "--shm-size=20.48gb",
-            # add available GPUs,
-            "--gpus=all",
+            f"--shm-size={docker_shm_size}",
             "--name",
             container_name,
             *docker_flags,
@@ -183,6 +220,16 @@ def run(
     default="",
     help="Weights and biases API key.",
 )
+@click.option(
+    "--use-gpu",
+    is_flag=True,
+    help="Whether to use GPUs from the host system in experiment container.",
+)
+@click.option(
+    "--docker-shm-size",
+    default="2gb",
+    help="Size of shared memory available to experiment container.",
+)
 @click.option(
     "--no-additional-container",
     is_flag=True,
@@ -199,6 +246,8 @@ def run_command(
     wandb_project: str,
     wandb_entity: str,
     wandb_api_key: str,
+    use_gpu: bool,
+    docker_shm_size: str,
     no_additional_container: bool,
     use_no_cache: bool,
 ) -> None:
@@ -216,6 +265,8 @@ def run_command(
         wandb_project=wandb_project,
         wandb_entity=wandb_entity,
         wandb_api_key=wandb_api_key,
+        use_gpu=use_gpu,
+        docker_shm_size=docker_shm_size,
         no_additional_container=no_additional_container,
         use_no_cache=use_no_cache,
     )