Merge pull request #1250 from SchoenTannenbaum/master

fix: regularizer serialization problem

Author: AsakusaRinne

src/TensorFlowNET.Core/Keras/Regularizers/IRegularizer.cs

@@ -1,7 +1,25 @@
-namespace Tensorflow.Keras
+using Newtonsoft.Json;
+using System.Collections.Generic;
+using Tensorflow.Keras.Saving.Common;
+
+namespace Tensorflow.Keras
 {
-    public interface IRegularizer
-    {
-        Tensor Apply(RegularizerArgs args);
-    }
+  [JsonConverter(typeof(CustomizedRegularizerJsonConverter))]
+  public interface IRegularizer
+  {
+    [JsonProperty("class_name")]
+    string ClassName { get; }
+    [JsonProperty("config")]
+    IDictionary<string, object> Config { get; }
+    Tensor Apply(RegularizerArgs args);
+  }
+
+  public interface IRegularizerApi
+  {
+    IRegularizer GetRegularizerFromName(string name);
+    IRegularizer L1 { get; }
+    IRegularizer L2 { get; }
+    IRegularizer L1L2 { get; }
+  }
+
 }

src/TensorFlowNET.Core/Keras/Saving/Json/CustomizedRegularizerJsonConverter.cs

@@ -0,0 +1,57 @@
+using Newtonsoft.Json.Linq;
+using Newtonsoft.Json;
+using System;
+using System.Collections.Generic;
+using System.Text;
+using Tensorflow.Operations.Regularizers;
+
+namespace Tensorflow.Keras.Saving.Common
+{
+  class RegularizerInfo
+  {
+    public string class_name { get; set; }
+    public JObject config { get; set; }
+  }
+
+  public class CustomizedRegularizerJsonConverter : JsonConverter
+    {
+        public override bool CanConvert(Type objectType)
+        {
+            return objectType == typeof(IRegularizer);
+        }
+
+        public override bool CanRead => true;
+
+        public override bool CanWrite => true;
+
+        public override void WriteJson(JsonWriter writer, object? value, JsonSerializer serializer)
+        {
+            var regularizer = value as IRegularizer;
+            if (regularizer is null)
+            {
+                JToken.FromObject(null).WriteTo(writer);
+                return;
+            }
+            JToken.FromObject(new RegularizerInfo()
+            {
+              class_name = regularizer.ClassName,
+              config = JObject.FromObject(regularizer.Config)
+            }, serializer).WriteTo(writer);
+        }
+
+        public override object? ReadJson(JsonReader reader, Type objectType, object? existingValue, JsonSerializer serializer)
+        {
+            var info = serializer.Deserialize<RegularizerInfo>(reader);
+            if (info is null)
+            {
+                return null;
+            }
+            return info.class_name switch
+            {
+                "L1L2" => new L1L2 (info.config["l1"].ToObject<float>(), info.config["l2"].ToObject<float>()),
+                "L1" => new L1(info.config["l1"].ToObject<float>()),
+                "L2" => new L2(info.config["l2"].ToObject<float>()),
+            };
+        }
+    }
+}

src/TensorFlowNET.Core/Operations/Regularizers/L1.cs

@@ -0,0 +1,33 @@
+using System;
+
+using Tensorflow.Keras;
+
+namespace Tensorflow.Operations.Regularizers
+{
+  public class L1 : IRegularizer
+  {
+    float _l1;
+    private readonly Dictionary<string, object> _config;
+
+    public string ClassName => "L1";
+    public virtual IDictionary<string, object> Config => _config;
+
+    public L1(float l1 = 0.01f)
+    {
+      //  l1 = 0.01 if l1 is None else l1
+      //  validate_float_arg(l1, name = "l1")
+      //  self.l1 = ops.convert_to_tensor(l1)
+      this._l1 = l1;
+
+      _config = new();
+      _config["l1"] = _l1;
+    }
+
+
+    public Tensor Apply(RegularizerArgs args)
+    {
+      //return self.l1 * ops.sum(ops.absolute(x))
+      return _l1 * math_ops.reduce_sum(math_ops.abs(args.X));
+    }
+  }
+}

