* When 'hasPeepholeConnections' is true, this is the "vanilla" variant in said paper
* When 'hasPeepholeConnections' is false, this is the "no peephole" variant
- * http://arxiv.org/pdf/1503.04069.pdf
+ * https://arxiv.org/pdf/1503.04069.pdf
*
*
* @author Alex Black (LSTM implementations)
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitIdentity.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitIdentity.java
index 076fa2ac8..b25121cd3 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitIdentity.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitIdentity.java
@@ -16,11 +16,14 @@
package org.deeplearning4j.nn.weights;
+import lombok.Data;
import lombok.EqualsAndHashCode;
+import lombok.NoArgsConstructor;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.indexing.INDArrayIndex;
import org.nd4j.linalg.indexing.NDArrayIndex;
+import org.nd4j.shade.jackson.annotation.JsonProperty;
import java.util.Arrays;
@@ -32,9 +35,17 @@ import java.util.Arrays;
*
* @author Adam Gibson
*/
-@EqualsAndHashCode
+@Data
+@NoArgsConstructor
public class WeightInitIdentity implements IWeightInit {
+ private Double scale;
+
+ public WeightInitIdentity(@JsonProperty("scale") Double scale){
+ this.scale = scale;
+ }
+
+
@Override
public INDArray init(double fanIn, double fanOut, long[] shape, char order, INDArray paramView) {
if (shape[0] != shape[1]) {
@@ -59,6 +70,11 @@ public class WeightInitIdentity implements IWeightInit {
} else {
ret = Nd4j.createUninitialized(shape, order).assign(Nd4j.eye(shape[0]));
}
+
+ if(scale != null){
+ ret.muli(scale);
+ }
+
INDArray flat = Nd4j.toFlattened(order, ret);
paramView.assign(flat);
return paramView.reshape(order, shape);
@@ -82,13 +98,16 @@ public class WeightInitIdentity implements IWeightInit {
indArrayIndices[i] = NDArrayIndex.point(shape[i] / 2);
}
- paramView.assign(Nd4j.zeros(paramView.shape()));
+ paramView.assign(0);
final INDArray params =paramView.reshape(order, shape);
for (int i = 0; i < shape[0]; i++) {
indArrayIndices[0] = NDArrayIndex.point(i);
indArrayIndices[1] = NDArrayIndex.point(i);
params.put(indArrayIndices, Nd4j.ones(1));
}
+ if(scale != null){
+ params.muli(scale);
+ }
return params;
}
}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitUtil.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitUtil.java
index b110bc5a0..17034d408 100755
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitUtil.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitUtil.java
@@ -19,6 +19,7 @@ package org.deeplearning4j.nn.weights;
import org.apache.commons.math3.util.FastMath;
import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.api.ops.random.impl.TruncatedNormalDistribution;
import org.nd4j.linalg.api.rng.distribution.Distribution;
import org.nd4j.linalg.api.rng.distribution.impl.OrthogonalDistribution;
import org.nd4j.linalg.factory.Nd4j;
@@ -146,14 +147,13 @@ public class WeightInitUtil {
paramView.assign(flat);
break;
case VAR_SCALING_NORMAL_FAN_IN:
- // TODO: needs to be truncated normal to match keras.
