mult state

Author: huailiang

Assets/Resources/Prefabs/Birds/Bird_1.prefab

@@ -903,7 +903,7 @@ Transform:
   m_PrefabInternal: {fileID: 0}
   m_GameObject: {fileID: 402347489}
   m_LocalRotation: {x: 0, y: 0, z: 0, w: 1}
-  m_LocalPosition: {x: 0, y: 2, z: 0}
+  m_LocalPosition: {x: 0, y: 1, z: 0}
   m_LocalScale: {x: 1, y: 1, z: 1}
   m_Children:
   - {fileID: 335634046}

Assets/Resources/tfModels/ppo.bytes

@@ -9,7 +9,7 @@ public abstract class BaseEnv : ScriptableObject
     protected int[] last_state;
     protected int total_r = 0;
     protected BirdAction last_action = BirdAction.PAD;
-    
+
     public int Score { get { return total_r; } }
 
     protected abstract bool birdFly { get; }
@@ -35,7 +35,7 @@ void OnScore(object arg)
 
     void OnDied(object arg)
     {
-        last_r = -100;
+        last_r = -2000;
     }
 
     public virtual void OnApplicationQuit() { }
@@ -71,7 +71,7 @@ public virtual void OnTick()
     public abstract BirdAction choose_action(int[] state);
 
     public abstract void UpdateState(int[] state, int[] state_, int rewd, BirdAction action);
-    
+
     public virtual void OnRestart(int[] state) { }
 
     public virtual void OnInspector() { }

Assets/Scene/3DBird.unity

@@ -49,7 +49,6 @@ public override void OnTick()
             {
                 UpdateState(last_state, state, last_r, last_action);
             }
-
             //do next loop
             BirdAction action = choose_action(state);
             GameMgr.S.RespondByDecision(action);
@@ -63,10 +62,12 @@ public override BirdAction choose_action(int[] state)
     {
 #if TensorFlow
         var runner = session.GetRunner();
-        float[,] sample = new float[1, 1];
-        // sample[0, 0] = state;
-        TFTensor t = new TFTensor(sample);
-        runner.AddInput(graph["state"][0], t);
+        float[,] fstate = new float[1, 3];
+        for (int i = 0; i < state.Length; i++)
+        {
+            fstate[0, i] = state[i];
+        }
+        runner.AddInput(graph["state"][0], fstate);
         runner.Fetch(graph["pi/probweights"][0]);
         TFTensor[] networkOutput;
         try
@@ -82,12 +83,11 @@ public override BirdAction choose_action(int[] state)
             }
             finally
             {
-                throw new System.Exception(errorMessage);
+                throw new System.Exception(errorMessage + "  \n" + e.StackTrace);
             }
         }
-        // Debug.Log(networkOutput.Length);
         float[,] output = (float[,])networkOutput[0].GetValue();
-        Debug.Log("choice action: " + output[0, 0] + "  " + output[0, 1]);
+        Debug.Log(string.Format("pi/probweights:{0},{1} ", output[0, 0], output[0, 1]));
         int rand = Random.Range(0, 100);
         return rand < (int)(output[0, 0] * 100) ? BirdAction.FLY : BirdAction.PAD;
 #else

Assets/Scripts/Core/BaseEnv.cs

@@ -3,7 +3,7 @@
 
 
 [CustomEditor(typeof(GameMgr))]
-public class GameManagerEditor : Editor
+public class GameMgrEditor : Editor
 {
     public override void OnInspectorGUI()
     {

Assets/Scripts/Core/InternalEnv.cs

@@ -48,7 +48,7 @@ void CreatePillar()
         currPillar = pillar;
         pillar.transform.position = new Vector3(EnvGlobalValue.PillarBornX, 0, 0);
         pillar.transform.localScale = Vector3.one;
-        int state = Random.Range(0, 2);
+        int state = 1;//Random.Range(0, 2);
         pillar.SetState(state);
         run_pool.Add(pillar);
     }

Assets/Scripts/Editor/GameManagerEditor.cs renamed to Assets/Scripts/Editor/GameMgrEditor.cs

@@ -29,13 +29,6 @@ public class GameMgr : MonoBehaviour
     public Bird mainBird;
     public PillarMgr pillMgr;
 
-    const float fpsMeasurePeriod = 0.5f;
-    private int m_FpsAccumulator = 0;
-    private float m_FpsNextPeriod = 0;
-    private float m_TotalTime = 0f;
-    private float m_CurrentFps;
-    private float m_SignTime = 0;
-
     public bool IsGameOver { get { return isGameOver; } }
 
     public bool IsGameStart { get { return isGameStart; } }
@@ -70,17 +63,10 @@ void Awake()
         env.Init();
     }
 
-
     void OnGUI()
     {
-        string str = string.Format("frame:{0} ", m_CurrentFps.ToString("f2"));
-        GUI.Label(new Rect(20, 20, 100, 30), str, style);
-        str = string.Format("runer:{0}", (Time.time - resetTime).ToString("f2"));
-        GUI.Label(new Rect(20, 50, 100, 30), str, style);
-        str = string.Format("epsln:{0}", epsilon);
-        GUI.Label(new Rect(20, 80, 100, 30), str, style);
-        str = string.Format("score:{0}", env.Score);
-        GUI.Label(new Rect(20, 110, 100, 30), str, style);
+        string str = string.Format("round:{0} timer:{1}", epsilon, (Time.time - resetTime).ToString("f2"));
+        GUI.Label(new Rect(30, 30, 100, 30), str, style);
     }
 
     void Update()
@@ -96,17 +82,6 @@ void Update()
         }
         env.OnUpdate(delta);
         pillMgr.Update(delta);
-
-        m_FpsAccumulator++;
-        m_TotalTime += Time.realtimeSinceStartup - m_SignTime;
-        m_SignTime = Time.realtimeSinceStartup;
-        if (Time.realtimeSinceStartup > m_FpsNextPeriod)
-        {
-            m_CurrentFps = m_FpsAccumulator / m_TotalTime;
-            m_TotalTime = 0;
-            m_FpsAccumulator = 0;
-            m_FpsNextPeriod = Time.realtimeSinceStartup + fpsMeasurePeriod;
-        }
     }
 
     public void ManuControl(bool fly)

Assets/Scripts/Editor/GameManagerEditor.cs.meta renamed to Assets/Scripts/Editor/GameMgrEditor.cs.meta

@@ -85,21 +85,21 @@ def _build_anet(self, name, trainable):
         with tf.variable_scope(name):
             l_1 = tf.layers.dense(self.tfs, 256, tf.nn.relu, trainable=trainable)
             a_prob = tf.layers.dense(l_1, A_DIM, tf.nn.softmax, trainable=trainable)
+            tf.identity(a_prob, name='probweights')
         params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=name)
         return a_prob, params
 
     def choose_action(self, s):
         prob_weights = self.sess.run(self.pi, feed_dict={self.tfs: s[None, :]})
         action = np.random.choice(range(prob_weights.shape[1]), p=prob_weights.ravel()) 
-        tf.identity(prob_weights, name='probweights')
         logger.info("action:{0} prob:{1}".format(str(action), str(prob_weights)))
         return action
 
     def get_v(self, s):
         return self.sess.run(self.v, {self.tfs: s})[0, 0]
 
     def output_nodes(self):
-        return ["state", "action", "advantage",  "critic/discounted_r", "probweights"]
+        return ["state", "action", "advantage",  "critic/discounted_r", "pi/probweights"]
 
     def freeze_graph(self):
         logger.info('**** Saved Model ****')