Merge branch 'master' into release_9_branch_merge

.github/ISSUE_TEMPLATE/bug_report.md

 - Unity Version: [e.g. Unity 2020.1f1]
 - OS + version: [e.g. Windows 10]
 - _ML-Agents version_: (e.g. ML-Agents v0.8, or latest `develop` branch from source)
-- _TensorFlow version_: (you can run `pip3 show tensorflow` to get this)
+- _Torch version_: (you can run `pip3 show torch` to get this)
 - _Environment_: (which example environment you used to reproduce the error)
 
 **NOTE:** We are unable to help reproduce bugs with custom environments.  Please attempt to reproduce your issue with one of the example environments, or provide a minimal patch to one of the environments needed to reproduce the issue.

.yamato/gym-interface-test.yml

-test_editors:
-  - version: 2019.4
+{% metadata_file .yamato/test_versions.metafile %}
 ---
 {% for editor in test_editors %}
 test_gym_interface_{{ editor.version }}:

.yamato/python-ll-api-test.yml

-test_editors:
-  - version: 2019.4
+{% metadata_file .yamato/test_versions.metafile %}
 ---
 {% for editor in test_editors %}
 test_mac_ll_api_{{ editor.version }}:

.yamato/standalone-build-test.yml

-test_editors:
-  - version: 2018.4
-  - version: 2019.3
+{% metadata_file .yamato/test_versions.metafile %}
 ---
 {% for editor in test_editors %}
 test_mac_standalone_{{ editor.version }}:
     UNITY_VERSION: {{ editor.version }}
   commands:
     - pip install pyyaml --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple
+    # TODO remove the "--user" command and the path prefix when we can migrate away from the custom bokken image
+    - python -m pip install unity-downloader-cli --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple --upgrade --user
+    - /Users/bokken/Library/Python/3.7/bin/unity-downloader-cli -u {{ editor.version }} -c editor --wait --fast
+    - python -u -m ml-agents.tests.yamato.standalone_build_tests --scene=Assets/ML-Agents/TestScenes/TestCompressedGrid/TestGridCompressed.unity
+    - python -u -m ml-agents.tests.yamato.standalone_build_tests --scene=Assets/ML-Agents/TestScenes/TestCompressedTexture/TestTextureCompressed.unity
   triggers:
     cancel_old_ci: true
     expression: |

.yamato/training-int-tests.yml

-test_editors:
-  - version: 2018.4
-  - version: 2019.4
-  - version: 2020.1
+{% metadata_file .yamato/test_versions.metafile %}
 ---
 {% for editor in test_editors %}
 test_mac_training_int_{{ editor.version }}:
     UNITY_VERSION: {{ editor.version }}
   commands:
     - pip install pyyaml --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple
+    # TODO remove the "--user" command and the path prefix when we can migrate away from the custom bokken image
+    - python -m pip install unity-downloader-cli --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple --upgrade --user
+    - /Users/bokken/Library/Python/3.7/bin/unity-downloader-cli -u {{ editor.version }} -c editor --wait --fast
-    - python -u -m ml-agents.tests.yamato.training_int_tests --csharp=1.0.0
+    - python -u -m ml-agents.tests.yamato.training_int_tests --csharp={{ editor.csharp_backcompat_version }}
   dependencies:
     - .yamato/standalone-build-test.yml#test_mac_standalone_{{ editor.version }}
   triggers:

Dockerfile

-# Based off of python:3.6-slim, except that we are using ubuntu instead of debian.
-FROM ubuntu:16.04
-
-
-# ensure local python is preferred over distribution python
-ENV PATH /usr/local/bin:$PATH
+FROM nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04
-# http://bugs.python.org/issue19846
-# > At the moment, setting "LANG=C" on a Linux system *fundamentally breaks Python 3*, and that's not OK.
-ENV LANG C.UTF-8
+RUN yes | unminimize
-# runtime dependencies
-RUN apt-get update && apt-get install -y --no-install-recommends \
-		ca-certificates \
-		libexpat1 \
-		libffi6 \
-		libgdbm3 \
-		libreadline6 \
-		libsqlite3-0 \
-		libssl1.0.0 \
-	&& rm -rf /var/lib/apt/lists/*
+RUN echo "deb http://packages.cloud.google.com/apt cloud-sdk-xenial main" | tee -a /etc/apt/sources.list.d/google-cloud-sdk.list
+RUN wget https://packages.cloud.google.com/apt/doc/apt-key.gpg && apt-key add apt-key.gpg
+RUN apt-get update && \
+  apt-get install -y --no-install-recommends wget curl tmux vim git gdebi-core \
+  build-essential python3-pip unzip google-cloud-sdk htop mesa-utils xorg-dev xorg \
+  libglvnd-dev libgl1-mesa-dev libegl1-mesa-dev libgles2-mesa-dev && \
+  wget http://security.ubuntu.com/ubuntu/pool/main/libx/libxfont/libxfont1_1.5.1-1ubuntu0.16.04.4_amd64.deb && \
+  wget http://security.ubuntu.com/ubuntu/pool/universe/x/xorg-server/xvfb_1.18.4-0ubuntu0.10_amd64.deb && \
+  yes | gdebi libxfont1_1.5.1-1ubuntu0.16.04.4_amd64.deb && \
+  yes | gdebi xvfb_1.18.4-0ubuntu0.10_amd64.deb
+RUN python3 -m pip install --upgrade pip
+RUN pip install setuptools==41.0.0
-ENV GPG_KEY 0D96DF4D4110E5C43FBFB17F2D347EA6AA65421D
-ENV PYTHON_VERSION 3.6.4
+ENV LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH
-RUN set -ex \
-	&& buildDeps=" \
-		dpkg-dev \
-		gcc \
-		libbz2-dev \
-		libc6-dev \
-		libexpat1-dev \
-		libffi-dev \
-		libgdbm-dev \
-		liblzma-dev \
-		libncursesw5-dev \
-		libreadline-dev \
-		libsqlite3-dev \
-		libssl-dev \
-		make \
-		tcl-dev \
-		tk-dev \
-		wget \
-		xz-utils \
-		zlib1g-dev \
-# as of Stretch, "gpg" is no longer included by default
-		$(command -v gpg > /dev/null || echo 'gnupg dirmngr') \
-	" \
-	&& apt-get update && apt-get install -y $buildDeps --no-install-recommends && rm -rf /var/lib/apt/lists/* \
-	\
-	&& wget -O python.tar.xz "https://www.python.org/ftp/python/${PYTHON_VERSION%%[a-z]*}/Python-$PYTHON_VERSION.tar.xz" \
-	&& wget -O python.tar.xz.asc "https://www.python.org/ftp/python/${PYTHON_VERSION%%[a-z]*}/Python-$PYTHON_VERSION.tar.xz.asc" \
-	&& export GNUPGHOME="$(mktemp -d)" \
-	&& gpg --keyserver ha.pool.sks-keyservers.net --recv-keys "$GPG_KEY" \
-	&& gpg --batch --verify python.tar.xz.asc python.tar.xz \
-	&& rm -rf "$GNUPGHOME" python.tar.xz.asc \
-	&& mkdir -p /usr/src/python \
-	&& tar -xJC /usr/src/python --strip-components=1 -f python.tar.xz \
-	&& rm python.tar.xz \
-	\
-	&& cd /usr/src/python \
-	&& gnuArch="$(dpkg-architecture --query DEB_BUILD_GNU_TYPE)" \
-	&& ./configure \
-		--build="$gnuArch" \
-		--enable-loadable-sqlite-extensions \
-		--enable-shared \
-		--with-system-expat \
-		--with-system-ffi \
-		--without-ensurepip \
-	&& make -j "$(nproc)" \
-	&& make install \
-	&& ldconfig \
-	\
-	&& apt-get purge -y --auto-remove $buildDeps \
-	\
-	&& find /usr/local -depth \
-		\( \
-			\( -type d -a \( -name test -o -name tests \) \) \
-			-o \
-			\( -type f -a \( -name '*.pyc' -o -name '*.pyo' \) \) \
-		\) -exec rm -rf '{}' + \
-	&& rm -rf /usr/src/python
-
-# make some useful symlinks that are expected to exist
-RUN cd /usr/local/bin \
-	&& ln -s idle3 idle \
-	&& ln -s pydoc3 pydoc \
-	&& ln -s python3 python \
-	&& ln -s python3-config python-config
-
-# if this is called "PIP_VERSION", pip explodes with "ValueError: invalid truth value '<VERSION>'"
-ENV PYTHON_PIP_VERSION 9.0.3
-
-RUN set -ex; \
-	\
-	apt-get update; \
-	apt-get install -y --no-install-recommends wget; \
-	rm -rf /var/lib/apt/lists/*; \
-	\
-	wget -O get-pip.py 'https://bootstrap.pypa.io/get-pip.py'; \
-	\
-	apt-get purge -y --auto-remove wget; \
-	\
-	python get-pip.py \
-		--disable-pip-version-check \
-		--no-cache-dir \
-		"pip==$PYTHON_PIP_VERSION" \
-	; \
-	pip --version; \
-	\
-	find /usr/local -depth \
-		\( \
-			\( -type d -a \( -name test -o -name tests \) \) \
-			-o \
-			\( -type f -a \( -name '*.pyc' -o -name '*.pyo' \) \) \
-		\) -exec rm -rf '{}' +; \
-	rm -f get-pip.py
-
-
-RUN apt-get update && apt-get -y upgrade
-
-# xvfb is used to do CPU based rendering of Unity
-RUN apt-get install -y xvfb
-
-# Install ml-agents-envs package locally
-COPY ml-agents-envs /ml-agents-envs
-WORKDIR /ml-agents-envs
-RUN pip install -e .
-
-# Install ml-agents package next
-COPY ml-agents /ml-agents
+#checkout ml-agents for SHA
+RUN mkdir /ml-agents
-RUN pip install -e .
-
-# Port 5004 is the port used in Editor training.
-# Environments will start from port 5005, 
-# so allow enough ports for several environments.
-EXPOSE 5004-5050
-
-ENTRYPOINT ["xvfb-run", "--auto-servernum", "--server-args='-screen 0 640x480x24'", "mlagents-learn"]
+ARG SHA
+RUN git init
+RUN git remote add origin https://github.com/Unity-Technologies/ml-agents.git
+RUN git fetch --depth 1 origin $SHA
+RUN git checkout FETCH_HEAD
+RUN pip install -e /ml-agents/ml-agents-envs
+RUN pip install -e /ml-agents/ml-agents

Project/Assets/ML-Agents/Examples/SharedAssets/Scripts/ModelOverrider.cs

             var bp = m_Agent.GetComponent<BehaviorParameters>();
             var behaviorName = bp.BehaviorName;
 
-            var nnModel = GetModelForBehaviorName(behaviorName);
+            NNModel nnModel = null;
+            try
+            {
+                nnModel = GetModelForBehaviorName(behaviorName);
+            }
+            catch (Exception e)
+            {
+                overrideError = $"Exception calling GetModelForBehaviorName: {e}";
+            }
+
-                overrideError =
-                    $"Didn't find a model for behaviorName {behaviorName}. Make " +
-                    $"sure the behaviorName is set correctly in the commandline " +
-                    $"and that the model file exists";
+                if (string.IsNullOrEmpty(overrideError))
+                {
+                    overrideError =
+                        $"Didn't find a model for behaviorName {behaviorName}. Make " +
+                        "sure the behaviorName is set correctly in the commandline " +
+                        "and that the model file exists";
+                }
             }
             else
             {

README.md

 
 **The Unity Machine Learning Agents Toolkit** (ML-Agents) is an open-source
 project that enables games and simulations to serve as environments for
-training intelligent agents. Agents can be trained using reinforcement learning,
-imitation learning, neuroevolution, or other machine learning methods through a
-simple-to-use Python API. We also provide implementations (based on TensorFlow)
+training intelligent agents. We provide implementations (based on PyTorch)
-train intelligent agents for 2D, 3D and VR/AR games. These trained agents can be
+train intelligent agents for 2D, 3D and VR/AR games. Researchers can also use the
+provided simple-to-use Python API to train Agents using reinforcement learning,
+imitation learning, neuroevolution, or any other methods. These trained agents can be
 used for multiple purposes, including controlling NPC behavior (in a variety of
 settings such as multi-agent and adversarial), automated testing of game builds
 and evaluating different game design decisions pre-release. The ML-Agents

com.unity.ml-agents.extensions/Runtime/Sensors/GridSensor.cs

         [Tooltip("Whether to show gizmos or not")]
         public bool ShowGizmos = false;
 
+        public SensorCompressionType CompressionType = SensorCompressionType.PNG;
+
         /// <summary>
         /// Array of colors displaying the DebugColors for each cell in OnDrawGizmos. Only updated if ShowGizmos.
         /// </summary>
         /// <inheritdoc/>
         public virtual SensorCompressionType GetCompressionType()
         {
-            return SensorCompressionType.PNG;
+            return CompressionType;
         }
 
         /// <summary>

com.unity.ml-agents/CHANGELOG.md

 [Semantic Versioning](http://semver.org/spec/v2.0.0.html).
 
 
-## [Unreleased] - 2020-11-04
+## [Unreleased]
+
+ - PyTorch trainers are now the default. See the
+ [installation docs](https://github.com/Unity-Technologies/ml-agents/blob/master/docs/Installation.md) for
+ more information on installing PyTorch. For the time being, TensorFlow is still available;
+ you can use the TensorFlow backend by adding `--tensorflow` to the CLI, or
+ adding `framework: tensorflow` in the configuration YAML. (#4517)
+- The Barracuda dependency was upgraded to 1.1.2 (#4571)
+- The `action_probs` node is no longer listed as an output in TensorFlow models (#4613).
+- `Agent.CollectObservations()` and `Agent.EndEpisode()` will now throw an exception
+if they are called recursively (for example, if they call `Agent.EndEpisode()`).
+Previously, this would result in an infinite loop and cause the editor to hang. (#4573)
- - Fixed an issue where runs could not be resumed when using TensorFlow and Ghost Training. (#4593)
+- Fixed an issue where runs could not be resumed when using TensorFlow and Ghost Training. (#4593)
+
 
 ## [1.5.0-preview] - 2020-10-14
 ### Major Changes

com.unity.ml-agents/Runtime/Academy.cs

         // Flag used to keep track of the first time the Academy is reset.
         bool m_HadFirstReset;
 
-        // Whether the Academy is in the middle of a step. This is used to detect and Academy
-        // step called by user code that is also called by the Academy.
-        bool m_IsStepping;
+        // Detect an Academy step called by user code that is also called by the Academy.
+        private RecursionChecker m_StepRecursionChecker = new RecursionChecker("EnvironmentStep");
 
         // Random seed used for inference.
         int m_InferenceSeed;
         /// </summary>
         public void EnvironmentStep()
         {
-            // Check whether we're already in the middle of a step.
-            // This shouldn't happen generally, but could happen if user code (e.g. CollectObservations)
-            // that is called by EnvironmentStep() also calls EnvironmentStep(). This would result
-            // in an infinite loop and/or stack overflow, so stop it before it happens.
-            if (m_IsStepping)
-            {
-                throw new UnityAgentsException(
-                    "Academy.EnvironmentStep() called recursively. " +
-                    "This might happen if you call EnvironmentStep() from custom code such as " +
-                    "CollectObservations() or OnActionReceived()."
-                );
-            }
-
-            m_IsStepping = true;
-
-            try
+            using (m_StepRecursionChecker.Start())
             {
                 if (!m_HadFirstReset)
                 {
                 {
                     AgentAct?.Invoke();
                 }
-            }
-            finally
-            {
-                // Reset m_IsStepping when we're done (or if an exception occurred).
-                m_IsStepping = false;
             }
         }
 

com.unity.ml-agents/Runtime/Agent.cs

         /// </summary>
         internal VectorSensor collectObservationsSensor;
 
+        private RecursionChecker m_CollectObservationsChecker = new RecursionChecker("CollectObservations");
+        private RecursionChecker m_OnEpisodeBeginChecker = new RecursionChecker("OnEpisodeBegin");
+
         /// <summary>
         /// List of IActuators that this Agent will delegate actions to if any exist.
         /// </summary>
             // episode when initializing until after the Academy had its first reset.
             if (Academy.Instance.TotalStepCount != 0)
             {
-                OnEpisodeBegin();
+                using (m_OnEpisodeBeginChecker.Start())
+                {
+                    OnEpisodeBegin();
+                }
             }
         }
 
             {
                 // Make sure the latest observations are being passed to training.
                 collectObservationsSensor.Reset();
-                CollectObservations(collectObservationsSensor);
+                using (m_CollectObservationsChecker.Start())
+                {
+                    CollectObservations(collectObservationsSensor);
+                }
             }
             // Request the last decision with no callbacks
             // We request a decision so Python knows the Agent is done immediately
             UpdateSensors();
             using (TimerStack.Instance.Scoped("CollectObservations"))
             {
-                CollectObservations(collectObservationsSensor);
+                using (m_CollectObservationsChecker.Start())
+                {
+                    CollectObservations(collectObservationsSensor);
+                }
             }
             using (TimerStack.Instance.Scoped("CollectDiscreteActionMasks"))
             {
         {
             ResetData();
             m_StepCount = 0;
-            OnEpisodeBegin();
+            using (m_OnEpisodeBeginChecker.Start())
+            {
+                OnEpisodeBegin();
+            }
+
         }
 