src/TensorFlowNET.Core/Operations/Regularizers/L1L2.cs

@@ -0,0 +1,48 @@
+using System;
+
+using Tensorflow.Keras;
+
+namespace Tensorflow.Operations.Regularizers
+{
+  public class L1L2 : IRegularizer
+  {
+    float _l1;
+    float _l2;
+    private readonly Dictionary<string, object> _config;
+
+    public string ClassName => "L1L2";
+    public virtual IDictionary<string, object> Config => _config;
+
+    public L1L2(float l1 = 0.0f, float l2 = 0.0f)
+    {
+      //l1 = 0.0 if l1 is None else l1
+      //l2 = 0.0 if l2 is None else l2
+      //  validate_float_arg(l1, name = "l1")
+      //  validate_float_arg(l2, name = "l2")
+
+      //  self.l1 = l1
+      //  self.l2 = l2
+      this._l1 = l1;
+      this._l2 = l2;
+
+      _config = new();
+      _config["l1"] = l1;
+      _config["l2"] = l2;
+    }
+
+    public Tensor Apply(RegularizerArgs args)
+    {
+        //regularization = ops.convert_to_tensor(0.0, dtype = x.dtype)
+        //if self.l1:
+        //    regularization += self.l1 * ops.sum(ops.absolute(x))
+        //if self.l2:
+        //    regularization += self.l2 * ops.sum(ops.square(x))
+        //return regularization
+
+        Tensor regularization = tf.constant(0.0, args.X.dtype);
+        regularization += _l1 * math_ops.reduce_sum(math_ops.abs(args.X));
+        regularization += _l2 * math_ops.reduce_sum(math_ops.square(args.X));
+        return regularization;
+    }
+  }
+}

src/TensorFlowNET.Core/Operations/Regularizers/L2.cs

@@ -0,0 +1,33 @@
+using System;
+
+using Tensorflow.Keras;
+
+namespace Tensorflow.Operations.Regularizers
+{
+  public class L2 : IRegularizer
+  {
+    float _l2;
+    private readonly Dictionary<string, object> _config;
+
+    public string ClassName => "L2";
+    public virtual IDictionary<string, object> Config => _config;
+
+    public L2(float l2 = 0.01f)
+    {
+      //  l2 = 0.01 if l2 is None else l2
+      //  validate_float_arg(l2, name = "l2")
+      //  self.l2 = l2
+      this._l2 = l2;
+
+      _config = new();
+      _config["l2"] = _l2;
+    }
+
+
+    public Tensor Apply(RegularizerArgs args)
+    {
+      //return self.l2 * ops.sum(ops.square(x))
+      return _l2 * math_ops.reduce_sum(math_ops.square(args.X));
+    }
+  }
+}

src/TensorFlowNET.Keras/Regularizers.cs

@@ -1,8 +1,51 @@
-namespace Tensorflow.Keras
+using Tensorflow.Operations.Regularizers;
+
+namespace Tensorflow.Keras
 {
-    public class Regularizers
+  public class Regularizers: IRegularizerApi
+  {
+    private static Dictionary<string, IRegularizer> _nameActivationMap;
+
+    public IRegularizer l1(float l1 = 0.01f)
+        => new L1(l1);
+    public IRegularizer l2(float l2 = 0.01f)
+        => new L2(l2);
+
+    //From TF source
+    //# The default value for l1 and l2 are different from the value in l1_l2
+    //# for backward compatibility reason. Eg, L1L2(l2=0.1) will only have l2
+    //# and no l1 penalty.
+    public IRegularizer l1l2(float l1 = 0.00f, float l2 = 0.00f)
+        => new L1L2(l1, l2);
+
+    static Regularizers()
     {
-        public IRegularizer l2(float l2 = 0.01f)
-            => new L2(l2);
+      _nameActivationMap = new Dictionary<string, IRegularizer>();
+      _nameActivationMap["L1"] = new L1();
+      _nameActivationMap["L1"] = new L2();
+      _nameActivationMap["L1"] = new L1L2();
     }
+
+    public IRegularizer L1 => l1();
+
+    public IRegularizer L2 => l2();
+
+    public IRegularizer L1L2 => l1l2();
+
+    public IRegularizer GetRegularizerFromName(string name)
+    {
+      if (name == null)
+      {
+        throw new Exception($"Regularizer name cannot be null");
+      }
+      if (!_nameActivationMap.TryGetValue(name, out var res))
+      {
+        throw new Exception($"Regularizer {name} not found");
+      }
+      else
+      {
+        return res;
+      }
+    }
+  }
 }

