Generic cleanup rest of framework, activations and initializers

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/Activation.java

@@ -20,13 +20,8 @@
 
 /**
  * Abstract base class for Activations
- *
- * <p><b>Note:</b> The {@link #tf} attribute must be set prior to invoking the call method. See
- * {@link #setTF(Ops)} and the constructor {@link #Activation(Ops)}.
- *
- * @param <T> the data type of the activation
  */
-public abstract class Activation<T extends TNumber> {
+public abstract class Activation {
 
   /** The TensorFlow Ops */
   protected Ops tf;
@@ -41,28 +36,29 @@ protected Activation(Ops tf) {
   }
 
   /**
-   * Sets the TensorFlow Ops
+   * Gets the TensorFlow Ops
    *
-   * @param tf the TensorFlow Ops
+   * @return the TensorFlow Ops
    */
-  protected void setTF(Ops tf) {
-    this.tf = tf;
+  protected Ops getTF() {
+    return this.tf;
   }
 
   /**
-   * Gets the TensorFlow Ops
+   * Sets the TensorFlow Ops
    *
-   * @return the TensorFlow Ops
+   * @param tf the TensorFlow Ops
    */
-  protected Ops getTF() {
-    return this.tf;
+  protected void setTF(Ops tf) {
+    this.tf = tf;
   }
 
   /**
    * Gets the calculation operation for the activation.
    *
    * @param input the input tensor
    * @return The operand for the activation
+   * @param <T> the data type of the input and result
    */
-  public abstract Operand<T> call(Operand<T> input);
+  public abstract <T extends TNumber> Operand<T> call(Operand<T> input);
 }

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/ELU.java

@@ -17,7 +17,7 @@
 import org.tensorflow.Operand;
 import org.tensorflow.op.Ops;
 import org.tensorflow.types.TBool;
-import org.tensorflow.types.family.TFloating;
+import org.tensorflow.types.family.TNumber;
 
 /**
  * Exponential linear unit.
@@ -44,11 +44,10 @@
  *     Operand<TFloat32> result = elu.call(input);
  * </pre>
  *
- * @param <T> the data type of the activation
  * @see <a href="https://arxiv.org/abs/1511.07289">Clevert et al, 2016, Fast and Accurate Deep
  *     Network Learning by Exponential Linear Units (ELUs)</a>
  */
-public class ELU<T extends TFloating> extends Activation<T> {
+public class ELU extends Activation {
 
   private static final double ALPHA_DEFAULT = 1.0;
 
@@ -83,11 +82,12 @@ public ELU(Ops tf, double alpha) {
    * @return The operand for the activation
    */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
 
     Operand<T> result = tf.nn.elu(input);
-    if (alpha == 1.0) return result;
-    else {
+    if (alpha == 1.0) {
+      return result;
+    } else {
       Class<T> inputType = input.type();
       Operand<T> y = tf.math.mul(result, tf.dtypes.cast(tf.constant(alpha), inputType));
       Operand<TBool> cond = tf.math.greater(result, tf.dtypes.cast(tf.constant(0), inputType));

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/Exponential.java

@@ -16,7 +16,7 @@
 
 import org.tensorflow.Operand;
 import org.tensorflow.op.Ops;
-import org.tensorflow.types.family.TFloating;
+import org.tensorflow.types.family.TNumber;
 
 /**
  * Exponential activation function.
@@ -30,10 +30,8 @@
  *   Operand<TFloat32> result = exp.call(input);
  *   // result is [0.04978707f,  0.36787945f,  1.f,  2.7182817f, 20.085537f]
  * </pre>
- *
- * @param <T> the data type of the activation
  */
-public class Exponential<T extends TFloating> extends Activation<T> {
+public class Exponential extends Activation {
 
   /**
    * Creates an Exponential activation.
@@ -49,9 +47,10 @@ public Exponential(Ops tf) {
    *
    * @param input the input tensor
    * @return an Operand for the exponential activation: <code>exp(x)</code>.
+   * @param <T> the data type of the activation
    */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
     return tf.math.exp(input);
   }
 }

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/HardSigmoid.java

@@ -16,7 +16,7 @@
 
 import org.tensorflow.Operand;
 import org.tensorflow.op.Ops;
-import org.tensorflow.types.family.TFloating;
+import org.tensorflow.types.family.TNumber;
 