         /// <summary>

com.unity.ml-agents/Runtime/Sensors/CameraSensor.cs

             using (TimerStack.Instance.Scoped("CameraSensor.WriteToTensor"))
             {
                 var texture = ObservationToTexture(m_Camera, m_Width, m_Height);
-                var numWritten = Utilities.TextureToTensorProxy(texture, writer, m_Grayscale);
+                var numWritten = writer.WriteTexture(texture, m_Grayscale);
                 DestroyTexture(texture);
                 return numWritten;
             }

com.unity.ml-agents/Runtime/Sensors/ObservationWriter.cs

             }
         }
     }
+
+    public static class ObservationWriterExtension
+    {
+
+        /// <summary>
+        /// Writes a Texture2D into a ObservationWriter.
+        /// </summary>
+        /// <param name="obsWriter">
+        /// Writer to fill with Texture data.
+        /// </param>
+        /// <param name="texture">
+        /// The texture to be put into the tensor.
+        /// </param>
+        /// <param name="grayScale">
+        /// If set to <c>true</c> the textures will be converted to grayscale before
+        /// being stored in the tensor.
+        /// </param>
+        /// <returns>The number of floats written</returns>
+        public static int WriteTexture(
+            this ObservationWriter obsWriter,
+            Texture2D texture,
+            bool grayScale)
+        {
+            var width = texture.width;
+            var height = texture.height;
+
+            var texturePixels = texture.GetPixels32();
+            // During training, we convert from Texture to PNG before sending to the trainer, which has the
+            // effect of flipping the image. We need another flip here at inference time to match this.
+            for (var h = height - 1; h >= 0; h--)
+            {
+                for (var w = 0; w < width; w++)
+                {
+                    var currentPixel = texturePixels[(height - h - 1) * width + w];
+                    if (grayScale)
+                    {
+                        obsWriter[h, w, 0] =
+                            (currentPixel.r + currentPixel.g + currentPixel.b) / 3f / 255.0f;
+                    }
+                    else
+                    {
+                        // For Color32, the r, g and b values are between 0 and 255.
+                        obsWriter[h, w, 0] = currentPixel.r / 255.0f;
+                        obsWriter[h, w, 1] = currentPixel.g / 255.0f;
+                        obsWriter[h, w, 2] = currentPixel.b / 255.0f;
+                    }
+                }
+            }
+
+            return height * width * (grayScale ? 1 : 3);
+        }
+    }
 }

com.unity.ml-agents/Runtime/Sensors/RenderTextureSensor.cs

             using (TimerStack.Instance.Scoped("RenderTextureSensor.Write"))
             {
                 var texture = ObservationToTexture(m_RenderTexture);
-                var numWritten = Utilities.TextureToTensorProxy(texture, writer, m_Grayscale);
+                var numWritten = writer.WriteTexture(texture, m_Grayscale);
                 DestroyTexture(texture);
                 return numWritten;
             }

com.unity.ml-agents/Runtime/Sensors/StackingSensor.cs

             int height = m_WrappedSensor.GetObservationShape()[0];
             int width = m_WrappedSensor.GetObservationShape()[1];
             var texture2D = new Texture2D(width, height, TextureFormat.RGB24, false);
+            Color32[] resetColorArray = texture2D.GetPixels32();
+            Color32 black = new Color32(0, 0, 0, 0);
+            for (int i = 0; i < resetColorArray.Length; i++)
+            {
+                resetColorArray[i] = black;
+            }
+            texture2D.SetPixels32(resetColorArray);
+            texture2D.Apply();
             return texture2D.EncodeToPNG();
         }
 

com.unity.ml-agents/Runtime/Utilities.cs

 {
     internal static class Utilities
     {
-        /// <summary>
-        /// Puts a Texture2D into a ObservationWriter.
-        /// </summary>
-        /// <param name="texture">
-        /// The texture to be put into the tensor.
-        /// </param>
-        /// <param name="obsWriter">
-        /// Writer to fill with Texture data.
-        /// </param>
-        /// <param name="grayScale">
-        /// If set to <c>true</c> the textures will be converted to grayscale before
-        /// being stored in the tensor.
-        /// </param>
-        /// <returns>The number of floats written</returns>
-        internal static int TextureToTensorProxy(
-            Texture2D texture,
-            ObservationWriter obsWriter,
-            bool grayScale)
-        {
-            var width = texture.width;
-            var height = texture.height;
-
-            var texturePixels = texture.GetPixels32();
-            // During training, we convert from Texture to PNG before sending to the trainer, which has the
-            // effect of flipping the image. We need another flip here at inference time to match this.
-            for (var h = height - 1; h >= 0; h--)
-            {
-                for (var w = 0; w < width; w++)
-                {
-                    var currentPixel = texturePixels[(height - h - 1) * width + w];
-                    if (grayScale)
-                    {
-                        obsWriter[h, w, 0] =
-                            (currentPixel.r + currentPixel.g + currentPixel.b) / 3f / 255.0f;
-                    }
-                    else
-                    {
-                        // For Color32, the r, g and b values are between 0 and 255.
-                        obsWriter[h, w, 0] = currentPixel.r / 255.0f;
-                        obsWriter[h, w, 1] = currentPixel.g / 255.0f;
-                        obsWriter[h, w, 2] = currentPixel.b / 255.0f;
-                    }
-                }
-            }
-
-            return height * width * (grayScale ? 1 : 3);
-        }
 
         /// <summary>
         /// Calculates the cumulative sum of an integer array. The result array will be one element

com.unity.ml-agents/Tests/Editor/MLAgentsEditModeTest.cs

             }
         }
     }
+
+    [TestFixture]
+    public class AgentRecursionTests
+    {
+        [SetUp]
+        public void SetUp()
+        {
+            if (Academy.IsInitialized)
+            {
+                Academy.Instance.Dispose();
+            }
+        }
+
+        class CollectObsEndEpisodeAgent : Agent
+        {
+            public override void CollectObservations(VectorSensor sensor)
+            {
+                // NEVER DO THIS IN REAL CODE!
+                EndEpisode();
+            }
+        }
+
+        class OnEpisodeBeginEndEpisodeAgent : Agent
+        {
+            public override void OnEpisodeBegin()
+            {
+                // NEVER DO THIS IN REAL CODE!
+                EndEpisode();
+            }
+        }
+
+        void TestRecursiveThrows<T>() where T : Agent
+        {
+            var gameObj = new GameObject();
+            var agent = gameObj.AddComponent<T>();
+            agent.LazyInitialize();
+            agent.RequestDecision();
+
+            Assert.Throws<UnityAgentsException>(() =>
+            {
+                Academy.Instance.EnvironmentStep();
+            });
+        }
+
+        [Test]
+        public void TestRecursiveCollectObsEndEpisodeThrows()
+        {
+            TestRecursiveThrows<CollectObsEndEpisodeAgent>();
+        }
+
+        [Test]
+        public void TestRecursiveOnEpisodeBeginEndEpisodeThrows()
+        {
+            TestRecursiveThrows<OnEpisodeBeginEndEpisodeAgent>();
+        }
+    }
 }

com.unity.ml-agents/package.json

   "unity": "2018.4",
   "description": "Use state-of-the-art machine learning to create intelligent character behaviors in any Unity environment (games, robotics, film, etc.).",
   "dependencies": {
-    "com.unity.barracuda": "1.1.1-preview",
+    "com.unity.barracuda": "1.1.2-preview",
     "com.unity.modules.imageconversion": "1.0.0",
     "com.unity.modules.jsonserialize": "1.0.0",
     "com.unity.modules.physics": "1.0.0",

docs/Background-Machine-Learning.md

 one where the number of observations an agent perceives and the number of
 actions they can take are large). Many of the algorithms we provide in ML-Agents
 use some form of deep learning, built on top of the open-source library,
-[TensorFlow](Background-TensorFlow.md).
+[PyTorch](Background-PyTorch.md).

docs/Getting-Started.md

 
 ## Running a pre-trained model
 
-We include pre-trained models for our agents (`.nn` files) and we use the
+We include pre-trained models for our agents (`.onnx` files) and we use the
 [Unity Inference Engine](Unity-Inference-Engine.md) to run these models inside
 Unity. In this section, we will use the pre-trained model for the 3D Ball
 example.
 
 ## Training a new model with Reinforcement Learning
 
-While we provide pre-trained `.nn` files for the agents in this environment, any
+While we provide pre-trained models for the agents in this environment, any
 environment you make yourself will require training agents from scratch to
 generate a new model file. In this section we will demonstrate how to use the
 reinforcement learning algorithms that are part of the ML-Agents Python package
 use it with compatible Agents (the Agents that generated the model). **Note:**
 Do not just close the Unity Window once the `Saved Model` message appears.
 Either wait for the training process to close the window or press `Ctrl+C` at
-the command-line prompt. If you close the window manually, the `.nn` file
+the command-line prompt. If you close the window manually, the `.onnx` file
 containing the trained model is not exported into the ml-agents folder.
 
 If you've quit the training early using `Ctrl+C` and want to resume training,
 mlagents-learn config/ppo/3DBall.yaml --run-id=first3DBallRun --resume
 ```
 
-Your trained model will be at `results/<run-identifier>/<behavior_name>.nn` where
+Your trained model will be at `results/<run-identifier>/<behavior_name>.onnx` where
 `<behavior_name>` is the name of the `Behavior Name` of the agents corresponding
 to the model. This file corresponds to your model's latest checkpoint. You can
 now embed this trained model into your Agents by following the steps below,
    `Project/Assets/ML-Agents/Examples/3DBall/TFModels/`.
 1. Open the Unity Editor, and select the **3DBall** scene as described above.
 1. Select the **3DBall** prefab Agent object.
-1. Drag the `<behavior_name>.nn` file from the Project window of the Editor to
+1. Drag the `<behavior_name>.onnx` file from the Project window of the Editor to
    the **Model** placeholder in the **Ball3DAgent** inspector window.
 1. Press the **Play** button at the top of the Editor.
 

docs/Installation.md

 [instructions](https://packaging.python.org/guides/installing-using-linux-tools/#installing-pip-setuptools-wheel-with-linux-package-managers)
 on installing it.
 
-Although we do not provide support for Anaconda installation on Windows, the
-previous
-[Windows Anaconda Installation (Deprecated) guide](Installation-Anaconda-Windows.md)
-is still available.
-
 ### Clone the ML-Agents Toolkit Repository (Optional)
 
 Now that you have installed Unity and Python, you can now install the Unity and
 dependencies for each project and are supported on Mac / Windows / Linux. We
 offer a dedicated [guide on Virtual Environments](Using-Virtual-Environment.md).
 
+#### (Windows) Installing PyTorch
+
+On Windows, you'll have to install the PyTorch package separately prior to
+installing ML-Agents. Activate your virtual environment and run from the command line:
+
+```sh
+pip3 install torch==1.7.0 -f https://download.pytorch.org/whl/torch_stable.html
+```
+
+Note that on Windows, you may also need Microsoft's
+[Visual C++ Redistributable](https://support.microsoft.com/en-us/help/2977003/the-latest-supported-visual-c-downloads)
+if you don't have it already. See the [PyTorch installation guide](https://pytorch.org/get-started/locally/)
+for more installation options and versions.
+
+#### Installing `mlagents`
+
 To install the `mlagents` Python package, activate your virtual environment and
 run from the command line:
 
 
 By installing the `mlagents` package, the dependencies listed in the
 [setup.py file](../ml-agents/setup.py) are also installed. These include
-[TensorFlow](Background-TensorFlow.md) (Requires a CPU w/ AVX support).
+[PyTorch](Background-PyTorch.md) (Requires a CPU w/ AVX support).
 
 #### Advanced: Local Installation for Development
 
 the repository's root directory, run:
 
 ```sh
+pip3 install torch -f https://download.pytorch.org/whl/torch_stable.html
 pip3 install -e ./ml-agents-envs
 pip3 install -e ./ml-agents
 ```

docs/Learning-Environment-Create-New.md

     }
 
     // Fell off platform
-    if (this.transform.localPosition.y < 0)
+    else if (this.transform.localPosition.y < 0)
     {
         EndEpisode();
     }
 1. Add the `Decision Requester` script with the Add Component button from the
    RollerAgent Inspector.
 1. Change **Decision Period** to `10`. For more information on decisions, see [the Agent documentation](Learning-Environment-Design-Agents.md#decisions)
-1. Drag the Target GameObject from the Hierarchy window to the RollerAgent
-   Target field.
 1. Add the `Behavior Parameters` script with the Add Component button from the
    RollerAgent Inspector.
 1. Modify the Behavior Parameters of the Agent :

docs/Learning-Environment-Design-Agents.md

     AddReward(1.0f);
     EndEpisode();
 }
-if (hitObjects.Where(col => col.gameObject.tag == "pit").ToArray().Length == 1)
+else if (hitObjects.Where(col => col.gameObject.tag == "pit").ToArray().Length == 1)
 {
     AddReward(-1f);
     EndEpisode();

docs/Learning-Environment-Executable.md

 ```
 
 You can press Ctrl+C to stop the training, and your trained model will be at
-`results/<run-identifier>/<behavior_name>.nn`, which corresponds to your model's
+`results/<run-identifier>/<behavior_name>.onnx`, which corresponds to your model's
 latest checkpoint. (**Note:** There is a known bug on Windows that causes the
 saving of the model to fail when you early terminate the training, it's
 recommended to wait until Step has reached the max_steps parameter you set in
    `Project/Assets/ML-Agents/Examples/3DBall/TFModels/`.
 1. Open the Unity Editor, and select the **3DBall** scene as described above.
 1. Select the **3DBall** prefab from the Project window and select **Agent**.
-1. Drag the `<behavior_name>.nn` file from the Project window of the Editor to
+1. Drag the `<behavior_name>.onnx` file from the Project window of the Editor to
    the **Model** placeholder in the **Ball3DAgent** inspector window.
 1. Press the **Play** button at the top of the Editor.

docs/ML-Agents-Overview.md

 for training intelligent agents. Agents can be trained using reinforcement
 learning, imitation learning, neuroevolution, or other machine learning methods
 through a simple-to-use Python API. We also provide implementations (based on
-TensorFlow) of state-of-the-art algorithms to enable game developers and
+PyTorch) of state-of-the-art algorithms to enable game developers and
 hobbyists to easily train intelligent agents for 2D, 3D and VR/AR games. These
 trained agents can be used for multiple purposes, including controlling NPC
 behavior (in a variety of settings such as multi-agent and adversarial),
 that include overviews and helpful resources on the
 [Unity Engine](Background-Unity.md),
 [machine learning](Background-Machine-Learning.md) and
-[TensorFlow](Background-TensorFlow.md). We **strongly** recommend browsing the
+[PyTorch](Background-PyTorch.md). We **strongly** recommend browsing the
-machine learning concepts or have not previously heard of TensorFlow.
+machine learning concepts or have not previously heard of PyTorch.
 
 The remainder of this page contains a deep dive into ML-Agents, its key
 components, different training modes and scenarios. By the end of it, you should
 
 ### Custom Training and Inference
 
-In the previous mode, the Agents were used for training to generate a TensorFlow
+In the previous mode, the Agents were used for training to generate a PyTorch
 model that the Agents can later use. However, any user of the ML-Agents Toolkit
 can leverage their own algorithms for training. In this case, the behaviors of
 all the Agents in the scene will be controlled within Python. You can even turn

docs/Readme.md

 - [ML-Agents Toolkit Overview](ML-Agents-Overview.md)
   - [Background: Unity](Background-Unity.md)
   - [Background: Machine Learning](Background-Machine-Learning.md)
-  - [Background: TensorFlow](Background-TensorFlow.md)
+  - [Background: PyTorch](Background-PyTorch.md)
 - [Example Environments](Learning-Environment-Examples.md)
 
 ## Creating Learning Environments

docs/Training-Configuration-File.md

 | `time_horizon`           | (default = `64`) How many steps of experience to collect per-agent before adding it to the experience buffer. When this limit is reached before the end of an episode, a value estimate is used to predict the overall expected reward from the agent's current state. As such, this parameter trades off between a less biased, but higher variance estimate (long time horizon) and more biased, but less varied estimate (short time horizon). In cases where there are frequent rewards within an episode, or episodes are prohibitively large, a smaller number can be more ideal. This number should be large enough to capture all the important behavior within a sequence of an agent's actions. <br><br> Typical range: `32` - `2048` |
 | `max_steps`              | (default = `500000`) Total number of steps (i.e., observation collected and action taken) that must be taken in the environment (or across all environments if using multiple in parallel) before ending the training process. If you have multiple agents with the same behavior name within your environment, all steps taken by those agents will contribute to the same `max_steps` count. <br><br>Typical range: `5e5` - `1e7`                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          |
 | `keep_checkpoints`         | (default = `5`) The maximum number of model checkpoints to keep. Checkpoints are saved after the number of steps specified by the checkpoint_interval option. Once the maximum number of checkpoints has been reached, the oldest checkpoint is deleted when saving a new checkpoint. |
-| `checkpoint_interval`         | (default = `500000`) The number of experiences collected between each checkpoint by the trainer. A maximum of `keep_checkpoints` checkpoints are saved before old ones are deleted. Each checkpoint saves the `.nn` (and `.onnx` if applicable) files in `results/` folder.|
+| `checkpoint_interval`         | (default = `500000`) The number of experiences collected between each checkpoint by the trainer. A maximum of `keep_checkpoints` checkpoints are saved before old ones are deleted. Each checkpoint saves the `.onnx` (and `.nn` if using TensorFlow) files in `results/` folder.|
 | `init_path`              | (default = None) Initialize trainer from a previously saved model. Note that the prior run should have used the same trainer configurations as the current run, and have been saved with the same version of ML-Agents. <br><br>You should provide the full path to the folder where the checkpoints were saved, e.g. `./models/{run-id}/{behavior_name}`. This option is provided in case you want to initialize different behaviors from different runs; in most cases, it is sufficient to use the `--initialize-from` CLI parameter to initialize all models from the same run.                                                                                                                                  |
 | `threaded`               | (default = `true`) By default, model updates can happen while the environment is being stepped. This violates the [on-policy](https://spinningup.openai.com/en/latest/user/algorithms.html#the-on-policy-algorithms) assumption of PPO slightly in exchange for a training speedup. To maintain the strict on-policyness of PPO, you can disable parallel updates by setting `threaded` to `false`. There is usually no reason to turn `threaded` off for SAC.                                                                                                                                                                                                                                                       |
 | `hyperparameters -> learning_rate`          | (default = `3e-4`) Initial learning rate for gradient descent. Corresponds to the strength of each gradient descent update step. This should typically be decreased if training is unstable, and the reward does not consistently increase. <br><br>Typical range: `1e-5` - `1e-3`                                                                                                                                                                                                                                                                                                                                                                                                                                                                |

docs/Training-ML-Agents.md

     - [Curriculum Learning](#curriculum)
       - [Training with a Curriculum](#training-with-a-curriculum)
   - [Training Using Concurrent Unity Instances](#training-using-concurrent-unity-instances)
-  - [Using PyTorch (Experimental)](#using-pytorch-experimental)
 