- Nd4j.randn(paramView).divi(FastMath.sqrt(fanIn));
+ Nd4j.exec(new TruncatedNormalDistribution(paramView, 0.0, Math.sqrt(1.0 / fanIn)));
break;
case VAR_SCALING_NORMAL_FAN_OUT:
- Nd4j.randn(paramView).divi(FastMath.sqrt(fanOut));
+ Nd4j.exec(new TruncatedNormalDistribution(paramView, 0.0, Math.sqrt(1.0 / fanOut)));
break;
case VAR_SCALING_NORMAL_FAN_AVG:
- Nd4j.randn(paramView).divi(FastMath.sqrt((fanIn + fanOut) / 2));
+ Nd4j.exec(new TruncatedNormalDistribution(paramView, 0.0, Math.sqrt(2.0 / (fanIn + fanOut))));
break;
case VAR_SCALING_UNIFORM_FAN_IN:
double scalingFanIn = 3.0 / Math.sqrt(fanIn);
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanAvg.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanAvg.java
index 0be5af0e9..3b9698f10 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanAvg.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanAvg.java
@@ -16,22 +16,39 @@
package org.deeplearning4j.nn.weights;
+import lombok.Data;
import lombok.EqualsAndHashCode;
+import lombok.NoArgsConstructor;
import org.apache.commons.math3.util.FastMath;
import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.api.ops.random.impl.TruncatedNormalDistribution;
import org.nd4j.linalg.factory.Nd4j;
/**
- * Gaussian distribution with mean 0, variance 1.0/((fanIn + fanOut)/2)
+ * Truncated aussian distribution with mean 0, variance 1.0/((fanIn + fanOut)/2)
*
* @author Adam Gibson
*/
-@EqualsAndHashCode
+@Data
+@NoArgsConstructor
public class WeightInitVarScalingNormalFanAvg implements IWeightInit {
+ private Double scale;
+
+ public WeightInitVarScalingNormalFanAvg(Double scale){
+ this.scale = scale;
+ }
+
@Override
public INDArray init(double fanIn, double fanOut, long[] shape, char order, INDArray paramView) {
- Nd4j.randn(paramView).divi(FastMath.sqrt((fanIn + fanOut) / 2));
+ double std;
+ if(scale == null){
+ std = Math.sqrt(2.0 / (fanIn + fanOut));
+ } else {
+ std = Math.sqrt(2.0 * scale / (fanIn + fanOut));
+ }
+
+ Nd4j.exec(new TruncatedNormalDistribution(paramView, 0.0, std));
return paramView.reshape(order, shape);
}
}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanIn.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanIn.java
index 3f89ff015..dca457de3 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanIn.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanIn.java
@@ -16,23 +16,38 @@
package org.deeplearning4j.nn.weights;
-import lombok.EqualsAndHashCode;
-import org.apache.commons.math3.util.FastMath;
+import lombok.Data;
+import lombok.NoArgsConstructor;
import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.api.ops.random.impl.TruncatedNormalDistribution;
import org.nd4j.linalg.factory.Nd4j;
/**
- * Gaussian distribution with mean 0, variance 1.0/(fanIn)
+ * Gaussian distribution with mean 0, variance {@code 1.0/(fanIn)}
+ * If a scale is provided, use variance {@code scale/(fanIn)} instead
*
* @author Adam Gibson
*/
-@EqualsAndHashCode
+@Data
+@NoArgsConstructor
public class WeightInitVarScalingNormalFanIn implements IWeightInit {
+ private Double scale;
+
+ public WeightInitVarScalingNormalFanIn(Double scale){
+ this.scale = scale;
+ }
+
@Override
public INDArray init(double fanIn, double fanOut, long[] shape, char order, INDArray paramView) {
- // TODO: needs to be truncated normal to match keras.