 /**
  * Hard sigmoid activation.
@@ -40,10 +40,8 @@
  *     Operand<TFloat32> result = hardSigmoid.call(input);
  *     // result is [0.f , 0.3f, 0.5f, 0.7f, 1.f]
  * </pre>
- *
- * @param <T> the data type of the result
  */
-public class HardSigmoid<T extends TFloating> extends Activation<T> {
+public class HardSigmoid extends Activation {
 
   /**
    * Creates Hard sigmoid activation.
@@ -59,9 +57,10 @@ public HardSigmoid(Ops tf) {
    *
    * @param input the input tensor
    * @return The operand for the activation
+   * @param <T> the data type of the result
    */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
     Class<T> inputType = input.type();
     Operand<T> point2 = tf.dtypes.cast(tf.constant(0.2), inputType);
     Operand<T> point5 = tf.dtypes.cast(tf.constant(0.5), inputType);

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/Linear.java

@@ -19,9 +19,9 @@
 import org.tensorflow.types.family.TNumber;
 
 /**
- * Linear activation function  (pass-through).
+ * Linear activation function (pass-through).
  *
- * <p>The linear activation returns its input. It is also known as the Identity activation function.</p>
+ * <p>The linear activation returns its input. It is also known as the Identity activation function.
  *
  * <p>For example:
  *
@@ -33,7 +33,7 @@
  *    // result is [-3.0f,-1.0f, 0.0f,1.0f,3.0f]
  * </pre>
  */
-public class Linear<U extends TNumber> extends Activation<U> {
+public class Linear extends Activation {
 
   /**
    * Creates a linear activation.
@@ -46,7 +46,7 @@ public Linear(Ops tf) {
 
   /** {@inheritDoc} */
   @Override
-  public Operand<U> call(Operand<U> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
     return input;
   }
 }

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/ReLU.java

@@ -55,10 +55,8 @@
  *     result = relu.call(input);
  *     // result is [-0.f, -0.f,  0.f,  0.f, 10.f]
  * </pre>
- *
- * @param <T> the data type of the result
  */
-public class ReLU<T extends TNumber> extends Activation<T> {
+public class ReLU extends Activation {
 
   public static final float ALPHA_DEFAULT = 0.0f;
   public static final float MAX_VALUE_DEFAULT = Float.NaN;
@@ -96,7 +94,7 @@ public ReLU(Ops tf, float alpha, float maxValue, float threshold) {
 
   /** {@inheritDoc} */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
     Class<T> inputType = input.type();
 
     boolean clipMax = !Float.isNaN(maxValue);

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/SELU.java

@@ -16,7 +16,7 @@
 
 import org.tensorflow.Operand;
 import org.tensorflow.op.Ops;
-import org.tensorflow.types.family.TFloating;
+import org.tensorflow.types.family.TNumber;
 
 /**
  * Scaled Exponential Linear Unit (SELU).
@@ -42,10 +42,9 @@
  * <p><b>Notes: </b> To be used together with the {@link
  * org.tensorflow.framework.initializers.LeCun} initializer with Normal Distribution.
  *
- * @param <T> the data type of the activation
  * @see <a href="https://arxiv.org/abs/1706.02515">Klambauer et al., 2017</a>
  */
-public class SELU<T extends TFloating> extends Activation<T> {
+public class SELU extends Activation {
 
   /**
    * Creates a Scaled Exponential Linear Unit (SELU) activation.
@@ -61,9 +60,10 @@ public SELU(Ops tf) {
    *
    * @param input the input tensor
    * @return The operand for the activation
+   * @param <T> the data type of the activation
    */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
     return tf.nn.selu(input);
   }
 }

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/Sigmoid.java

@@ -16,7 +16,7 @@
 
 import org.tensorflow.Operand;
 import org.tensorflow.op.Ops;
-import org.tensorflow.types.family.TFloating;
+import org.tensorflow.types.family.TNumber;
 
 /**
  * Sigmoid activation. <code>sigmoid(x) = 1 / (1 + exp(-x))</code>.
@@ -38,10 +38,8 @@
  *     // result is [2.0611537e-09f, 2.6894143e-01f,
  *     //                 5.0000000e-01f,7.3105860e-01f, 1.f]
  * </pre>
- *
- * @param <T> the data type of the activation
  */
-public class Sigmoid<T extends TFloating> extends Activation<T> {
+public class Sigmoid extends Activation {
 