 For a broad overview of reinforcement learning, imitation learning and all the
 training scenarios, methods and options within the ML-Agents Toolkit, see
    values. See [Using TensorBoard](Using-Tensorboard.md) for more details on how
    to visualize the training metrics.
 1. Models: these contain the model checkpoints that
-   are updated throughout training and the final model file (`.nn`). This final
+   are updated throughout training and the final model file (`.onnx`). This final
    model file is generated once either when training completes or is
    interrupted.
 1. Timers file (under `results/<run-identifier>/run_logs`): this contains aggregated
 - **Result Variation Using Concurrent Unity Instances** - If you keep all the
   hyperparameters the same, but change `--num-envs=<n>`, the results and model
   would likely change.
-
-### Using PyTorch (Experimental)
-
-ML-Agents, by default, uses TensorFlow as its backend, but experimental support
-for PyTorch has been added. To use PyTorch, the `torch` Python package must
-be installed, and PyTorch must be enabled for your trainer.
-
-#### Installing PyTorch
-
-If you've already installed ML-Agents, follow the
-[official PyTorch install instructions](https://pytorch.org/get-started/locally/) for
-your platform and configuration. Note that on Windows, you may also need Microsoft's
-[Visual C++ Redistributable](https://support.microsoft.com/en-us/help/2977003/the-latest-supported-visual-c-downloads) if you don't have it already.
-
-If you're installing or upgrading ML-Agents on Linux or Mac, you can also run
-`pip3 install mlagents[torch]` instead of `pip3 install mlagents`
-during [installation](Installation.md). On Windows, install ML-Agents first and then
-separately install PyTorch.
-
-#### Enabling PyTorch
-
-PyTorch can be enabled in one of two ways. First, by adding `--torch` to the
-`mlagents-learn` command. This will make all behaviors train with PyTorch.
-
-Second, by changing the `framework` option for your agent behavior in the
-configuration YAML as below. This will use PyTorch just for that behavior.
-
-```yaml
-behaviors:
-  YourAgentBehavior:
-    framework: pytorch
-```

docs/Training-on-Amazon-Web-Service.md

     # Download and install the latest Nvidia driver for ubuntu
     # Please refer to http://download.nvidia.com/XFree86/Linux-#x86_64/latest.txt
     $ wget http://download.nvidia.com/XFree86/Linux-x86_64/390.87/NVIDIA-Linux-x86_64-390.87.run
-    $ sudo /bin/bash ./NVIDIA-Linux-x86_64-390.67.run --accept-license --no-questions --ui=none
+    $ sudo /bin/bash ./NVIDIA-Linux-x86_64-390.87.run --accept-license --no-questions --ui=none
 
     # Disable Nouveau as it will clash with the Nvidia driver
     $ sudo echo 'blacklist nouveau'  | sudo tee -a /etc/modprobe.d/blacklist.conf

docs/Unity-Inference-Engine.md

   [industry-standard open format](https://onnx.ai/about.html) produced by the
   [tf2onnx package](https://github.com/onnx/tensorflow-onnx).
 
-Export to ONNX is currently considered beta. To enable it, make sure
-`tf2onnx>=1.5.5` is installed in pip. tf2onnx does not currently support
-tensorflow 2.0.0 or later, or earlier than 1.12.0.
+Export to ONNX is used if using PyTorch (the default). To enable it
+while using TensorFlow, make sure `tf2onnx>=1.6.1` is installed in pip.
 
 ## Using the Unity Inference Engine
 

gym-unity/gym_unity/__init__.py

 # Version of the library that will be used to upload to pypi
-__version__ = "0.21.1"
+__version__ = "0.22.0.dev0"
-__release_tag__ = "release_9"
+__release_tag__ = None

gym-unity/gym_unity/envs/__init__.py

         self._previous_decision_step = decision_steps
 
         # Set action spaces
-        if self.group_spec.is_action_discrete():
-            branches = self.group_spec.discrete_action_branches
-            if self.group_spec.action_size == 1:
+        if self.group_spec.action_spec.is_discrete():
+            self.action_size = self.group_spec.action_spec.discrete_size
+            branches = self.group_spec.action_spec.discrete_branches
+            if self.group_spec.action_spec.discrete_size == 1:
                 self._action_space = spaces.Discrete(branches[0])
             else:
                 if flatten_branched:
                     self._action_space = spaces.MultiDiscrete(branches)
 
-        else:
+        elif self.group_spec.action_spec.is_continuous():
-            high = np.array([1] * self.group_spec.action_shape)
+
+            self.action_size = self.group_spec.action_spec.continuous_size
+            high = np.array([1] * self.group_spec.action_spec.continuous_size)
+        else:
+            raise UnityGymException(
+                "The gym wrapper does not provide explicit support for both discrete "
+                "and continuous actions."
+            )
 
         # Set observations space
         list_spaces: List[gym.Space] = []
             # Translate action into list
             action = self._flattener.lookup_action(action)
 
-        spec = self.group_spec
-        action = np.array(action).reshape((1, spec.action_size))
+        action = np.array(action).reshape((1, self.action_size))
         self._env.set_actions(self.name, action)
 
         self._env.step()

gym-unity/gym_unity/tests/test_gym.py

 from gym_unity.envs import UnityToGymWrapper
 from mlagents_envs.base_env import (
     BehaviorSpec,
-    ActionType,
+    ActionSpec,
     DecisionSteps,
     TerminalSteps,
     BehaviorMapping,
     Creates a mock BrainParameters object with parameters.
     """
     # Avoid using mutable object as default param
-    act_type = ActionType.DISCRETE
-        act_type = ActionType.CONTINUOUS
+        action_spec = ActionSpec.create_continuous(vector_action_space_size)
+        action_spec = ActionSpec.create_discrete(vector_action_space_size)
-    return BehaviorSpec(obs_shapes, act_type, vector_action_space_size)
+    return BehaviorSpec(obs_shapes, action_spec)
 
 
 def create_mock_vector_steps(specs, num_agents=1, number_visual_observations=0):

ml-agents-envs/mlagents_envs/__init__.py

 # Version of the library that will be used to upload to pypi
-__version__ = "0.21.1"
+__version__ = "0.22.0.dev0"
-__release_tag__ = "release_9"
+__release_tag__ = None

ml-agents-envs/mlagents_envs/base_env.py

     NamedTuple,
     Tuple,
     Optional,
-    Union,
     Dict,
     Iterator,
     Any,
-from enum import Enum
+
+from mlagents_envs.exception import UnityActionException
 
 AgentId = int
 BehaviorName = str
         )
 
 
-class ActionType(Enum):
-    DISCRETE = 0
-    CONTINUOUS = 1
-
-
-class BehaviorSpec(NamedTuple):
+class ActionSpec(NamedTuple):
-    A NamedTuple to containing information about the observations and actions
-    spaces for a group of Agents under the same behavior.
-     - observation_shapes is a List of Tuples of int : Each Tuple corresponds
-     to an observation's dimensions. The shape tuples have the same ordering as
-     the ordering of the DecisionSteps and TerminalSteps.
-     - action_type is the type of data of the action. it can be discrete or
-     continuous. If discrete, the action tensors are expected to be int32. If
-     continuous, the actions are expected to be float32.
-     - action_shape is:
-       - An int in continuous action space corresponding to the number of
-     floats that constitute the action.
-       - A Tuple of int in discrete action space where each int corresponds to
-       the number of discrete actions available to the agent.
+    A NamedTuple containing utility functions and information about the action spaces
+    for a group of Agents under the same behavior.
+    - num_continuous_actions is an int corresponding to the number of floats which
+    constitute the action.
+    - discrete_branch_sizes is a Tuple of int where each int corresponds to
+    the number of discrete actions available to the agent on an independent action branch.
-    observation_shapes: List[Tuple]
-    action_type: ActionType
-    action_shape: Union[int, Tuple[int, ...]]
+    continuous_size: int
+    discrete_branches: Tuple[int, ...]
+
+    def __eq__(self, other):
+        return (
+            self.continuous_size == other.continuous_size
+            and self.discrete_branches == other.discrete_branches
+        )
+
+    def __str__(self):
+        return f"Continuous: {self.continuous_size}, Discrete: {self.discrete_branches}"
-    def is_action_discrete(self) -> bool:
+    # For backwards compatibility
+    def is_discrete(self) -> bool:
-        return self.action_type == ActionType.DISCRETE
+        return self.discrete_size > 0 and self.continuous_size == 0
-    def is_action_continuous(self) -> bool:
+    # For backwards compatibility
+    def is_continuous(self) -> bool:
-        return self.action_type == ActionType.CONTINUOUS
+        return self.discrete_size == 0 and self.continuous_size > 0
-    def action_size(self) -> int:
+    def discrete_size(self) -> int:
-        Returns the dimension of the action.
-         - In the continuous case, will return the number of continuous actions.
-         - In the (multi-)discrete case, will return the number of action.
-         branches.
+        Returns a an int corresponding to the number of discrete branches.
-        if self.action_type == ActionType.DISCRETE:
-            return len(self.action_shape)  # type: ignore
-        else:
-            return self.action_shape  # type: ignore
-
-    @property
-    def discrete_action_branches(self) -> Optional[Tuple[int, ...]]:
-        """
-        Returns a Tuple of int corresponding to the number of possible actions
-        for each branch (only for discrete actions). Will return None in
-        for continuous actions.
-        """
-        if self.action_type == ActionType.DISCRETE:
-            return self.action_shape  # type: ignore
-        else:
-            return None
+        return len(self.discrete_branches)
-    def create_empty_action(self, n_agents: int) -> np.ndarray:
+    def empty_action(self, n_agents: int) -> np.ndarray:
-        if self.action_type == ActionType.DISCRETE:
-            return np.zeros((n_agents, self.action_size), dtype=np.int32)
-        else:
-            return np.zeros((n_agents, self.action_size), dtype=np.float32)
+        if self.is_continuous():
+            return np.zeros((n_agents, self.continuous_size), dtype=np.float32)
+        return np.zeros((n_agents, self.discrete_size), dtype=np.int32)
-    def create_random_action(self, n_agents: int) -> np.ndarray:
+    def random_action(self, n_agents: int) -> np.ndarray:
-        :param generator: The random number generator used for creating random action
-        if self.is_action_continuous():
+        if self.is_continuous():
-                low=-1.0, high=1.0, size=(n_agents, self.action_size)
+                low=-1.0, high=1.0, size=(n_agents, self.continuous_size)
-            return action
-        elif self.is_action_discrete():
-            branch_size = self.discrete_action_branches
+        else:
+            branch_size = self.discrete_branches
             action = np.column_stack(
                 [
                     np.random.randint(
                         dtype=np.int32,
                     )
-                    for i in range(self.action_size)
+                    for i in range(self.discrete_size)
-            return action
+        return action
+
+    def _validate_action(
+        self, actions: np.ndarray, n_agents: int, name: str
+    ) -> np.ndarray:
+        """
+        Validates that action has the correct action dim
+        for the correct number of agents and ensures the type.
+        """
+        if self.continuous_size > 0:
+            _size = self.continuous_size
+        else:
+            _size = self.discrete_size
+        _expected_shape = (n_agents, _size)
+        if actions.shape != _expected_shape:
+            raise UnityActionException(
+                f"The behavior {name} needs an input of dimension "
+                f"{_expected_shape} for (<number of agents>, <action size>) but "
+                f"received input of dimension {actions.shape}"
+            )
+        _expected_type = np.float32 if self.is_continuous() else np.int32
+        if actions.dtype != _expected_type:
+            actions = actions.astype(_expected_type)
+        return actions
+
+    @staticmethod
+    def create_continuous(continuous_size: int) -> "ActionSpec":
+        """
+        Creates an ActionSpec that is homogenously continuous
+        """
+        return ActionSpec(continuous_size, ())
+
+    @staticmethod
+    def create_discrete(discrete_branches: Tuple[int]) -> "ActionSpec":
+        """
+        Creates an ActionSpec that is homogenously discrete
+        """
+        return ActionSpec(0, discrete_branches)
+
+
+class BehaviorSpec(NamedTuple):
+    """
+    A NamedTuple containing information about the observation and action
+    spaces for a group of Agents under the same behavior.
+    - observation_shapes is a List of Tuples of int : Each Tuple corresponds
+    to an observation's dimensions. The shape tuples have the same ordering as
+    the ordering of the DecisionSteps and TerminalSteps.
+    - action_spec is an ActionSpec NamedTuple
+    """
+
+    observation_shapes: List[Tuple]
+    action_spec: ActionSpec
 
 
 class BehaviorMapping(Mapping):

ml-agents-envs/mlagents_envs/environment.py

                     n_agents = len(self._env_state[group_name][0])
                 self._env_actions[group_name] = self._env_specs[
                     group_name
-                ].create_empty_action(n_agents)
+                ].action_spec.empty_action(n_agents)
         step_input = self._generate_step_input(self._env_actions)
         with hierarchical_timer("communicator.exchange"):
             outputs = self._communicator.exchange(step_input)
         self._assert_behavior_exists(behavior_name)
         if behavior_name not in self._env_state:
             return
-        spec = self._env_specs[behavior_name]
-        expected_type = np.float32 if spec.is_action_continuous() else np.int32
-        expected_shape = (len(self._env_state[behavior_name][0]), spec.action_size)
-        if action.shape != expected_shape:
-            raise UnityActionException(
-                f"The behavior {behavior_name} needs an input of dimension "
-                f"{expected_shape} for (<number of agents>, <action size>) but "
-                f"received input of dimension {action.shape}"
-            )
-        if action.dtype != expected_type:
-            action = action.astype(expected_type)
+        action_spec = self._env_specs[behavior_name].action_spec
+        num_agents = len(self._env_state[behavior_name][0])
+        action = action_spec._validate_action(action, num_agents, behavior_name)
         self._env_actions[behavior_name] = action
 
     def set_action_for_agent(
         if behavior_name not in self._env_state:
             return
-        spec = self._env_specs[behavior_name]
-        expected_shape = (spec.action_size,)
-        if action.shape != expected_shape:
-            raise UnityActionException(
-                f"The Agent {agent_id} with BehaviorName {behavior_name} needs "
-                f"an input of dimension {expected_shape} but received input of "
-                f"dimension {action.shape}"
-            )
-        expected_type = np.float32 if spec.is_action_continuous() else np.int32
-        if action.dtype != expected_type:
-            action = action.astype(expected_type)
-
+        action_spec = self._env_specs[behavior_name].action_spec
+        num_agents = len(self._env_state[behavior_name][0])
+        action = action_spec._validate_action(action, num_agents, behavior_name)
-            self._env_actions[behavior_name] = spec.create_empty_action(
-                len(self._env_state[behavior_name][0])
-            )
+            self._env_actions[behavior_name] = action_spec.empty_action(num_agents)
         try:
             index = np.where(self._env_state[behavior_name][0].agent_id == agent_id)[0][
                 0

ml-agents-envs/mlagents_envs/rpc_utils.py

 from mlagents_envs.base_env import (
     BehaviorSpec,
-    ActionType,
+    ActionSpec,
     DecisionSteps,
     TerminalSteps,
 )
 from mlagents_envs.communicator_objects.brain_parameters_pb2 import BrainParametersProto
 import numpy as np
 import io
-from typing import cast, List, Tuple, Union, Collection, Optional, Iterable
+from typing import cast, List, Tuple, Collection, Optional, Iterable
 from PIL import Image
 
 
     :return: BehaviorSpec object.
     """
     observation_shape = [tuple(obs.shape) for obs in agent_info.observations]
-    action_type = (
-        ActionType.DISCRETE
-        if brain_param_proto.vector_action_space_type == 0
-        else ActionType.CONTINUOUS
-    )
-    if action_type == ActionType.CONTINUOUS:
-        action_shape: Union[
-            int, Tuple[int, ...]
-        ] = brain_param_proto.vector_action_size[0]
+    if brain_param_proto.vector_action_space_type == 1:
+        action_spec = ActionSpec(brain_param_proto.vector_action_size[0], ())
-        action_shape = tuple(brain_param_proto.vector_action_size)
-    return BehaviorSpec(observation_shape, action_type, action_shape)
+        action_spec = ActionSpec(0, tuple(brain_param_proto.vector_action_size))
+    return BehaviorSpec(observation_shape, action_spec)
 
 
 class OffsetBytesIO:
         [agent_info.id for agent_info in terminal_agent_info_list], dtype=np.int32
     )
     action_mask = None
-    if behavior_spec.is_action_discrete():
+    if behavior_spec.action_spec.discrete_size > 0:
-            a_size = np.sum(behavior_spec.discrete_action_branches)
+            a_size = np.sum(behavior_spec.action_spec.discrete_branches)
             mask_matrix = np.ones((n_agents, a_size), dtype=np.bool)
             for agent_index, agent_info in enumerate(decision_agent_info_list):
                 if agent_info.action_mask is not None:
                             for k in range(a_size)
                         ]
             action_mask = (1 - mask_matrix).astype(np.bool)
-            indices = _generate_split_indices(behavior_spec.discrete_action_branches)
+            indices = _generate_split_indices(
+                behavior_spec.action_spec.discrete_branches
+            )
             action_mask = np.split(action_mask, indices, axis=1)
     return (
         DecisionSteps(

ml-agents-envs/mlagents_envs/tests/test_envs.py

 from unittest import mock
 import pytest
 
-import numpy as np
-
 from mlagents_envs.environment import UnityEnvironment
 from mlagents_envs.base_env import DecisionSteps, TerminalSteps
 from mlagents_envs.exception import UnityEnvironmentException, UnityActionException
     env.step()
     decision_steps, terminal_steps = env.get_steps("RealFakeBrain")
     n_agents = len(decision_steps)
-    env.set_actions(
-        "RealFakeBrain", np.zeros((n_agents, spec.action_size), dtype=np.float32)
-    )
+    env.set_actions("RealFakeBrain", spec.action_spec.empty_action(n_agents))
-        env.set_actions(
-            "RealFakeBrain",
-            np.zeros((n_agents - 1, spec.action_size), dtype=np.float32),
-        )
+        env.set_actions("RealFakeBrain", spec.action_spec.empty_action(n_agents - 1))
-    env.set_actions(
-        "RealFakeBrain", -1 * np.ones((n_agents, spec.action_size), dtype=np.float32)
-    )
+    env.set_actions("RealFakeBrain", spec.action_spec.empty_action(n_agents) - 1)
     env.step()
 