src/TensorFlowNET.Keras/Regularizers/L1.cs

@@ -1,6 +1,7 @@
 using Microsoft.VisualStudio.TestPlatform.Utilities;
 using Microsoft.VisualStudio.TestTools.UnitTesting;
 using Newtonsoft.Json.Linq;
+using System.Collections.Generic;
 using System.Linq;
 using System.Xml.Linq;
 using Tensorflow.Keras.Engine;
@@ -129,6 +130,53 @@ public void TestModelBeforeTF2_5()
     }
 
 
+    [TestMethod]
+    public void BiasRegularizerSaveAndLoad()
+    {
+        var savemodel = keras.Sequential(new List<ILayer>()
+            {
+                tf.keras.layers.InputLayer((227, 227, 3)),
+                tf.keras.layers.Conv2D(96, (11, 11), (4, 4), activation:"relu", padding:"valid"),
+                tf.keras.layers.BatchNormalization(),
+                tf.keras.layers.MaxPooling2D((3, 3), strides:(2, 2)),
+
+                tf.keras.layers.Conv2D(256, (5, 5), (1, 1), "same", activation: keras.activations.Relu, bias_regularizer:keras.regularizers.L1L2),
+                tf.keras.layers.BatchNormalization(),
+
+                tf.keras.layers.Conv2D(256, (5, 5), (1, 1), "same", activation: keras.activations.Relu, bias_regularizer:keras.regularizers.L2),
+                tf.keras.layers.BatchNormalization(),
+
+                tf.keras.layers.Conv2D(256, (5, 5), (1, 1), "same", activation: keras.activations.Relu, bias_regularizer:keras.regularizers.L1),
+                tf.keras.layers.BatchNormalization(),
+                tf.keras.layers.MaxPooling2D((3, 3), (2, 2)),
+
+                tf.keras.layers.Flatten(),
+
+                tf.keras.layers.Dense(1000, activation: "linear"),
+                tf.keras.layers.Softmax(1)
+            });
+
+        savemodel.compile(tf.keras.optimizers.Adam(), tf.keras.losses.SparseCategoricalCrossentropy(from_logits: true), new string[] { "accuracy" });
+
+        var num_epochs = 1;
+        var batch_size = 8;
+
+        var trainDataset = new RandomDataSet(new Shape(227, 227, 3), 16);
+
+        savemodel.fit(trainDataset.Data, trainDataset.Labels, batch_size, num_epochs);
+
+        savemodel.save(@"./bias_regularizer_save_and_load", save_format: "tf");
+
+        var loadModel = tf.keras.models.load_model(@"./bias_regularizer_save_and_load");
+        loadModel.summary();
+
+        loadModel.compile(tf.keras.optimizers.Adam(), tf.keras.losses.SparseCategoricalCrossentropy(from_logits: true), new string[] { "accuracy" });
+
+        var fitDataset = new RandomDataSet(new Shape(227, 227, 3), 16);
+
+        loadModel.fit(fitDataset.Data, fitDataset.Labels, batch_size, num_epochs);
+    }
+
 
     [TestMethod]
     public void CreateConcatenateModelSaveAndLoad()