   /**
    * Creates a Sigmoid activation.
@@ -57,9 +55,10 @@ public Sigmoid(Ops tf) {
    *
    * @param input the input tensor
    * @return The operand for the activation
+   * @param <T> the data type of the activation
    */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
     return tf.math.sigmoid(input);
   }
 }

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/Softmax.java

@@ -19,7 +19,7 @@
 import org.tensorflow.op.Ops;
 import org.tensorflow.op.core.ReduceMax;
 import org.tensorflow.op.core.ReduceSum;
-import org.tensorflow.types.family.TFloating;
+import org.tensorflow.types.family.TNumber;
 
 /**
  * Softmax converts a real vector to a vector of categorical probabilities.
@@ -35,10 +35,8 @@
  * <p>The softmax of each vector x is computed as: <code>exp(x) / tf.sum(exp(x))</code>.
  *
  * <p>The input values in are the log-odds of the resulting probability.
- *
- * @param <T> the data type of the activation
  */
-public class Softmax<T extends TFloating> extends Activation<T> {
+public class Softmax extends Activation {
 
   private static final int AXIS_DEFAULT = -1;
 
@@ -70,9 +68,10 @@ public Softmax(Ops tf, int axis) {
    *
    * @param input the input tensor
    * @return The operand for the activation
+   * @param <T> the data type of the activation
    */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
     Shape shape = input.shape();
     int numDimensions = shape.numDimensions();
     if (numDimensions == 2) {

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/Softplus.java

@@ -16,7 +16,7 @@
 
 import org.tensorflow.Operand;
 import org.tensorflow.op.Ops;
-import org.tensorflow.types.family.TFloating;
+import org.tensorflow.types.family.TNumber;
 
 /**
  * Softplus activation function, <code>softplus(x) = log(exp(x) + 1)</code>.
@@ -32,7 +32,7 @@
  *     //                 1.3132616e+00f, 2.0000000e+01f]
  * </pre>
  */
-public class Softplus<T extends TFloating> extends Activation<T> {
+public class Softplus extends Activation {
 
   /**
    * Creates a Softplus activation function.
@@ -50,7 +50,7 @@ public Softplus(Ops tf) {
    * @return The operand for the activation
    */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
     return tf.math.softplus(input);
   }
 }

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/Softsign.java

@@ -16,7 +16,7 @@
 
 import org.tensorflow.Operand;
 import org.tensorflow.op.Ops;
-import org.tensorflow.types.family.TFloating;
+import org.tensorflow.types.family.TNumber;
 
 /**
  * Softsign activation function, <code>softsign(x) = x / (abs(x) + 1)</code>.
@@ -30,10 +30,8 @@
  *     Operand&lt;TFloat32&gt; result = softsign.call(input);
  *     // result is [-0.5f, 0.f, 0.5f]
  * </pre>
- *
- * @param <T> the data type of the activation
  */
-public class Softsign<T extends TFloating> extends Activation<T> {
+public class Softsign extends Activation {
 
   /**
    * Creates a Softsign activation.
@@ -49,9 +47,10 @@ public Softsign(Ops tf) {
    *
    * @param input the input tensor
    * @return The operand for the activation
+   * @param <T> the data type of the activation
    */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
     return tf.nn.softsign(input);
   }
 }

tensorflow-framework/src/main/java/org/tensorflow/framework/activations/Swish.java

@@ -16,7 +16,7 @@
 
 import org.tensorflow.Operand;
 import org.tensorflow.op.Ops;
-import org.tensorflow.types.family.TFloating;
+import org.tensorflow.types.family.TNumber;
 
 /**
  * Swish activation function. <code>swish(x) = x * sigmoid(x)</code>.
@@ -37,10 +37,9 @@
  *
  * </pre>
  *
- * @param <T> the data type of the activation
  * @see <a href="https://arxiv.org/abs/1710.05941">Ramachandran et al., 2017</a>
  */
-public class Swish<T extends TFloating> extends Activation<T> {
+public class Swish extends Activation {
 
   /**
    * Creates a Swish activation, <code>swish(x) = x * sigmoid(x)</code>.
@@ -57,7 +56,7 @@ public Swish(Ops tf) {
 
   /** {@inheritDoc} */
   @Override
-  public Operand<T> call(Operand<T> input) {
+  public <T extends TNumber> Operand<T> call(Operand<T> input) {
 
     // TODO Python Keras returns a "grad", which is an optimization not implemented in Java.
     return tf.math.mul(input, tf.math.sigmoid(input));