     env.close()

ml-agents-envs/mlagents_envs/tests/test_rpc_utils.py

 from mlagents_envs.communicator_objects.agent_action_pb2 import AgentActionProto
 from mlagents_envs.base_env import (
     BehaviorSpec,
-    ActionType,
+    ActionSpec,
     DecisionSteps,
     TerminalSteps,
 )
 def test_batched_step_result_from_proto():
     n_agents = 10
     shapes = [(3,), (4,)]
-    spec = BehaviorSpec(shapes, ActionType.CONTINUOUS, 3)
+    spec = BehaviorSpec(shapes, ActionSpec.create_continuous(3))
     ap_list = generate_list_agent_proto(n_agents, shapes)
     decision_steps, terminal_steps = steps_from_proto(ap_list, spec)
     for agent_id in range(n_agents):
 def test_action_masking_discrete():
     n_agents = 10
     shapes = [(3,), (4,)]
-    behavior_spec = BehaviorSpec(shapes, ActionType.DISCRETE, (7, 3))
+    behavior_spec = BehaviorSpec(shapes, ActionSpec.create_discrete((7, 3)))
     ap_list = generate_list_agent_proto(n_agents, shapes)
     decision_steps, terminal_steps = steps_from_proto(ap_list, behavior_spec)
     masks = decision_steps.action_mask
 def test_action_masking_discrete_1():
     n_agents = 10
     shapes = [(3,), (4,)]
-    behavior_spec = BehaviorSpec(shapes, ActionType.DISCRETE, (10,))
+    behavior_spec = BehaviorSpec(shapes, ActionSpec.create_discrete((10,)))
     ap_list = generate_list_agent_proto(n_agents, shapes)
     decision_steps, terminal_steps = steps_from_proto(ap_list, behavior_spec)
     masks = decision_steps.action_mask
 def test_action_masking_discrete_2():
     n_agents = 10
     shapes = [(3,), (4,)]
-    behavior_spec = BehaviorSpec(shapes, ActionType.DISCRETE, (2, 2, 6))
+    behavior_spec = BehaviorSpec(shapes, ActionSpec.create_discrete((2, 2, 6)))
     ap_list = generate_list_agent_proto(n_agents, shapes)
     decision_steps, terminal_steps = steps_from_proto(ap_list, behavior_spec)
     masks = decision_steps.action_mask
 def test_action_masking_continuous():
     n_agents = 10
     shapes = [(3,), (4,)]
-    behavior_spec = BehaviorSpec(shapes, ActionType.CONTINUOUS, 10)
+    behavior_spec = BehaviorSpec(shapes, ActionSpec.create_continuous(10))
     ap_list = generate_list_agent_proto(n_agents, shapes)
     decision_steps, terminal_steps = steps_from_proto(ap_list, behavior_spec)
     masks = decision_steps.action_mask
     bp.vector_action_size.extend([5, 4])
     bp.vector_action_space_type = 0
     behavior_spec = behavior_spec_from_proto(bp, agent_proto)
-    assert behavior_spec.is_action_discrete()
-    assert not behavior_spec.is_action_continuous()
+    assert behavior_spec.action_spec.is_discrete()
+    assert not behavior_spec.action_spec.is_continuous()
-    assert behavior_spec.discrete_action_branches == (5, 4)
-    assert behavior_spec.action_size == 2
+    assert behavior_spec.action_spec.discrete_branches == (5, 4)
+    assert behavior_spec.action_spec.discrete_size == 2
-    assert not behavior_spec.is_action_discrete()
-    assert behavior_spec.is_action_continuous()
-    assert behavior_spec.action_size == 6
+    assert not behavior_spec.action_spec.is_discrete()
+    assert behavior_spec.action_spec.is_continuous()
+    assert behavior_spec.action_spec.continuous_size == 6
-    behavior_spec = BehaviorSpec(shapes, ActionType.CONTINUOUS, 3)
+    behavior_spec = BehaviorSpec(shapes, ActionSpec.create_continuous(3))
     ap_list = generate_list_agent_proto(n_agents, shapes, infinite_rewards=True)
     with pytest.raises(RuntimeError):
         steps_from_proto(ap_list, behavior_spec)
     n_agents = 10
     shapes = [(3,), (4,)]
-    behavior_spec = BehaviorSpec(shapes, ActionType.CONTINUOUS, 3)
+    behavior_spec = BehaviorSpec(shapes, ActionSpec.create_continuous(3))
     ap_list = generate_list_agent_proto(n_agents, shapes, nan_observations=True)
     with pytest.raises(RuntimeError):
         steps_from_proto(ap_list, behavior_spec)

ml-agents-envs/mlagents_envs/tests/test_steps.py

 from mlagents_envs.base_env import (
     DecisionSteps,
     TerminalSteps,
-    ActionType,
+    ActionSpec,
     BehaviorSpec,
 )
 
 
 def test_empty_decision_steps():
     specs = BehaviorSpec(
-        observation_shapes=[(3, 2), (5,)],
-        action_type=ActionType.CONTINUOUS,
-        action_shape=3,
+        observation_shapes=[(3, 2), (5,)], action_spec=ActionSpec.create_continuous(3)
     )
     ds = DecisionSteps.empty(specs)
     assert len(ds.obs) == 2
 
 def test_empty_terminal_steps():
     specs = BehaviorSpec(
-        observation_shapes=[(3, 2), (5,)],
-        action_type=ActionType.CONTINUOUS,
-        action_shape=3,
+        observation_shapes=[(3, 2), (5,)], action_spec=ActionSpec.create_continuous(3)
     )
     ts = TerminalSteps.empty(specs)
     assert len(ts.obs) == 2
 
 def test_specs():
-    specs = BehaviorSpec(
-        observation_shapes=[(3, 2), (5,)],
-        action_type=ActionType.CONTINUOUS,
-        action_shape=3,
-    )
-    assert specs.discrete_action_branches is None
-    assert specs.action_size == 3
-    assert specs.create_empty_action(5).shape == (5, 3)
-    assert specs.create_empty_action(5).dtype == np.float32
+    specs = ActionSpec.create_continuous(3)
+    assert specs.discrete_branches == ()
+    assert specs.discrete_size == 0
+    assert specs.continuous_size == 3
+    assert specs.empty_action(5).shape == (5, 3)
+    assert specs.empty_action(5).dtype == np.float32
-    specs = BehaviorSpec(
-        observation_shapes=[(3, 2), (5,)],
-        action_type=ActionType.DISCRETE,
-        action_shape=(3,),
-    )
-    assert specs.discrete_action_branches == (3,)
-    assert specs.action_size == 1
-    assert specs.create_empty_action(5).shape == (5, 1)
-    assert specs.create_empty_action(5).dtype == np.int32
+    specs = ActionSpec.create_discrete((3,))
+    assert specs.discrete_branches == (3,)
+    assert specs.discrete_size == 1
+    assert specs.continuous_size == 0
+    assert specs.empty_action(5).shape == (5, 1)
+    assert specs.empty_action(5).dtype == np.int32
-    specs = BehaviorSpec(
-        observation_shapes=[(5,)],
-        action_type=ActionType.CONTINUOUS,
-        action_shape=action_len,
-    )
-    zero_action = specs.create_empty_action(4)
+    specs = ActionSpec.create_continuous(action_len)
+    zero_action = specs.empty_action(4)
-    random_action = specs.create_random_action(4)
+    random_action = specs.random_action(4)
     assert random_action.dtype == np.float32
     assert random_action.shape == (4, action_len)
     assert np.min(random_action) >= -1
     action_shape = (10, 20, 30)
-    specs = BehaviorSpec(
-        observation_shapes=[(5,)],
-        action_type=ActionType.DISCRETE,
-        action_shape=action_shape,
-    )
-    zero_action = specs.create_empty_action(4)
+    specs = ActionSpec.create_discrete(action_shape)
+    zero_action = specs.empty_action(4)
-    random_action = specs.create_random_action(4)
+    random_action = specs.random_action(4)
     assert random_action.dtype == np.int32
     assert random_action.shape == (4, len(action_shape))
     assert np.min(random_action) >= 0

ml-agents/mlagents/tf_utils/__init__.py

 from mlagents.tf_utils.tf import tf as tf  # noqa
 from mlagents.tf_utils.tf import set_warnings_enabled  # noqa
 from mlagents.tf_utils.tf import generate_session_config  # noqa
+from mlagents.tf_utils.tf import is_available  # noqa

ml-agents/mlagents/tf_utils/tf.py

 # This should be the only place that we import tensorflow directly.
 # Everywhere else is caught by the banned-modules setting for flake8
-import tensorflow as tf  # noqa I201
+
+try:
+    import tensorflow as tf  # noqa I201
-# LooseVersion handles things "1.2.3a" or "4.5.6-rc7" fairly sensibly.
-_is_tensorflow2 = LooseVersion(tf.__version__) >= LooseVersion("2.0.0")
+    # LooseVersion handles things "1.2.3a" or "4.5.6-rc7" fairly sensibly.
+    _is_tensorflow2 = LooseVersion(tf.__version__) >= LooseVersion("2.0.0")
-if _is_tensorflow2:
-    import tensorflow.compat.v1 as tf
+    if _is_tensorflow2:
+        import tensorflow.compat.v1 as tf
-    tf.disable_v2_behavior()
-    tf_logging = tf.logging
-else:
-    try:
-        # Newer versions of tf 1.x will complain that tf.logging is deprecated
-        tf_logging = tf.compat.v1.logging
-    except AttributeError:
-        # Fall back to the safe import, even if it might generate a warning or two.
+        tf.disable_v2_behavior()
+    else:
+        try:
+            # Newer versions of tf 1.x will complain that tf.logging is deprecated
+            tf_logging = tf.compat.v1.logging
+        except AttributeError:
+            # Fall back to the safe import, even if it might generate a warning or two.
+            tf_logging = tf.logging
+except ImportError:
+    tf = None
+
+
+def is_available():
+    """
+    Returns whether Torch is available in this Python environment
+    """
+    return tf is not None
 
 
 def set_warnings_enabled(is_enabled: bool) -> None:
     """
-    level = tf_logging.WARN if is_enabled else tf_logging.ERROR
-    tf_logging.set_verbosity(level)
+    if is_available():
+        level = tf_logging.WARN if is_enabled else tf_logging.ERROR
+        tf_logging.set_verbosity(level)
-def generate_session_config() -> tf.ConfigProto:
+def generate_session_config() -> "tf.ConfigProto":
-    config = tf.ConfigProto()
-    config.gpu_options.allow_growth = True
-    # For multi-GPU training, set allow_soft_placement to True to allow
-    # placing the operation into an alternative device automatically
-    # to prevent from exceptions if the device doesn't suppport the operation
-    # or the device does not exist
-    config.allow_soft_placement = True
-    return config
+    if is_available():
+        config = tf.ConfigProto()
+        config.gpu_options.allow_growth = True
+        # For multi-GPU training, set allow_soft_placement to True to allow
+        # placing the operation into an alternative device automatically
+        # to prevent from exceptions if the device doesn't suppport the operation
+        # or the device does not exist
+        config.allow_soft_placement = True
+        return config
+    else:
+        return None

ml-agents/mlagents/torch_utils/__init__.py

 from mlagents.torch_utils.torch import torch as torch  # noqa
 from mlagents.torch_utils.torch import nn  # noqa
-from mlagents.torch_utils.torch import is_available  # noqa
 from mlagents.torch_utils.torch import default_device  # noqa

ml-agents/mlagents/torch_utils/torch.py

 import os
 
+from distutils.version import LooseVersion
+import pkg_resources
-# Detect availability of torch package here.
-# NOTE: this try/except is temporary until torch is required for ML-Agents.
-try:
-    # This should be the only place that we import torch directly.
-    # Everywhere else is caught by the banned-modules setting for flake8
-    import torch  # noqa I201
-    torch.set_num_threads(cpu_utils.get_num_threads_to_use())
-    os.environ["KMP_BLOCKTIME"] = "0"
+def assert_torch_installed():
+    # Check that torch version 1.6.0 or later has been installed. If not, refer
+    # user to the PyTorch webpage for install instructions.
+    torch_pkg = None
+    try:
+        torch_pkg = pkg_resources.get_distribution("torch")
+    except pkg_resources.DistributionNotFound:
+        pass
+    assert torch_pkg is not None and LooseVersion(torch_pkg.version) >= LooseVersion(
+        "1.6.0"
+    ), (
+        "A compatible version of PyTorch was not installed. Please visit the PyTorch homepage "
+        + "(https://pytorch.org/get-started/locally/) and follow the instructions to install. "
+        + "Version 1.6.0 and later are supported."
+    )
-    # Known PyLint compatibility with PyTorch https://github.com/pytorch/pytorch/issues/701
-    # pylint: disable=E1101
-    if torch.cuda.is_available():
-        torch.set_default_tensor_type(torch.cuda.FloatTensor)
-        device = torch.device("cuda")
-    else:
-        torch.set_default_tensor_type(torch.FloatTensor)
-        device = torch.device("cpu")
-    nn = torch.nn
-    # pylint: disable=E1101
-except ImportError:
-    torch = None
-    nn = None
-    device = None
+
+assert_torch_installed()
+
+# This should be the only place that we import torch directly.
+# Everywhere else is caught by the banned-modules setting for flake8
+import torch  # noqa I201
+
+
+torch.set_num_threads(cpu_utils.get_num_threads_to_use())
+os.environ["KMP_BLOCKTIME"] = "0"
+
+# Known PyLint compatibility with PyTorch https://github.com/pytorch/pytorch/issues/701
+# pylint: disable=E1101
+if torch.cuda.is_available():
+    torch.set_default_tensor_type(torch.cuda.FloatTensor)
+    device = torch.device("cuda")
+else:
+    torch.set_default_tensor_type(torch.FloatTensor)
+    device = torch.device("cpu")
+nn = torch.nn
-
-
-def is_available():
-    """
-    Returns whether Torch is available in this Python environment
-    """
-    return torch is not None

ml-agents/mlagents/trainers/__init__.py

 # Version of the library that will be used to upload to pypi
-__version__ = "0.21.1"
+__version__ = "0.22.0.dev0"
-__release_tag__ = "release_9"
+__release_tag__ = None

ml-agents/mlagents/trainers/cli_utils.py

         "--torch",
         default=False,
         action=DetectDefaultStoreTrue,
-        help="(Experimental) Use the PyTorch framework instead of TensorFlow. Install PyTorch "
-        "before using this option",
+        help="Use the PyTorch framework. Note that this option is not required anymore as PyTorch is the"
+        "default framework, and will be removed in the next release.",
+    )
+    argparser.add_argument(
+        "--tensorflow",
+        default=False,
+        action=DetectDefaultStoreTrue,
+        help="(Deprecated) Use the TensorFlow framework instead of PyTorch. Install TensorFlow "
+        "before using this option.",
     )
 
     eng_conf = argparser.add_argument_group(title="Engine Configuration")

ml-agents/mlagents/trainers/demo_loader.py

     demo_buffer = make_demo_buffer(info_action_pair, behavior_spec, sequence_length)
     if expected_behavior_spec:
         # check action dimensions in demonstration match
-        if behavior_spec.action_shape != expected_behavior_spec.action_shape:
+        if behavior_spec.action_spec != expected_behavior_spec.action_spec:
-                "The action dimensions {} in demonstration do not match the policy's {}.".format(
-                    behavior_spec.action_shape, expected_behavior_spec.action_shape
-                )
-            )
-        # check the action types in demonstration match
-        if behavior_spec.action_type != expected_behavior_spec.action_type:
-            raise RuntimeError(
-                "The action type of {} in demonstration do not match the policy's {}.".format(
-                    behavior_spec.action_type, expected_behavior_spec.action_type
+                "The action spaces {} in demonstration do not match the policy's {}.".format(
+                    behavior_spec.action_spec, expected_behavior_spec.action_spec
                 )
             )
         # check observations match

ml-agents/mlagents/trainers/learn.py

 # # Unity ML-Agents Toolkit
+from mlagents import torch_utils
 import yaml
 
 import os
   ml-agents: {mlagents.trainers.__version__},
   ml-agents-envs: {mlagents_envs.__version__},
   Communicator API: {UnityEnvironment.API_VERSION},
-  TensorFlow: {tf_utils.tf.__version__}"""
+  PyTorch: {torch_utils.torch.__version__}"""
 
 
 def parse_command_line(argv: Optional[List[str]] = None) -> RunOptions:
             init_path=maybe_init_path,
             multi_gpu=False,
             force_torch="torch" in DetectDefault.non_default_args,
+            force_tensorflow="tensorflow" in DetectDefault.non_default_args,
         )
         # Create controller and begin training.
         tc = TrainerController(
     add_timer_metadata("mlagents_version", mlagents.trainers.__version__)
     add_timer_metadata("mlagents_envs_version", mlagents_envs.__version__)
     add_timer_metadata("communication_protocol_version", UnityEnvironment.API_VERSION)
-    add_timer_metadata("tensorflow_version", tf_utils.tf.__version__)
+    add_timer_metadata("pytorch_version", torch_utils.torch.__version__)
     add_timer_metadata("numpy_version", np.__version__)
 
     if options.env_settings.seed == -1:

ml-agents/mlagents/trainers/policy/policy.py

         self.trainer_settings = trainer_settings
         self.network_settings: NetworkSettings = trainer_settings.network_settings
         self.seed = seed
+        if (
+            self.behavior_spec.action_spec.continuous_size > 0
+            and self.behavior_spec.action_spec.discrete_size > 0
+        ):
+            raise UnityPolicyException("Trainers do not support mixed action spaces.")
-            list(behavior_spec.discrete_action_branches)
-            if behavior_spec.is_action_discrete()
-            else [behavior_spec.action_size]
+            list(self.behavior_spec.action_spec.discrete_branches)
+            if self.behavior_spec.action_spec.is_discrete()
+            else [self.behavior_spec.action_spec.continuous_size]
         )
         self.vec_obs_size = sum(
             shape[0] for shape in behavior_spec.observation_shapes if len(shape) == 1
         )
-        self.use_continuous_act = behavior_spec.is_action_continuous()
-        self.num_branches = self.behavior_spec.action_size
+        self.use_continuous_act = self.behavior_spec.action_spec.is_continuous()
+        # This line will be removed in the ActionBuffer change
+        self.num_branches = (
+            self.behavior_spec.action_spec.continuous_size
+            + self.behavior_spec.action_spec.discrete_size
+        )
         self.previous_action_dict: Dict[str, np.array] = {}
         self.memory_dict: Dict[str, np.ndarray] = {}
         self.normalize = trainer_settings.network_settings.normalize

ml-agents/mlagents/trainers/policy/tf_policy.py

             mask = np.ones(
                 (
                     len(batched_step_result),
-                    sum(self.behavior_spec.discrete_action_branches),
+                    sum(self.behavior_spec.action_spec.discrete_branches),
                 ),
                 dtype=np.float32,
             )
             self.mask = tf.cast(self.mask_input, tf.int32)
 
             tf.Variable(
-                int(self.behavior_spec.is_action_continuous()),
+                int(self.behavior_spec.action_spec.is_continuous()),
                 name="is_continuous_control",
                 trainable=False,
                 dtype=tf.int32,
             tf.Variable(
                 self.m_size, name="memory_size", trainable=False, dtype=tf.int32
             )
-            if self.behavior_spec.is_action_continuous():
+            if self.behavior_spec.action_spec.is_continuous():
                 tf.Variable(
                     self.act_size[0],
                     name="action_output_shape",

ml-agents/mlagents/trainers/policy/torch_policy.py

         self.actor_critic = ac_class(
             observation_shapes=self.behavior_spec.observation_shapes,
             network_settings=trainer_settings.network_settings,
-            act_type=behavior_spec.action_type,
-            act_size=self.act_size,
+            action_spec=behavior_spec.action_spec,
             stream_names=reward_signal_names,
             conditional_sigma=self.condition_sigma_on_obs,
             tanh_squash=tanh_squash,
             actions = actions[:, :, 0]
         else:
             actions = actions[:, 0, :]
-
-        return (actions, all_logs if all_log_probs else log_probs, entropies, memories)
+        # Use the sum of entropy across actions, not the mean
+        entropy_sum = torch.sum(entropies, dim=1)
+        return (
+            actions,
+            all_logs if all_log_probs else log_probs,
+            entropy_sum,
+            memories,
+        )
 
     def evaluate_actions(
         self,
         )
         action_list = [actions[..., i] for i in range(actions.shape[-1])]
         log_probs, entropies, _ = ModelUtils.get_probs_and_entropy(action_list, dists)
-
-        return log_probs, entropies, value_heads
+        # Use the sum of entropy across actions, not the mean
+        entropy_sum = torch.sum(entropies, dim=1)
+        return log_probs, entropy_sum, value_heads
 
