save model

com.unity.ml-agents/Runtime/Training/MyTimeScaleSetting.cs

         set { m_Train = value; }
     }
 
+    [SerializeField]
+    bool m_LoadFile = true;
+    public bool LoadFile
+    {
+        get { return m_LoadFile; }
+        set { m_LoadFile = value; }
+    }
+
     // Start is called before the first frame update
     void Start()
     {

com.unity.ml-agents/Runtime/Training/Trainer.cs

                 return;
             }
 
-            for (int i = 0; i < m_Config.numSamplingAndUpdates; i++)
+            float loss = 0f;
+            for (var i = 0; i < m_Config.numSamplingAndUpdates; i++)
-                m_ModelRunner.UpdateModel(samples);
+                loss += m_ModelRunner.UpdateModel(samples);
-            UnityEngine.Debug.Log("Update !");
+            Debug.Log($"Loss: {loss/m_Config.numSamplingAndUpdates}");
+            m_ModelRunner.SaveModelToFile();
 
             // Update target network
             // if (m_TrainingStep % m_Config.updateTargetFreq == 0)

com.unity.ml-agents/Runtime/Training/TrainingModelRunner.cs

 using UnityEngine;
 using Unity.MLAgents.Inference.Utils;
 using System.Linq;
+using System.Runtime.Serialization.Formatters.Binary;
+using System.IO;
+using System;
 
 namespace Unity.MLAgents
 {
         bool m_TrainingObservationsInitialized;
 
         ReplayBuffer m_Buffer;
+        string m_ModelFileName = "Assets/model.dat";
 
         /// <summary>
         /// Initializes the Brain with the Model that it will use when selecting actions for
         void InitializeTrainingState()
         {
             var initState = m_Model.GetTensorByName(TensorNames.InitialTrainingState);
+            int[] stateShape = initState.shape.ToArray();
+            if (MyTimeScaleSetting.instance.LoadFile)
+            {
+                Debug.Log("load model");
+                initState = LoadModelFromFile(stateShape);
+            }
+
-                shape = initState.shape.ToArray().Select(i => (long)i).ToArray()
+                shape = stateShape.Select(i => (long)i).ToArray()
             };
         }
 
         //         Debug.Log(string.Join(", ", message));
         //     }
         // }
+
+        public void SaveModelToFile()
+        {
+            float[] array = m_TrainingState.data.ToReadOnlyArray();
+            var byteArray = new byte[array.Length * 4];
+            Buffer.BlockCopy(array, 0, byteArray, 0, byteArray.Length);
+            File.WriteAllBytes(m_ModelFileName, byteArray);
+            Debug.Log($"Save ModelParam: {m_TrainingState.data[0]}, {m_TrainingState.data[1]}, {m_TrainingState.data[2]}, " +
+                $"{m_TrainingState.data[3]}, {m_TrainingState.data[4]}, {m_TrainingState.data[5]}, " +
+                $"{m_TrainingState.data[6]}, {m_TrainingState.data[7]}, {m_TrainingState.data[8]}, {m_TrainingState.data[9]}");
+        }
+
+        public Tensor LoadModelFromFile(int[] shape)
+        {
+            var byteArray = File.ReadAllBytes(m_ModelFileName);
+            float[] array = new float[byteArray.Length / 4];
+            Buffer.BlockCopy(byteArray, 0, array, 0, byteArray.Length);
+            return new Tensor(shape, array);
+        }
     }
 }