     @timed
     def evaluate(

ml-agents/mlagents/trainers/ppo/optimizer_torch.py

 
         self.optimizer.step()
         update_stats = {
-            "Losses/Policy Loss": policy_loss.item(),
+            # NOTE: abs() is not technically correct, but matches the behavior in TensorFlow.
+            # TODO: After PyTorch is default, change to something more correct.
+            "Losses/Policy Loss": torch.abs(policy_loss).item(),
             "Losses/Value Loss": value_loss.item(),
             "Policy/Learning Rate": decay_lr,
             "Policy/Epsilon": decay_eps,

ml-agents/mlagents/trainers/ppo/trainer.py

 from mlagents_envs.base_env import BehaviorSpec
 from mlagents.trainers.trainer.rl_trainer import RLTrainer
 from mlagents.trainers.policy import Policy
-from mlagents.trainers.policy.tf_policy import TFPolicy
-from mlagents.trainers.ppo.optimizer_tf import PPOOptimizer
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.ppo.optimizer_torch import TorchPPOOptimizer
-from mlagents.trainers.tf.components.reward_signals import RewardSignal
-from mlagents import torch_utils
+from mlagents.trainers.torch.components.reward_providers.base_reward_provider import (
+    BaseRewardProvider,
+)
+from mlagents import tf_utils
-if torch_utils.is_available():
-    from mlagents.trainers.policy.torch_policy import TorchPolicy
-    from mlagents.trainers.ppo.optimizer_torch import TorchPPOOptimizer
+if tf_utils.is_available():
+    from mlagents.trainers.policy.tf_policy import TFPolicy
+    from mlagents.trainers.ppo.optimizer_tf import PPOOptimizer
-    TorchPolicy = None  # type: ignore
-    TorchPPOOptimizer = None  # type: ignore
+    TFPolicy = None  # type: ignore
+    PPOOptimizer = None  # type: ignore
 
 
 logger = get_logger(__name__)
 
         for name, v in value_estimates.items():
             agent_buffer_trajectory[f"{name}_value_estimates"].extend(v)
-            if isinstance(self.optimizer.reward_signals[name], RewardSignal):
+            if isinstance(self.optimizer.reward_signals[name], BaseRewardProvider):
-                    self.optimizer.reward_signals[name].value_name, np.mean(v)
+                    f"Policy/{self.optimizer.reward_signals[name].name.capitalize()} Value Estimate",
+                    np.mean(v),
-                    f"Policy/{self.optimizer.reward_signals[name].name.capitalize()} Value Estimate",
-                    np.mean(v),
+                    self.optimizer.reward_signals[name].value_name, np.mean(v)
                 )
 
         # Evaluate all reward functions
         for name, reward_signal in self.optimizer.reward_signals.items():
-            if isinstance(reward_signal, RewardSignal):
-                evaluate_result = reward_signal.evaluate_batch(
-                    agent_buffer_trajectory
-                ).scaled_reward
-            else:
+            # BaseRewardProvider is a PyTorch-based reward signal
+            if isinstance(reward_signal, BaseRewardProvider):
+            else:  # reward_signal is a TensorFlow-based RewardSignal class
+                evaluate_result = reward_signal.evaluate_batch(
+                    agent_buffer_trajectory
+                ).scaled_reward
             agent_buffer_trajectory[f"{name}_rewards"].extend(evaluate_result)
             # Report the reward signals
             self.collected_rewards[name][agent_id] += np.sum(evaluate_result)
             behavior_spec,
             self.trainer_settings,
             condition_sigma_on_obs=False,  # Faster training for PPO
-            separate_critic=behavior_spec.is_action_continuous(),
+            separate_critic=behavior_spec.action_spec.is_continuous(),
         )
         return policy
 

ml-agents/mlagents/trainers/sac/optimizer_torch.py

 from mlagents.torch_utils import torch, nn, default_device
 
 from mlagents_envs.logging_util import get_logger
-from mlagents_envs.base_env import ActionType
+from mlagents_envs.base_env import ActionSpec
 from mlagents.trainers.optimizer.torch_optimizer import TorchOptimizer
 from mlagents.trainers.policy.torch_policy import TorchPolicy
 from mlagents.trainers.settings import NetworkSettings
             stream_names: List[str],
             observation_shapes: List[Tuple[int, ...]],
             network_settings: NetworkSettings,
-            act_type: ActionType,
-            act_size: List[int],
+            action_spec: ActionSpec,
-            if act_type == ActionType.CONTINUOUS:
+            self.action_spec = action_spec
+            if self.action_spec.is_continuous():
+                self.act_size = self.action_spec.continuous_size
-                num_action_ins = sum(act_size)
+                num_action_ins = self.act_size
+
-                num_value_outs = sum(act_size)
+                self.act_size = self.action_spec.discrete_branches
+                num_value_outs = sum(self.act_size)
                 num_action_ins = 0
             self.q1_network = ValueNetwork(
                 stream_names,
             self.stream_names,
             self.policy.behavior_spec.observation_shapes,
             policy_network_settings,
-            self.policy.behavior_spec.action_type,
-            self.act_size,
+            self.policy.behavior_spec.action_spec,
         )
 
         self.target_network = ValueNetwork(
                     for i, (_lp, _qt) in enumerate(
                         zip(branched_per_action_ent, branched_q_term)
                     )
-                ]
+                ],
+                dim=1,
-        policy_loss = torch.mean(loss_masks * batch_policy_loss)
+            policy_loss = ModelUtils.masked_mean(batch_policy_loss, loss_masks)
         return policy_loss
 
     def sac_entropy_loss(
             with torch.no_grad():
                 target_current_diff = torch.sum(log_probs + self.target_entropy, dim=1)
-            entropy_loss = -torch.mean(
-                self._log_ent_coef * loss_masks * target_current_diff
+            entropy_loss = -1 * ModelUtils.masked_mean(
+                self._log_ent_coef * target_current_diff, loss_masks
             )
         else:
             with torch.no_grad():

ml-agents/mlagents/trainers/sac/trainer.py

 from mlagents_envs.logging_util import get_logger
 from mlagents_envs.timers import timed
 from mlagents_envs.base_env import BehaviorSpec
-from mlagents.trainers.policy.tf_policy import TFPolicy
-from mlagents.trainers.sac.optimizer_tf import SACOptimizer
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.sac.optimizer_torch import TorchSACOptimizer
-from mlagents.trainers.tf.components.reward_signals import RewardSignal
-from mlagents import torch_utils
+from mlagents.trainers.torch.components.reward_providers import BaseRewardProvider
+from mlagents import tf_utils
-if torch_utils.is_available():
-    from mlagents.trainers.policy.torch_policy import TorchPolicy
-    from mlagents.trainers.sac.optimizer_torch import TorchSACOptimizer
+if tf_utils.is_available():
+    from mlagents.trainers.policy.tf_policy import TFPolicy
+    from mlagents.trainers.sac.optimizer_tf import SACOptimizer
-    TorchPolicy = None  # type: ignore
-    TorchSACOptimizer = None  # type: ignore
+    TFPolicy = None  # type: ignore
+    SACOptimizer = None  # type: ignore
 
 logger = get_logger(__name__)
 
 
         self.seed = seed
         self.policy: Policy = None  # type: ignore
-        self.optimizer: SACOptimizer = None  # type: ignore
+        self.optimizer: TorchSACOptimizer = None  # type: ignore
         self.hyperparameters: SACSettings = cast(
             SACSettings, trainer_settings.hyperparameters
         )
             agent_buffer_trajectory["environment_rewards"]
         )
         for name, reward_signal in self.optimizer.reward_signals.items():
-            if isinstance(reward_signal, RewardSignal):
-                evaluate_result = reward_signal.evaluate_batch(
-                    agent_buffer_trajectory
-                ).scaled_reward
-            else:
+            # BaseRewardProvider is a PyTorch-based reward signal
+            if isinstance(reward_signal, BaseRewardProvider):
+            else:  # reward_signal uses TensorFlow
+                evaluate_result = reward_signal.evaluate_batch(
+                    agent_buffer_trajectory
+                ).scaled_reward
+
             # Report the reward signals
             self.collected_rewards[name][agent_id] += np.sum(evaluate_result)
 
         )
         for name, v in value_estimates.items():
-            if isinstance(self.optimizer.reward_signals[name], RewardSignal):
-                self._stats_reporter.add_stat(
-                    self.optimizer.reward_signals[name].value_name, np.mean(v)
-                )
-            else:
+            # BaseRewardProvider is a PyTorch-based reward signal
+            if isinstance(self.optimizer.reward_signals[name], BaseRewardProvider):
+                )
+            else:  # TensorFlow reward signal
+                self._stats_reporter.add_stat(
+                    self.optimizer.reward_signals[name].value_name, np.mean(v)
                 )
 
         # Bootstrap using the last step rather than the bootstrap step if max step is reached.
                 )
                 # Get rewards for each reward
                 for name, signal in self.optimizer.reward_signals.items():
-                    if isinstance(signal, RewardSignal):
+                    # BaseRewardProvider is a PyTorch-based reward signal
+                    if isinstance(signal, BaseRewardProvider):
+                        sampled_minibatch[f"{name}_rewards"] = (
+                            signal.evaluate(sampled_minibatch) * signal.strength
+                        )
+                    else:  # reward_signal is a TensorFlow-based RewardSignal class
-                    else:
-                        sampled_minibatch[f"{name}_rewards"] = (
-                            signal.evaluate(sampled_minibatch) * signal.strength
-                        )
 
                 update_stats = self.optimizer.update(sampled_minibatch, n_sequences)
                 for stat_name, value in update_stats.items():
             reward_signal_minibatches = {}
             for name, signal in self.optimizer.reward_signals.items():
                 logger.debug(f"Updating {name} at step {self.step}")
-                if isinstance(signal, RewardSignal):
+                # BaseRewardProvider is a PyTorch-based reward signal
+                if not isinstance(signal, BaseRewardProvider):
                     # Some signals don't need a minibatch to be sampled - so we don't!
                     if signal.update_dict:
                         reward_signal_minibatches[name] = buffer.sample_mini_batch(
-                else:
+                else:  # TensorFlow reward signal
                     if name != "extrinsic":
                         reward_signal_minibatches[name] = buffer.sample_mini_batch(
                             self.hyperparameters.batch_size,
             for stat, stat_list in batch_update_stats.items():
                 self._stats_reporter.add_stat(stat, np.mean(stat_list))
 
-    def create_sac_optimizer(self) -> SACOptimizer:
+    def create_sac_optimizer(self) -> TorchSACOptimizer:
         if self.framework == FrameworkType.PYTORCH:
             return TorchSACOptimizer(  # type: ignore
                 cast(TorchPolicy, self.policy), self.trainer_settings  # type: ignore

ml-agents/mlagents/trainers/settings.py

     threaded: bool = True
     self_play: Optional[SelfPlaySettings] = None
     behavioral_cloning: Optional[BehavioralCloningSettings] = None
-    framework: FrameworkType = FrameworkType.TENSORFLOW
+    framework: FrameworkType = FrameworkType.PYTORCH
 
     cattr.register_structure_hook(
         Dict[RewardSignalType, RewardSignalSettings], RewardSignalSettings.structure

ml-agents/mlagents/trainers/stats.py

 from collections import defaultdict
 from enum import Enum
-from typing import List, Dict, NamedTuple, Any, Optional
+from typing import List, Dict, NamedTuple, Any
 import numpy as np
 import abc
 import os
 from mlagents_envs.logging_util import get_logger
 from mlagents_envs.timers import set_gauge
-from mlagents.tf_utils import tf, generate_session_config
+from torch.utils.tensorboard import SummaryWriter
+
+
+def _dict_to_str(param_dict: Dict[str, Any], num_tabs: int) -> str:
+    """
+    Takes a parameter dictionary and converts it to a human-readable string.
+    Recurses if there are multiple levels of dict. Used to print out hyperparameters.
+    param: param_dict: A Dictionary of key, value parameters.
+    return: A string version of this dictionary.
+    """
+    if not isinstance(param_dict, dict):
+        return str(param_dict)
+    else:
+        append_newline = "\n" if num_tabs > 0 else ""
+        return append_newline + "\n".join(
+            [
+                "\t"
+                + "  " * num_tabs
+                + "{}:\t{}".format(x, _dict_to_str(param_dict[x], num_tabs + 1))
+                for x in param_dict
+            ]
+        )
 
 
 class StatsSummary(NamedTuple):
         if property_type == StatsPropertyType.HYPERPARAMETERS:
             logger.info(
                 """Hyperparameters for behavior name {}: \n{}""".format(
-                    category, self._dict_to_str(value, 0)
+                    category, _dict_to_str(value, 0)
                 )
             )
         elif property_type == StatsPropertyType.SELF_PLAY:
-    def _dict_to_str(self, param_dict: Dict[str, Any], num_tabs: int) -> str:
-        """
-        Takes a parameter dictionary and converts it to a human-readable string.
-        Recurses if there are multiple levels of dict. Used to print out hyperparameters.
-        param: param_dict: A Dictionary of key, value parameters.
-        return: A string version of this dictionary.
-        """
-        if not isinstance(param_dict, dict):
-            return str(param_dict)
-        else:
-            append_newline = "\n" if num_tabs > 0 else ""
-            return append_newline + "\n".join(
-                [
-                    "\t"
-                    + "  " * num_tabs
-                    + "{}:\t{}".format(
-                        x, self._dict_to_str(param_dict[x], num_tabs + 1)
-                    )
-                    for x in param_dict
-                ]
-            )
-
 
 class TensorboardWriter(StatsWriter):
     def __init__(self, base_dir: str, clear_past_data: bool = False):
         :param clear_past_data: Whether or not to clean up existing Tensorboard files associated with the base_dir and
             category.
         """
-        self.summary_writers: Dict[str, tf.summary.FileWriter] = {}
+        self.summary_writers: Dict[str, SummaryWriter] = {}
         self.base_dir: str = base_dir
         self._clear_past_data = clear_past_data
 
         self._maybe_create_summary_writer(category)
         for key, value in values.items():
-            summary = tf.Summary()
-            summary.value.add(tag=f"{key}", simple_value=value.mean)
-            self.summary_writers[category].add_summary(summary, step)
+            self.summary_writers[category].add_scalar(f"{key}", value.mean, step)
             self.summary_writers[category].flush()
 
     def _maybe_create_summary_writer(self, category: str) -> None:
             os.makedirs(filewriter_dir, exist_ok=True)
             if self._clear_past_data:
                 self._delete_all_events_files(filewriter_dir)
-            self.summary_writers[category] = tf.summary.FileWriter(filewriter_dir)
+            self.summary_writers[category] = SummaryWriter(filewriter_dir)
 
     def _delete_all_events_files(self, directory_name: str) -> None:
         for file_name in os.listdir(directory_name):
     ) -> None:
         if property_type == StatsPropertyType.HYPERPARAMETERS:
             assert isinstance(value, dict)
-            summary = self._dict_to_tensorboard("Hyperparameters", value)
+            summary = _dict_to_str(value, 0)
-                self.summary_writers[category].add_summary(summary, 0)
-
-    def _dict_to_tensorboard(
-        self, name: str, input_dict: Dict[str, Any]
-    ) -> Optional[bytes]:
-        """
-        Convert a dict to a Tensorboard-encoded string.
-        :param name: The name of the text.
-        :param input_dict: A dictionary that will be displayed in a table on Tensorboard.
-        """
-        try:
-            with tf.Session(config=generate_session_config()) as sess:
-                s_op = tf.summary.text(
-                    name,
-                    tf.convert_to_tensor(
-                        [[str(x), str(input_dict[x])] for x in input_dict]
-                    ),
-                )
-                s = sess.run(s_op)
-                return s
-        except Exception:
-            logger.warning(
-                f"Could not write {name} summary for Tensorboard: {input_dict}"
-            )
-            return None
+                self.summary_writers[category].add_text("Hyperparameters", summary)
+                self.summary_writers[category].flush()
 
 
 class StatsReporter:

ml-agents/mlagents/trainers/tests/mock_brain.py

-from typing import List, Tuple, Union
-from collections.abc import Iterable
+from typing import List, Tuple
 import numpy as np
 
 from mlagents.trainers.buffer import AgentBuffer
     TerminalSteps,
     BehaviorSpec,
-    ActionType,
+    ActionSpec,
 )
 
 
-    action_shape: Union[int, Tuple[int]] = None,
-    discrete: bool = False,
+    action_spec: ActionSpec,
     done: bool = False,
 ) -> Tuple[DecisionSteps, TerminalSteps]:
     """
     :bool discrete: Whether or not action space is discrete
     :bool done: Whether all the agents in the batch are done
     """
-    if action_shape is None:
-        action_shape = 2
-
-    if discrete and isinstance(action_shape, Iterable):
+    if action_spec.is_discrete():
-            for action_size in action_shape  # type: ignore
+            for action_size in action_spec.discrete_branches  # type: ignore
-    behavior_spec = BehaviorSpec(
-        observation_shapes,
-        ActionType.DISCRETE if discrete else ActionType.CONTINUOUS,
-        action_shape,
-    )
+    behavior_spec = BehaviorSpec(observation_shapes, action_spec)
     if done:
         return (
             DecisionSteps.empty(behavior_spec),
     return create_mock_steps(
         num_agents=num_agents,
         observation_shapes=behavior_spec.observation_shapes,
-        action_shape=behavior_spec.action_shape,
-        discrete=behavior_spec.is_action_discrete(),
+        action_spec=behavior_spec.action_spec,
     )
 
 
+    action_spec: ActionSpec,
-    action_space: Union[int, Tuple[int]] = 2,
-    is_discrete: bool = True,
 ) -> Trajectory:
     """
     Makes a fake trajectory of length length. If max_step_complete,
+
+    action_size = action_spec.discrete_size + action_spec.continuous_size
+    action_probs = np.ones(
+        int(np.sum(action_spec.discrete_branches) + action_spec.continuous_size),
+        dtype=np.float32,
+    )
     for _i in range(length - 1):
         obs = []
         for _shape in observation_shapes:
-        if is_discrete:
-            action_size = len(action_space)  # type: ignore
-            action_probs = np.ones(np.sum(action_space), dtype=np.float32)
-        else:
-            action_size = int(action_space)  # type: ignore
-            action_probs = np.ones((action_size), dtype=np.float32)
-            [[False for _ in range(branch)] for branch in action_space]  # type: ignore
-            if is_discrete
+            [
+                [False for _ in range(branch)]
+                for branch in action_spec.discrete_branches
+            ]  # type: ignore
+            if action_spec.is_discrete()
             else None
         )
         prev_action = np.ones(action_size, dtype=np.float32)
     memory_size: int = 10,
     exclude_key_list: List[str] = None,
 ) -> AgentBuffer:
-    action_space = behavior_spec.action_shape
-    is_discrete = behavior_spec.is_action_discrete()
-
-        action_space=action_space,
+        action_spec=behavior_spec.action_spec,
-        is_discrete=is_discrete,
     )
     buffer = trajectory.to_agentbuffer()
     # If a key_list was given, remove those keys
 def setup_test_behavior_specs(
     use_discrete=True, use_visual=False, vector_action_space=2, vector_obs_space=8
 ):
+    if use_discrete:
+        action_spec = ActionSpec.create_discrete(tuple(vector_action_space))
+    else:
+        action_spec = ActionSpec.create_continuous(vector_action_space)
-        [(84, 84, 3)] * int(use_visual) + [(vector_obs_space,)],
-        ActionType.DISCRETE if use_discrete else ActionType.CONTINUOUS,
-        tuple(vector_action_space) if use_discrete else vector_action_space,
+        [(84, 84, 3)] * int(use_visual) + [(vector_obs_space,)], action_spec
     )
     return behavior_spec
 

ml-agents/mlagents/trainers/tests/simple_test_envs.py

 import numpy as np
 
 from mlagents_envs.base_env import (
+    ActionSpec,
-    ActionType,
     BehaviorMapping,
 )
 from mlagents_envs.tests.test_rpc_utils import proto_from_steps_and_action
         self.num_vector = num_vector
         self.vis_obs_size = vis_obs_size
         self.vec_obs_size = vec_obs_size
-        action_type = ActionType.DISCRETE if use_discrete else ActionType.CONTINUOUS
-        self.behavior_spec = BehaviorSpec(
-            self._make_obs_spec(),
-            action_type,
-            tuple(2 for _ in range(action_size)) if use_discrete else action_size,
-        )
+        if use_discrete:
+            action_spec = ActionSpec.create_discrete(
+                tuple(2 for _ in range(action_size))
+            )
+        else:
+            action_spec = ActionSpec.create_continuous(action_size)
+        self.behavior_spec = BehaviorSpec(self._make_obs_spec(), action_spec)
         self.action_size = action_size
         self.names = brain_names
         self.positions: Dict[str, List[float]] = {}

ml-agents/mlagents/trainers/tests/tensorflow/test_ghost.py

         length=time_horizon,
         max_step_complete=True,
         observation_shapes=[(1,)],
-        action_space=[2],
+        action_spec=mock_specs.action_spec,
     )
     trajectory_queue0.put(trajectory)
     trainer.advance()

ml-agents/mlagents/trainers/tests/tensorflow/test_models.py

 
 from mlagents.trainers.tf.models import ModelUtils
 from mlagents.tf_utils import tf
-from mlagents_envs.base_env import BehaviorSpec, ActionType
+from mlagents_envs.base_env import BehaviorSpec, ActionSpec
-        ActionType.DISCRETE,
-        (1,),
+        ActionSpec.create_discrete((1,)),
     )
     return behavior_spec
 

ml-agents/mlagents/trainers/tests/tensorflow/test_nn_policy.py

         length=time_horizon,
         max_step_complete=True,
         observation_shapes=[(1,)],
-        action_space=[2],
+        action_spec=behavior_spec.action_spec,
     )
     for i in range(time_horizon):
         trajectory.steps[i].obs[0] = np.array([large_obs1[i]], dtype=np.float32)
         length=time_horizon,
         max_step_complete=True,
         observation_shapes=[(1,)],
-        action_space=[2],
+        action_spec=behavior_spec.action_spec,
     )
     for i in range(time_horizon):
         trajectory.steps[i].obs[0] = np.array([large_obs2[i]], dtype=np.float32)
         length=time_horizon,
         max_step_complete=True,
         observation_shapes=[(1,)],
-        action_space=[2],
+        action_spec=behavior_spec.action_spec,
     )
     # Change half of the obs to 0
     for i in range(3):
         length=time_horizon,
         max_step_complete=True,
         observation_shapes=[(1,)],
-        action_space=[2],
+        action_spec=behavior_spec.action_spec,
     )
     trajectory_buffer = trajectory.to_agentbuffer()
     policy.update_normalization(trajectory_buffer["vector_obs"])

ml-agents/mlagents/trainers/tests/tensorflow/test_ppo.py

     ppo_dummy_config,
 )
 
+from mlagents_envs.base_env import ActionSpec
+
 
 @pytest.fixture
 def dummy_config():
 DISCRETE_ACTION_SPACE = [3, 3, 3, 2]
 BUFFER_INIT_SAMPLES = 64
 NUM_AGENTS = 12
+
+CONTINUOUS_ACTION_SPEC = ActionSpec.create_continuous(VECTOR_ACTION_SPACE)
+DISCRETE_ACTION_SPEC = ActionSpec.create_discrete(tuple(DISCRETE_ACTION_SPACE))
 
 
 def _create_ppo_optimizer_ops_mock(dummy_config, use_rnn, use_discrete, use_visual):
         length=time_horizon,
         observation_shapes=optimizer.policy.behavior_spec.observation_shapes,
         max_step_complete=True,
-        action_space=DISCRETE_ACTION_SPACE if discrete else VECTOR_ACTION_SPACE,
-        is_discrete=discrete,
+        action_spec=DISCRETE_ACTION_SPEC if discrete else CONTINUOUS_ACTION_SPEC,
     )
     run_out, final_value_out = optimizer.get_trajectory_value_estimates(
         trajectory.to_agentbuffer(), trajectory.next_obs, done=False
         length=time_horizon,
         observation_shapes=behavior_spec.observation_shapes,
         max_step_complete=True,
-        action_space=[2],
+        action_spec=behavior_spec.action_spec,
     )
     trajectory_queue.put(trajectory)
     trainer.advance()
         length=time_horizon + 1,
         max_step_complete=False,
         observation_shapes=behavior_spec.observation_shapes,
-        action_space=[2],
+        action_spec=behavior_spec.action_spec,
     )
     trajectory_queue.put(trajectory)
     trainer.advance()

ml-agents/mlagents/trainers/tests/tensorflow/test_sac.py

         length=15,
         observation_shapes=specs.observation_shapes,
         max_step_complete=True,
-        action_space=2,
-        is_discrete=False,
+        action_spec=specs.action_spec,
     )
     trajectory_queue.put(trajectory)
     trainer.advance()
         length=6,
         observation_shapes=specs.observation_shapes,
         max_step_complete=False,
-        action_space=2,
-        is_discrete=False,
+        action_spec=specs.action_spec,
     )
     trajectory_queue.put(trajectory)
     trainer.advance()
     trajectory = make_fake_trajectory(
         length=5,
         observation_shapes=specs.observation_shapes,
+        action_spec=specs.action_spec,
-        action_space=2,
-        is_discrete=False,
     )
     trajectory_queue.put(trajectory)
     trainer.advance()

ml-agents/mlagents/trainers/tests/tensorflow/test_saver.py

         length=time_horizon,
         max_step_complete=True,
         observation_shapes=[(1,)],
-        action_space=[2],
+        action_spec=behavior_spec.action_spec,
     )
     # Change half of the obs to 0
     for i in range(3):
         length=time_horizon,
         max_step_complete=True,
         observation_shapes=[(1,)],
-        action_space=[2],
+        action_spec=behavior_spec.action_spec,
     )
     trajectory_buffer = trajectory.to_agentbuffer()
     policy1.update_normalization(trajectory_buffer["vector_obs"])

ml-agents/mlagents/trainers/tests/tensorflow/test_simple_rl.py

         PPO_TF_CONFIG,
         hyperparameters=new_hyperparams,
         network_settings=new_networksettings,
-        max_steps=500,
+        max_steps=300,
         summary_freq=100,
         framework=FrameworkType.TENSORFLOW,
     )
 
 @pytest.mark.parametrize("use_discrete", [True, False])
 def test_recurrent_sac(use_discrete):
-    step_size = 0.5 if use_discrete else 0.2
+    step_size = 0.2 if use_discrete else 0.5
-        memory=NetworkSettings.MemorySettings(memory_size=16, sequence_length=16),
+        memory=NetworkSettings.MemorySettings(memory_size=16),
     )
     new_hyperparams = attr.evolve(
         SAC_TF_CONFIG.hyperparameters,
         SAC_TF_CONFIG,
         hyperparameters=new_hyperparams,
         network_settings=new_networksettings,
-        max_steps=5000,
+        max_steps=4000,
         framework=FrameworkType.TENSORFLOW,
     )
     _check_environment_trains(env, {BRAIN_NAME: config})

ml-agents/mlagents/trainers/tests/tensorflow/test_tf_policy.py

 from unittest.mock import MagicMock
 from mlagents.trainers.settings import TrainerSettings
 import numpy as np
+from mlagents_envs.base_env import ActionSpec
-def basic_mock_brain():
-    mock_brain = MagicMock()
-    mock_brain.vector_action_space_type = "continuous"
-    mock_brain.vector_observation_space_size = 1
-    mock_brain.vector_action_space_size = [1]
-    mock_brain.brain_name = "MockBrain"
-    return mock_brain
+def basic_behavior_spec():
+    dummy_actionspec = ActionSpec.create_continuous(1)
+    dummy_groupspec = BehaviorSpec([(1,)], dummy_actionspec)
+    return dummy_groupspec
 
 
 class FakePolicy(TFPolicy):
 
 def test_take_action_returns_empty_with_no_agents():
     test_seed = 3
-    policy = FakePolicy(test_seed, basic_mock_brain(), TrainerSettings(), "output")
-    # Doesn't really matter what this is
-    dummy_groupspec = BehaviorSpec([(1,)], "continuous", 1)
-    no_agent_step = DecisionSteps.empty(dummy_groupspec)
+    behavior_spec = basic_behavior_spec()
+    policy = FakePolicy(test_seed, behavior_spec, TrainerSettings(), "output")
+    no_agent_step = DecisionSteps.empty(behavior_spec)
     result = policy.get_action(no_agent_step)
     assert result == ActionInfo.empty()
 
-    policy = FakePolicy(test_seed, basic_mock_brain(), TrainerSettings(), "output")
+    behavior_spec = basic_behavior_spec()
+    policy = FakePolicy(test_seed, behavior_spec, TrainerSettings(), "output")
     policy.evaluate = MagicMock(return_value={})
     policy.save_memories = MagicMock()
     step_with_agents = DecisionSteps(
 
 def test_take_action_returns_action_info_when_available():
     test_seed = 3
-    policy = FakePolicy(test_seed, basic_mock_brain(), TrainerSettings(), "output")
+    behavior_spec = basic_behavior_spec()
+    policy = FakePolicy(test_seed, behavior_spec, TrainerSettings(), "output")
     policy_eval_out = {
         "action": np.array([1.0], dtype=np.float32),
         "memory_out": np.array([[2.5]], dtype=np.float32),

ml-agents/mlagents/trainers/tests/test_agent_processor.py

 from mlagents.trainers.behavior_id_utils import get_global_agent_id
 from mlagents_envs.side_channel.stats_side_channel import StatsAggregationMethod
 
+from mlagents_envs.base_env import ActionSpec
+
 
 def create_mock_policy():
     mock_policy = mock.Mock()
     mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
         num_agents=2,
         observation_shapes=[(8,)] + num_vis_obs * [(84, 84, 3)],
-        action_shape=2,
+        action_spec=ActionSpec.create_continuous(2),
     )
     fake_action_info = ActionInfo(
         action=[0.1, 0.1],
     mock_decision_steps, mock_terminal_steps = mb.create_mock_steps(
         num_agents=0,
         observation_shapes=[(8,)] + num_vis_obs * [(84, 84, 3)],
-        action_shape=2,
+        action_spec=ActionSpec.create_continuous(2),
     )
     processor.add_experiences(
         mock_decision_steps, mock_terminal_steps, 0, ActionInfo([], [], {}, [])
         "log_probs": [0.1],
     }
     mock_decision_step, mock_terminal_step = mb.create_mock_steps(
-        num_agents=1, observation_shapes=[(8,)], action_shape=2
+        num_agents=1,
+        observation_shapes=[(8,)],
+        action_spec=ActionSpec.create_continuous(2),
-        num_agents=1, observation_shapes=[(8,)], action_shape=2, done=True
+        num_agents=1,
+        observation_shapes=[(8,)],
+        action_spec=ActionSpec.create_continuous(2),
+        done=True,
     )
     fake_action_info = ActionInfo(
         action=[0.1],
         "log_probs": [0.1],
     }
     mock_decision_step, mock_terminal_step = mb.create_mock_steps(
-        num_agents=1, observation_shapes=[(8,)], action_shape=2
+        num_agents=1,
+        observation_shapes=[(8,)],
+        action_spec=ActionSpec.create_continuous(2),
     )
     fake_action_info = ActionInfo(
         action=[0.1],

ml-agents/mlagents/trainers/tests/test_rl_trainer.py

 from mlagents.trainers.agent_processor import AgentManagerQueue
 from mlagents.trainers.settings import TrainerSettings, FrameworkType
 
+from mlagents_envs.base_env import ActionSpec
+
 
 # Add concrete implementations of abstract methods
 class FakeTrainer(RLTrainer):
         length=time_horizon,
         observation_shapes=[(1,)],
         max_step_complete=True,
-        action_space=[2],
+        action_spec=ActionSpec.create_discrete((2,)),
     )
     trajectory_queue.put(trajectory)
 
         length=time_horizon,
         observation_shapes=[(1,)],
         max_step_complete=True,
-        action_space=[2],
+        action_spec=ActionSpec.create_discrete((2,)),
     )
     # Check that we can turn off the trainer and that the buffer is cleared
     num_trajectories = 5

ml-agents/mlagents/trainers/tests/test_stats.py

     )
 
 
-@mock.patch("mlagents.tf_utils.tf.Summary")
-@mock.patch("mlagents.tf_utils.tf.summary.FileWriter")
-def test_tensorboard_writer(mock_filewriter, mock_summary):
+@mock.patch("mlagents.trainers.stats.SummaryWriter")
+def test_tensorboard_writer(mock_summary):
     # Test write_stats
     category = "category1"
     with tempfile.TemporaryDirectory(prefix="unittest-") as base_dir:
             basedir=base_dir, category=category
         )
         assert os.path.exists(filewriter_dir)
-        mock_filewriter.assert_called_once_with(filewriter_dir)
+        mock_summary.assert_called_once_with(filewriter_dir)
-        mock_summary.return_value.value.add.assert_called_once_with(
-            tag="key1", simple_value=1.0
-        )
-        mock_filewriter.return_value.add_summary.assert_called_once_with(
-            mock_summary.return_value, 10
-        )
-        mock_filewriter.return_value.flush.assert_called_once()
+        mock_summary.return_value.add_scalar.assert_called_once_with("key1", 1.0, 10)
+        mock_summary.return_value.flush.assert_called_once()
-        assert mock_filewriter.return_value.add_summary.call_count > 1
+        assert mock_summary.return_value.add_text.call_count >= 1
 
 
 def test_tensorboard_writer_clear(tmp_path):
                 },
                 10,
             )
-            # Test hyperparameter writing - no good way to parse the TB string though.
+            # Test hyperparameter writing
             console_writer.add_property(
                 "category1", StatsPropertyType.HYPERPARAMETERS, {"example": 1.0}
             )

ml-agents/mlagents/trainers/tests/test_trajectory.py

 from mlagents.trainers.trajectory import SplitObservations
 from mlagents.trainers.tests.mock_brain import make_fake_trajectory
 
+from mlagents_envs.base_env import ActionSpec
+
 VEC_OBS_SIZE = 6
 ACTION_SIZE = 4
 
     trajectory = make_fake_trajectory(
         length=length,
         observation_shapes=[(VEC_OBS_SIZE,), (84, 84, 3)],
-        action_space=[ACTION_SIZE],
+        action_spec=ActionSpec.create_continuous(ACTION_SIZE),
     )
     agentbuffer = trajectory.to_agentbuffer()
     seen_keys = set()

ml-agents/mlagents/trainers/tests/torch/test_ghost.py

         length=time_horizon,
         max_step_complete=True,
         observation_shapes=[(1,)],
-        action_space=[2],
+        action_spec=mock_specs.action_spec,
     )
     trajectory_queue0.put(trajectory)
     trainer.advance()

ml-agents/mlagents/trainers/tests/torch/test_networks.py

     SeparateActorCritic,
 )
 from mlagents.trainers.settings import NetworkSettings
-from mlagents_envs.base_env import ActionType
+
+from mlagents_envs.base_env import ActionSpec
 
 
 def test_networkbody_vector():
             assert _out[0] == pytest.approx(1.0, abs=0.1)
 
 
-@pytest.mark.parametrize("action_type", [ActionType.DISCRETE, ActionType.CONTINUOUS])
-def test_simple_actor(action_type):
+@pytest.mark.parametrize("use_discrete", [True, False])
+def test_simple_actor(use_discrete):
-    masks = None if action_type == ActionType.CONTINUOUS else torch.ones((1, 1))
-    actor = SimpleActor(obs_shapes, network_settings, action_type, act_size)
+    if use_discrete:
+        masks = torch.ones((1, 1))
+        action_spec = ActionSpec.create_discrete(tuple(act_size))
+    else:
+        masks = None
+        action_spec = ActionSpec.create_continuous(act_size[0])
+    actor = SimpleActor(obs_shapes, network_settings, action_spec)
-        if action_type == ActionType.CONTINUOUS:
-            assert isinstance(dist, GaussianDistInstance)
+        if use_discrete:
+            assert isinstance(dist, CategoricalDistInstance)
-            assert isinstance(dist, CategoricalDistInstance)
+            assert isinstance(dist, GaussianDistInstance)
-        if action_type == ActionType.CONTINUOUS:
-            assert act.shape == (1, act_size[0])
-        else:
+        if use_discrete:
+        else:
+            assert act.shape == (1, act_size[0])
 
     # Test forward
     actions, ver_num, mem_size, is_cont, act_size_vec = actor.forward(
         # This is different from above for ONNX export
-        if action_type == ActionType.CONTINUOUS:
+        if use_discrete:
+            assert act.shape == tuple(act_size)
+        else:
-        else:
-            assert act.shape == tuple(act_size)
-    assert is_cont == int(action_type == ActionType.CONTINUOUS)
+    assert is_cont == int(not use_discrete)
     assert act_size_vec == torch.tensor(act_size)
 
 
     obs_shapes = [(obs_size,)]
     act_size = [2]
     stream_names = [f"stream_name{n}" for n in range(4)]
-    actor = ac_type(
-        obs_shapes, network_settings, ActionType.CONTINUOUS, act_size, stream_names
-    )
+    action_spec = ActionSpec.create_continuous(act_size[0])
+    actor = ac_type(obs_shapes, network_settings, action_spec, stream_names)
     if lstm:
         sample_obs = torch.ones((1, network_settings.memory.sequence_length, obs_size))
         memories = torch.ones(

ml-agents/mlagents/trainers/tests/torch/test_policy.py

         memories=memories,
         seq_len=policy.sequence_length,
     )
-    assert log_probs.shape == (64, policy.behavior_spec.action_size)
-    assert entropy.shape == (64, policy.behavior_spec.action_size)
+    if discrete:
+        _size = policy.behavior_spec.action_spec.discrete_size
+    else:
+        _size = policy.behavior_spec.action_spec.continuous_size
+
+    assert log_probs.shape == (64, _size)
+    assert entropy.shape == (64,)
     for val in values.values():
         assert val.shape == (64,)
 
     if discrete:
         assert log_probs.shape == (
             64,
-            sum(policy.behavior_spec.discrete_action_branches),
+            sum(policy.behavior_spec.action_spec.discrete_branches),
-        assert log_probs.shape == (64, policy.behavior_spec.action_shape)
-    assert entropies.shape == (64, policy.behavior_spec.action_size)
+        assert log_probs.shape == (64, policy.behavior_spec.action_spec.continuous_size)
+    assert entropies.shape == (64,)
 
     if rnn:
         assert memories.shape == (1, 1, policy.m_size)

ml-agents/mlagents/trainers/tests/torch/test_ppo.py

     gail_dummy_config,
 )
 
+from mlagents_envs.base_env import ActionSpec
+
 
 @pytest.fixture
 def dummy_config():
 DISCRETE_ACTION_SPACE = [3, 3, 3, 2]
 BUFFER_INIT_SAMPLES = 64
 NUM_AGENTS = 12
+
+CONTINUOUS_ACTION_SPEC = ActionSpec.create_continuous(VECTOR_ACTION_SPACE)
+DISCRETE_ACTION_SPEC = ActionSpec.create_discrete(tuple(DISCRETE_ACTION_SPACE))
 
 
 def create_test_ppo_optimizer(dummy_config, use_rnn, use_discrete, use_visual):
     trajectory = make_fake_trajectory(
         length=time_horizon,
         observation_shapes=optimizer.policy.behavior_spec.observation_shapes,
+        action_spec=DISCRETE_ACTION_SPEC if discrete else CONTINUOUS_ACTION_SPEC,
-        action_space=DISCRETE_ACTION_SPACE if discrete else VECTOR_ACTION_SPACE,
-        is_discrete=discrete,
     )
     run_out, final_value_out = optimizer.get_trajectory_value_estimates(
         trajectory.to_agentbuffer(), trajectory.next_obs, done=False

ml-agents/mlagents/trainers/tests/torch/test_reward_providers/test_curiosity.py

     CuriosityRewardProvider,
     create_reward_provider,
 )
-from mlagents_envs.base_env import BehaviorSpec, ActionType
+from mlagents_envs.base_env import BehaviorSpec, ActionSpec
 from mlagents.trainers.settings import CuriositySettings, RewardSignalType
 from mlagents.trainers.tests.torch.test_reward_providers.utils import (
     create_agent_buffer,
 SEED = [42]
 
+ACTIONSPEC_CONTINUOUS = ActionSpec.create_continuous(5)
+ACTIONSPEC_TWODISCRETE = ActionSpec.create_discrete((2, 3))
+ACTIONSPEC_DISCRETE = ActionSpec.create_discrete((2,))
+
-        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+        BehaviorSpec([(10,)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,)], ACTIONSPEC_TWODISCRETE),
     ],
 )
 def test_construction(behavior_spec: BehaviorSpec) -> None:
 @pytest.mark.parametrize(
     "behavior_spec",
     [
-        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,), (64, 66, 3), (84, 86, 1)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,), (64, 66, 1)], ActionType.DISCRETE, (2, 3)),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2,)),
+        BehaviorSpec([(10,)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,), (64, 66, 3), (84, 86, 1)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,), (64, 66, 1)], ACTIONSPEC_TWODISCRETE),
+        BehaviorSpec([(10,)], ACTIONSPEC_DISCRETE),
     ],
 )
 def test_factory(behavior_spec: BehaviorSpec) -> None:
 @pytest.mark.parametrize(
     "behavior_spec",
     [
-        BehaviorSpec([(10,), (64, 66, 3), (24, 26, 1)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2,)),
+        BehaviorSpec([(10,), (64, 66, 3), (24, 26, 1)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,)], ACTIONSPEC_TWODISCRETE),
+        BehaviorSpec([(10,)], ACTIONSPEC_DISCRETE),
     ],
 )
 def test_reward_decreases(behavior_spec: BehaviorSpec, seed: int) -> None:
 
 @pytest.mark.parametrize("seed", SEED)
 @pytest.mark.parametrize(
-    "behavior_spec", [BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5)]
+    "behavior_spec", [BehaviorSpec([(10,)], ACTIONSPEC_CONTINUOUS)]
 )
 def test_continuous_action_prediction(behavior_spec: BehaviorSpec, seed: int) -> None:
     np.random.seed(seed)
 @pytest.mark.parametrize(
     "behavior_spec",
     [
-        BehaviorSpec([(10,), (64, 66, 3)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2,)),
+        BehaviorSpec([(10,), (64, 66, 3), (24, 26, 1)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,)], ACTIONSPEC_TWODISCRETE),
+        BehaviorSpec([(10,)], ACTIONSPEC_DISCRETE),
     ],
 )
 def test_next_state_prediction(behavior_spec: BehaviorSpec, seed: int) -> None:

ml-agents/mlagents/trainers/tests/torch/test_reward_providers/test_extrinsic.py

     ExtrinsicRewardProvider,
     create_reward_provider,
 )
-from mlagents_envs.base_env import BehaviorSpec, ActionType
+from mlagents_envs.base_env import BehaviorSpec, ActionSpec
 from mlagents.trainers.settings import RewardSignalSettings, RewardSignalType
 from mlagents.trainers.tests.torch.test_reward_providers.utils import (
     create_agent_buffer,
+ACTIONSPEC_CONTINUOUS = ActionSpec.create_continuous(5)
+ACTIONSPEC_TWODISCRETE = ActionSpec.create_discrete((2, 3))
+
+
-        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+        BehaviorSpec([(10,)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,)], ACTIONSPEC_TWODISCRETE),
     ],
 )
 def test_construction(behavior_spec: BehaviorSpec) -> None:
 @pytest.mark.parametrize(
     "behavior_spec",
     [
-        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+        BehaviorSpec([(10,)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,)], ACTIONSPEC_TWODISCRETE),
     ],
 )
 def test_factory(behavior_spec: BehaviorSpec) -> None:
 @pytest.mark.parametrize(
     "behavior_spec",
     [
-        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+        BehaviorSpec([(10,)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,)], ACTIONSPEC_TWODISCRETE),
     ],
 )
 def test_reward(behavior_spec: BehaviorSpec, reward: float) -> None:

ml-agents/mlagents/trainers/tests/torch/test_reward_providers/test_gail.py

     GAILRewardProvider,
     create_reward_provider,
 )
-from mlagents_envs.base_env import BehaviorSpec, ActionType
+from mlagents_envs.base_env import BehaviorSpec, ActionSpec
 from mlagents.trainers.settings import GAILSettings, RewardSignalType
 from mlagents.trainers.tests.torch.test_reward_providers.utils import (
     create_agent_buffer,
 )
+
 
 CONTINUOUS_PATH = (
     os.path.join(os.path.dirname(os.path.abspath(__file__)), os.pardir, os.pardir)
 )
 SEED = [42]
 
+ACTIONSPEC_CONTINUOUS = ActionSpec.create_continuous(2)
+ACTIONSPEC_FOURDISCRETE = ActionSpec.create_discrete((2, 3, 3, 3))
+ACTIONSPEC_DISCRETE = ActionSpec.create_discrete((20,))
-@pytest.mark.parametrize(
-    "behavior_spec", [BehaviorSpec([(8,)], ActionType.CONTINUOUS, 2)]
-)
+
+@pytest.mark.parametrize("behavior_spec", [BehaviorSpec([(8,)], ACTIONSPEC_CONTINUOUS)])
 def test_construction(behavior_spec: BehaviorSpec) -> None:
     gail_settings = GAILSettings(demo_path=CONTINUOUS_PATH)
     gail_rp = GAILRewardProvider(behavior_spec, gail_settings)
-@pytest.mark.parametrize(
-    "behavior_spec", [BehaviorSpec([(8,)], ActionType.CONTINUOUS, 2)]
-)
+@pytest.mark.parametrize("behavior_spec", [BehaviorSpec([(8,)], ACTIONSPEC_CONTINUOUS)])
 def test_factory(behavior_spec: BehaviorSpec) -> None:
     gail_settings = GAILSettings(demo_path=CONTINUOUS_PATH)
     gail_rp = create_reward_provider(
 @pytest.mark.parametrize(
     "behavior_spec",
     [
-        BehaviorSpec([(8,), (24, 26, 1)], ActionType.CONTINUOUS, 2),
-        BehaviorSpec([(50,)], ActionType.DISCRETE, (2, 3, 3, 3)),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (20,)),
+        BehaviorSpec([(8,), (24, 26, 1)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(50,)], ACTIONSPEC_FOURDISCRETE),
+        BehaviorSpec([(10,)], ACTIONSPEC_DISCRETE),
     ],
 )
 @pytest.mark.parametrize("use_actions", [False, True])
 @pytest.mark.parametrize(
     "behavior_spec",
     [
-        BehaviorSpec([(8,)], ActionType.CONTINUOUS, 2),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3, 3, 3)),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (20,)),
+        BehaviorSpec([(8,), (24, 26, 1)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(50,)], ACTIONSPEC_FOURDISCRETE),
+        BehaviorSpec([(10,)], ACTIONSPEC_DISCRETE),
     ],
 )
 @pytest.mark.parametrize("use_actions", [False, True])
         RewardSignalType.GAIL, behavior_spec, gail_settings
     )
 
-    for _ in range(200):
+    for _ in range(300):
         gail_rp.update(buffer_policy)
         reward_expert = gail_rp.evaluate(buffer_expert)[0]
         reward_policy = gail_rp.evaluate(buffer_policy)[0]

ml-agents/mlagents/trainers/tests/torch/test_reward_providers/test_rnd.py

     RNDRewardProvider,
     create_reward_provider,
 )
-from mlagents_envs.base_env import BehaviorSpec, ActionType
+from mlagents_envs.base_env import BehaviorSpec, ActionSpec
+
+
+ACTIONSPEC_CONTINUOUS = ActionSpec.create_continuous(5)
+ACTIONSPEC_TWODISCRETE = ActionSpec.create_discrete((2, 3))
+ACTIONSPEC_DISCRETE = ActionSpec.create_discrete((2,))
-        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
+        BehaviorSpec([(10,)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,)], ACTIONSPEC_TWODISCRETE),
     ],
 )
 def test_construction(behavior_spec: BehaviorSpec) -> None:
 @pytest.mark.parametrize(
     "behavior_spec",
     [
-        BehaviorSpec([(10,)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,), (64, 66, 3), (84, 86, 1)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,), (64, 66, 1)], ActionType.DISCRETE, (2, 3)),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2,)),
+        BehaviorSpec([(10,)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,), (64, 66, 3), (84, 86, 1)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,), (64, 66, 1)], ACTIONSPEC_TWODISCRETE),
+        BehaviorSpec([(10,)], ACTIONSPEC_DISCRETE),
     ],
 )
 def test_factory(behavior_spec: BehaviorSpec) -> None:
 @pytest.mark.parametrize(
     "behavior_spec",
     [
-        BehaviorSpec([(10,), (64, 66, 3), (24, 26, 1)], ActionType.CONTINUOUS, 5),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2, 3)),
-        BehaviorSpec([(10,)], ActionType.DISCRETE, (2,)),
+        BehaviorSpec([(10,), (64, 66, 3), (24, 26, 1)], ACTIONSPEC_CONTINUOUS),
+        BehaviorSpec([(10,)], ACTIONSPEC_TWODISCRETE),
+        BehaviorSpec([(10,)], ACTIONSPEC_DISCRETE),
     ],
 )
 def test_reward_decreases(behavior_spec: BehaviorSpec, seed: int) -> None:

ml-agents/mlagents/trainers/tests/torch/test_reward_providers/utils.py

 ) -> AgentBuffer:
     buffer = AgentBuffer()
     curr_observations = [
-        np.random.normal(size=shape) for shape in behavior_spec.observation_shapes
+        np.random.normal(size=shape).astype(np.float32)
+        for shape in behavior_spec.observation_shapes
-        np.random.normal(size=shape) for shape in behavior_spec.observation_shapes
+        np.random.normal(size=shape).astype(np.float32)
+        for shape in behavior_spec.observation_shapes
-    action = behavior_spec.create_random_action(1)[0, :]
+    action = behavior_spec.action_spec.random_action(1)[0, :]
     for _ in range(number):
         curr_split_obs = SplitObservations.from_observations(curr_observations)
         next_split_obs = SplitObservations.from_observations(next_observations)

ml-agents/mlagents/trainers/tests/torch/test_simple_rl.py

         PPO_TORCH_CONFIG,
         hyperparameters=new_hyperparams,
         network_settings=new_networksettings,
-        max_steps=700,
+        max_steps=900,
         summary_freq=100,
     )
     # The number of steps is pretty small for these encoders
     )
     new_hyperparams = attr.evolve(
         SAC_TORCH_CONFIG.hyperparameters,
-        batch_size=128,
+        batch_size=256,
         learning_rate=1e-3,
         buffer_init_steps=1000,
         steps_per_update=2,
         hyperparameters=new_hyperparams,
         network_settings=new_networksettings,
-        max_steps=5000,
+        max_steps=2000,
     )
     check_environment_trains(env, {BRAIN_NAME: config})
 

ml-agents/mlagents/trainers/tests/torch/test_utils.py

 
 def test_list_to_tensor():
     # Test converting pure list
-    unconverted_list = [[1, 2], [1, 3], [1, 4]]
+    unconverted_list = [[1.0, 2], [1, 3], [1, 4]]
     tensor = ModelUtils.list_to_tensor(unconverted_list)
     # Should be equivalent to torch.tensor conversion
     assert torch.equal(tensor, torch.tensor(unconverted_list))
     list_of_np = [np.asarray(_el) for _el in unconverted_list]
     tensor = ModelUtils.list_to_tensor(list_of_np)
     # Should be equivalent to torch.tensor conversion
-    assert torch.equal(tensor, torch.tensor(unconverted_list))
+    assert torch.equal(tensor, torch.tensor(unconverted_list, dtype=torch.float32))
 
 
 def test_break_into_branches():

ml-agents/mlagents/trainers/tf/components/bc/model.py

         self.done_expert = tf.placeholder(shape=[None, 1], dtype=tf.float32)
         self.done_policy = tf.placeholder(shape=[None, 1], dtype=tf.float32)
 
-        if self.policy.behavior_spec.is_action_continuous():
+        if self.policy.behavior_spec.action_spec.is_continuous():
             action_length = self.policy.act_size[0]
             self.action_in_expert = tf.placeholder(
                 shape=[None, action_length], dtype=tf.float32

ml-agents/mlagents/trainers/tf/components/bc/module.py

             self.policy.sequence_length_ph: self.policy.sequence_length,
         }
         feed_dict[self.model.action_in_expert] = mini_batch_demo["actions"]
-        if self.policy.behavior_spec.is_action_discrete():
+        if self.policy.behavior_spec.action_spec.is_discrete():
-                    sum(self.policy.behavior_spec.discrete_action_branches),
+                    sum(self.policy.behavior_spec.action_spec.discrete_branches),
                 ),
                 dtype=np.float32,
             )

ml-agents/mlagents/trainers/tf/components/reward_signals/curiosity/model.py

         """
         combined_input = tf.concat([encoded_state, encoded_next_state], axis=1)
         hidden = tf.layers.dense(combined_input, 256, activation=ModelUtils.swish)
-        if self.policy.behavior_spec.is_action_continuous():
+        if self.policy.behavior_spec.action_spec.is_continuous():
             pred_action = tf.layers.dense(
                 hidden, self.policy.act_size[0], activation=None
             )

ml-agents/mlagents/trainers/tf/components/reward_signals/gail/model.py

         self.done_expert = tf.expand_dims(self.done_expert_holder, -1)
         self.done_policy = tf.expand_dims(self.done_policy_holder, -1)
 
-        if self.policy.behavior_spec.is_action_continuous():
+        if self.policy.behavior_spec.action_spec.is_continuous():
             action_length = self.policy.act_size[0]
             self.action_in_expert = tf.placeholder(
                 shape=[None, action_length], dtype=tf.float32

ml-agents/mlagents/trainers/tf/model_serialization.py

     ]
 )
 
-POSSIBLE_OUTPUT_NODES = frozenset(
-    ["action", "action_probs", "recurrent_out", "value_estimate"]
-)
+POSSIBLE_OUTPUT_NODES = frozenset(["action", "recurrent_out", "value_estimate"])
 
 MODEL_CONSTANTS = frozenset(
     [

ml-agents/mlagents/trainers/torch/components/bc/module.py

                 np.ones(
                     (
                         self.n_sequences * self.policy.sequence_length,
-                        sum(self.policy.behavior_spec.discrete_action_branches),
+                        sum(self.policy.behavior_spec.action_spec.discrete_branches),
                     ),
                     dtype=np.float32,
                 )

ml-agents/mlagents/trainers/torch/components/reward_providers/curiosity_reward_provider.py

 
     def __init__(self, specs: BehaviorSpec, settings: CuriositySettings) -> None:
         super().__init__()
-        self._policy_specs = specs
+        self._action_spec = specs.action_spec
         state_encoder_settings = NetworkSettings(
             normalize=False,
             hidden_units=settings.encoding_size,
             specs.observation_shapes, state_encoder_settings
         )
 
-        self._action_flattener = ModelUtils.ActionFlattener(specs)
+        self._action_flattener = ModelUtils.ActionFlattener(self._action_spec)
 
         self.inverse_model_action_prediction = torch.nn.Sequential(
             LinearEncoder(2 * settings.encoding_size, 1, 256),
             (self.get_current_state(mini_batch), self.get_next_state(mini_batch)), dim=1
         )
         hidden = self.inverse_model_action_prediction(inverse_model_input)
-        if self._policy_specs.is_action_continuous():
+        if self._action_spec.is_continuous():
-                hidden, self._policy_specs.discrete_action_branches
+                hidden, self._action_spec.discrete_branches
             )
             branches = [torch.softmax(b, dim=1) for b in branches]
             return torch.cat(branches, dim=1)
         Uses the current state embedding and the action of the mini_batch to predict
         the next state embedding.
         """
-        if self._policy_specs.is_action_continuous():
+        if self._action_spec.is_continuous():
-                    self._policy_specs.discrete_action_branches,
+                    self._action_spec.discrete_branches,
                 ),
                 dim=1,
             )
         action prediction (given the current and next state).
         """
         predicted_action = self.predict_action(mini_batch)
-        if self._policy_specs.is_action_continuous():
+        if self._action_spec.is_continuous():
             sq_difference = (
                 ModelUtils.list_to_tensor(mini_batch["actions"], dtype=torch.float)
                 - predicted_action
             true_action = torch.cat(
                 ModelUtils.actions_to_onehot(
                     ModelUtils.list_to_tensor(mini_batch["actions"], dtype=torch.long),
-                    self._policy_specs.discrete_action_branches,
+                    self._action_spec.discrete_branches,
                 ),
                 dim=1,
             )

ml-agents/mlagents/trainers/torch/components/reward_providers/gail_reward_provider.py

 
     def __init__(self, specs: BehaviorSpec, settings: GAILSettings) -> None:
         super().__init__()
-        self._policy_specs = specs
         self._use_vail = settings.use_vail
         self._settings = settings
 
             vis_encode_type=EncoderType.SIMPLE,
             memory=None,
         )
-        self._action_flattener = ModelUtils.ActionFlattener(specs)
+        self._action_flattener = ModelUtils.ActionFlattener(specs.action_spec)
         unencoded_size = (
             self._action_flattener.flattened_size + 1 if settings.use_actions else 0
         )  # +1 is for dones

ml-agents/mlagents/trainers/torch/components/reward_providers/rnd_reward_provider.py

 
     def __init__(self, specs: BehaviorSpec, settings: RNDSettings) -> None:
         super().__init__()
-        self._policy_specs = specs
         state_encoder_settings = NetworkSettings(
             normalize=True,
             hidden_units=settings.encoding_size,

ml-agents/mlagents/trainers/torch/distributions.py

         if self.conditional_sigma:
             log_sigma = torch.clamp(self.log_sigma(inputs), min=-20, max=2)
         else:
-            log_sigma = self.log_sigma
+            # Expand so that entropy matches batch size. Note that we're using
+            # torch.cat here instead of torch.expand() becuase it is not supported in the
+            # verified version of Barracuda (1.0.2).
+            log_sigma = torch.cat([self.log_sigma] * inputs.shape[0], axis=0)
         if self.tanh_squash:
             return [TanhGaussianDistInstance(mu, torch.exp(log_sigma))]
         else:

ml-agents/mlagents/trainers/torch/model_serialization.py

             for shape in self.policy.behavior_spec.observation_shapes
             if len(shape) == 3
         ]
-        dummy_masks = torch.ones(batch_dim + [sum(self.policy.actor_critic.act_size)])
+        dummy_masks = torch.ones(
+            batch_dim + [sum(self.policy.behavior_spec.action_spec.discrete_branches)]
+        )
         dummy_memories = torch.zeros(
             batch_dim + seq_len_dim + [self.policy.export_memory_size]
         )

ml-agents/mlagents/trainers/torch/networks.py

 
 from mlagents.torch_utils import torch, nn
 
-from mlagents_envs.base_env import ActionType
+from mlagents_envs.base_env import ActionSpec
 from mlagents.trainers.torch.distributions import (
     GaussianDistribution,
     MultiCategoricalDistribution,
         self,
         observation_shapes: List[Tuple[int, ...]],
         network_settings: NetworkSettings,
-        act_type: ActionType,
-        act_size: List[int],
+        action_spec: ActionSpec,
-        self.act_type = act_type
-        self.act_size = act_size
+        self.action_spec = action_spec
-            torch.Tensor([int(act_type == ActionType.CONTINUOUS)])
+            torch.Tensor([int(self.action_spec.is_continuous())])
-            torch.Tensor([sum(act_size)]), requires_grad=False
+            torch.Tensor(
+                [
+                    self.action_spec.continuous_size
+                    + sum(self.action_spec.discrete_branches)
+                ]
+            ),
+            requires_grad=False,
         )
         self.network_body = NetworkBody(observation_shapes, network_settings)
         if network_settings.memory is not None:
 
-        if self.act_type == ActionType.CONTINUOUS:
+        if self.action_spec.is_continuous():
-                act_size[0],
+                self.action_spec.continuous_size,
-                self.encoding_size, act_size
+                self.encoding_size, self.action_spec.discrete_branches
             )
 
     @property
         encoding, memories = self.network_body(
             vec_inputs, vis_inputs, memories=memories, sequence_length=sequence_length
         )
-        if self.act_type == ActionType.CONTINUOUS:
+        if self.action_spec.is_continuous():
             dists = self.distribution(encoding)
         else:
             dists = self.distribution(encoding, masks)
         Note: This forward() method is required for exporting to ONNX. Don't modify the inputs and outputs.
         """
         dists, _ = self.get_dists(vec_inputs, vis_inputs, masks, memories, 1)
-        if self.act_type == ActionType.CONTINUOUS:
+        if self.action_spec.is_continuous():
             action_list = self.sample_action(dists)
             action_out = torch.stack(action_list, dim=-1)
         else:
         self,
         observation_shapes: List[Tuple[int, ...]],
         network_settings: NetworkSettings,
-        act_type: ActionType,
-        act_size: List[int],
+        action_spec: ActionSpec,
         stream_names: List[str],
         conditional_sigma: bool = False,
         tanh_squash: bool = False,
             network_settings,
-            act_type,
-            act_size,
+            action_spec,
             conditional_sigma,
             tanh_squash,
         )
         encoding, memories = self.network_body(
             vec_inputs, vis_inputs, memories=memories, sequence_length=sequence_length
         )
-        if self.act_type == ActionType.CONTINUOUS:
+        if self.action_spec.is_continuous():
             dists = self.distribution(encoding)
         else:
             dists = self.distribution(encoding, masks=masks)
         self,
         observation_shapes: List[Tuple[int, ...]],
         network_settings: NetworkSettings,
-        act_type: ActionType,
-        act_size: List[int],
+        action_spec: ActionSpec,
         stream_names: List[str],
         conditional_sigma: bool = False,
         tanh_squash: bool = False,
         super().__init__(
             observation_shapes,
             network_settings,
-            act_type,
-            act_size,
+            action_spec,
             conditional_sigma,
             tanh_squash,
         )

ml-agents/mlagents/trainers/torch/utils.py

 )
 from mlagents.trainers.settings import EncoderType, ScheduleType
 from mlagents.trainers.exception import UnityTrainerException
-from mlagents_envs.base_env import BehaviorSpec
+from mlagents_envs.base_env import ActionSpec
 from mlagents.trainers.torch.distributions import DistInstance, DiscreteDistInstance
 
 
     }
 
     class ActionFlattener:
-        def __init__(self, behavior_spec: BehaviorSpec):
-            self._specs = behavior_spec
+        def __init__(self, action_spec: ActionSpec):
+            self._specs = action_spec
-            if self._specs.is_action_continuous():
-                return self._specs.action_size
+            if self._specs.is_continuous():
+                return self._specs.continuous_size
-                return sum(self._specs.discrete_action_branches)
+                return sum(self._specs.discrete_branches)
-            if self._specs.is_action_continuous():
+            if self._specs.is_continuous():
-                        self._specs.discrete_action_branches,
+                        self._specs.discrete_branches,
                     ),
                     dim=1,
                 )
 
     @staticmethod
     def list_to_tensor(
-        ndarray_list: List[np.ndarray], dtype: Optional[torch.dtype] = None
+        ndarray_list: List[np.ndarray], dtype: Optional[torch.dtype] = torch.float32
     ) -> torch.Tensor:
         """
         Converts a list of numpy arrays into a tensor. MUCH faster than

ml-agents/mlagents/trainers/trainer/rl_trainer.py

 from mlagents.trainers.optimizer import Optimizer
 from mlagents.trainers.buffer import AgentBuffer
 from mlagents.trainers.trainer import Trainer
-from mlagents.trainers.tf.components.reward_signals import (
-    RewardSignalResult,
-    RewardSignal,
+from mlagents.trainers.torch.components.reward_providers.base_reward_provider import (
+    BaseRewardProvider,
-from mlagents.trainers.policy.tf_policy import TFPolicy
+from mlagents.trainers.policy.torch_policy import TorchPolicy
+from mlagents.trainers.model_saver.torch_model_saver import TorchModelSaver
 from mlagents.trainers.behavior_id_utils import BehaviorIdentifiers
 from mlagents.trainers.agent_processor import AgentManagerQueue
 from mlagents.trainers.trajectory import Trajectory
-from mlagents.trainers.model_saver.tf_model_saver import TFModelSaver
+
-from mlagents import torch_utils
+from mlagents import tf_utils
-if torch_utils.is_available():
-    from mlagents.trainers.policy.torch_policy import TorchPolicy
-    from mlagents.trainers.model_saver.torch_model_saver import TorchModelSaver
+if tf_utils.is_available():
+    from mlagents.trainers.policy.tf_policy import TFPolicy
+    from mlagents.trainers.model_saver.tf_model_saver import TFModelSaver
-    TorchPolicy = None  # type: ignore
-    TorchSaver = None  # type: ignore
-
-RewardSignalResults = Dict[str, RewardSignalResult]
+    TFPolicy = None  # type: ignore
+    TFModelSaver = None  # type: ignore
 
 logger = get_logger(__name__)
 
             StatsPropertyType.HYPERPARAMETERS, self.trainer_settings.as_dict()
         )
         self.framework = self.trainer_settings.framework
-        if self.framework == FrameworkType.PYTORCH and not torch_utils.is_available():
+        if self.framework == FrameworkType.TENSORFLOW and not tf_utils.is_available():
-                "To use the experimental PyTorch backend, install the PyTorch Python package first."
+                "To use the TensorFlow backend, install the TensorFlow Python package first."
             )
 
         logger.debug(f"Using framework {self.framework.value}")
                 self.reward_buffer.appendleft(rewards.get(agent_id, 0))
                 rewards[agent_id] = 0
             else:
-                if isinstance(optimizer.reward_signals[name], RewardSignal):
+                if isinstance(optimizer.reward_signals[name], BaseRewardProvider):
-                        optimizer.reward_signals[name].stat_name,
+                        f"Policy/{optimizer.reward_signals[name].name.capitalize()} Reward",
-                        f"Policy/{optimizer.reward_signals[name].name.capitalize()} Reward",
+                        optimizer.reward_signals[name].stat_name,
                         rewards.get(agent_id, 0),
                     )
                 rewards[agent_id] = 0

ml-agents/mlagents/trainers/trainer/trainer_factory.py

         init_path: str = None,
         multi_gpu: bool = False,
         force_torch: bool = False,
+        force_tensorflow: bool = False,
     ):
         """
         The TrainerFactory generates the Trainers based on the configuration passed as
         :param init_path: Path from which to load model.
         :param multi_gpu: If True, multi-gpu will be used. (currently not available)
         :param force_torch: If True, the Trainers will all use the PyTorch framework
-        instead of the TensorFlow framework.
+        instead of what is specified in the config YAML.
+        :param force_tensorflow: If True, thee Trainers will all use the TensorFlow
+        framework.
         """
         self.trainer_config = trainer_config
         self.output_path = output_path
         self.multi_gpu = multi_gpu
         self.ghost_controller = GhostController()
         self._force_torch = force_torch
+        self._force_tf = force_tensorflow
 
     def generate(self, behavior_name: str) -> Trainer:
         if behavior_name not in self.trainer_config.keys():
         trainer_settings = self.trainer_config[behavior_name]
         if self._force_torch:
             trainer_settings.framework = FrameworkType.PYTORCH
+            logger.warning(
+                "Note that specifying --torch is not required anymore as PyTorch is the default framework."
+            )
+        if self._force_tf:
+            trainer_settings.framework = FrameworkType.TENSORFLOW
+            logger.warning(
+                "Setting the framework to TensorFlow. TensorFlow trainers will be deprecated in the future."
+            )
+            if self._force_torch:
+                logger.warning(
+                    "Both --torch and --tensorflow CLI options were specified. Using TensorFlow."
+                )
         return TrainerFactory._initialize_trainer(
             trainer_settings,
             behavior_name,

ml-agents/mlagents/trainers/trainer_controller.py

 
 import numpy as np
 from mlagents.tf_utils import tf
+from mlagents import tf_utils
 
 from mlagents_envs.logging_util import get_logger
 from mlagents.trainers.env_manager import EnvManager, EnvironmentStep
         self.trainer_threads: List[threading.Thread] = []
         self.kill_trainers = False
         np.random.seed(training_seed)
-        tf.set_random_seed(training_seed)
-        if torch_utils.is_available():
-            torch_utils.torch.manual_seed(training_seed)
+        if tf_utils.is_available():
+            tf.set_random_seed(training_seed)
+        torch_utils.torch.manual_seed(training_seed)
         self.rank = get_rank()
 
     @timed
     @timed
     def start_learning(self, env_manager: EnvManager) -> None:
         self._create_output_path(self.output_path)
-        tf.reset_default_graph()
+        if tf_utils.is_available():
+            tf.reset_default_graph()
         try:
             # Initial reset
             self._reset_env(env_manager)

ml-agents/mlagents/trainers/training_status.py

 import attr
 import cattr
 
-from mlagents.tf_utils import tf
+from mlagents.torch_utils import torch
+from mlagents.tf_utils import tf, is_available as tf_is_available
 from mlagents_envs.logging_util import get_logger
 from mlagents.trainers import __version__
 from mlagents.trainers.exception import TrainerError
-STATUS_FORMAT_VERSION = "0.1.0"
+STATUS_FORMAT_VERSION = "0.2.0"
 
 
 class StatusType(Enum):
 class StatusMetaData:
     stats_format_version: str = STATUS_FORMAT_VERSION
     mlagents_version: str = __version__
-    tensorflow_version: str = tf.__version__
+    torch_version: str = torch.__version__
+    tensorflow_version: str = tf.__version__ if tf_is_available() else -1
 
     def to_dict(self) -> Dict[str, str]:
         return cattr.unstructure(self)
         if self.tensorflow_version != other.tensorflow_version:
             logger.warning(
                 "Tensorflow checkpoint was saved with a different version of Tensorflow. Model may not resume properly."
+            )
+        if self.torch_version != other.torch_version:
+            logger.warning(
+                "PyTorch checkpoint was saved with a different version of PyTorch. Model may not resume properly."
             )