- Nd4j.randn(paramView).divi(FastMath.sqrt(fanIn));
+ double std;
+ if(scale == null){
+ std = Math.sqrt(1.0 / fanIn);
+ } else {
+ std = Math.sqrt(scale / fanIn);
+ }
+
+ Nd4j.exec(new TruncatedNormalDistribution(paramView, 0.0, std));
return paramView.reshape(order, shape);
}
}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanOut.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanOut.java
index 6369a19c6..0af43ac88 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanOut.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingNormalFanOut.java
@@ -16,22 +16,40 @@
package org.deeplearning4j.nn.weights;
+import lombok.Data;
import lombok.EqualsAndHashCode;
+import lombok.NoArgsConstructor;
import org.apache.commons.math3.util.FastMath;
import org.nd4j.linalg.api.ndarray.INDArray;
+import org.nd4j.linalg.api.ops.random.impl.TruncatedNormalDistribution;
import org.nd4j.linalg.factory.Nd4j;
/**
- * Gaussian distribution with mean 0, variance 1.0/(fanOut)
+ * Truncated normal distribution with mean 0, variance 1.0/(fanOut)
+ * If a scale is provided, variance is scale / fanOut
*
* @author Adam Gibson
*/
-@EqualsAndHashCode
+@Data
+@NoArgsConstructor
public class WeightInitVarScalingNormalFanOut implements IWeightInit {
+ private Double scale;
+
+ public WeightInitVarScalingNormalFanOut(Double scale){
+ this.scale = scale;
+ }
+
@Override
public INDArray init(double fanIn, double fanOut, long[] shape, char order, INDArray paramView) {
- Nd4j.randn(paramView).divi(FastMath.sqrt(fanOut));
+ double std;
+ if(scale == null){
+ std = Math.sqrt(1.0 / fanOut);
+ } else {
+ std = Math.sqrt(scale / fanOut);
+ }
+
+ Nd4j.exec(new TruncatedNormalDistribution(paramView, 0.0, std));
return paramView.reshape(order, shape);
}
}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanAvg.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanAvg.java
index afb1a1dc8..f2e050e6e 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanAvg.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanAvg.java
@@ -16,7 +16,9 @@
package org.deeplearning4j.nn.weights;
+import lombok.Data;
import lombok.EqualsAndHashCode;
+import lombok.NoArgsConstructor;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.factory.Nd4j;
@@ -25,12 +27,22 @@ import org.nd4j.linalg.factory.Nd4j;
*
* @author Adam Gibson
*/
-@EqualsAndHashCode
+@Data
+@NoArgsConstructor
public class WeightInitVarScalingUniformFanAvg implements IWeightInit {
+ private Double scale;
+
+ public WeightInitVarScalingUniformFanAvg(Double scale){
+ this.scale = scale;
+ }
+
@Override
public INDArray init(double fanIn, double fanOut, long[] shape, char order, INDArray paramView) {
double scalingFanAvg = 3.0 / Math.sqrt((fanIn + fanOut) / 2);
+ if(scale != null)
+ scalingFanAvg *= scale;
+
Nd4j.rand(paramView, Nd4j.getDistributions().createUniform(-scalingFanAvg, scalingFanAvg));
return paramView.reshape(order, shape);
}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanIn.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanIn.java
index 0cf26ecc6..7135394a7 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanIn.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanIn.java
@@ -16,21 +16,34 @@
package org.deeplearning4j.nn.weights;
+import lombok.Data;
import lombok.EqualsAndHashCode;
+import lombok.NoArgsConstructor;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.factory.Nd4j;
/**
- * Uniform U[-a,a] with a=3.0/(fanIn)
+ * Uniform U[-a,a] with a=3.0/(fanIn)
+ * If a scale is provided, a = 3.0 * scale / (fanIn)
*
* @author Adam Gibson
*/
-@EqualsAndHashCode
+@NoArgsConstructor
+@Data
public class WeightInitVarScalingUniformFanIn implements IWeightInit {
+ private Double scale;
+
+ public WeightInitVarScalingUniformFanIn(Double scale){
+ this.scale = scale;
+ }
+
@Override
public INDArray init(double fanIn, double fanOut, long[] shape, char order, INDArray paramView) {
double scalingFanIn = 3.0 / Math.sqrt(fanIn);
+ if(scale != null)
+ scalingFanIn *= scale;
+
Nd4j.rand(paramView, Nd4j.getDistributions().createUniform(-scalingFanIn, scalingFanIn));
return paramView.reshape(order, shape);
}
diff --git a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanOut.java b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanOut.java
index 2d3b116fc..09bf2053d 100644
--- a/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanOut.java
+++ b/deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/weights/WeightInitVarScalingUniformFanOut.java
@@ -16,21 +16,33 @@
package org.deeplearning4j.nn.weights;
+import lombok.Data;
import lombok.EqualsAndHashCode;
+import lombok.NoArgsConstructor;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.factory.Nd4j;
/**
- * Uniform U[-a,a] with a=3.0/(fanOut)
+ * Uniform U[-a,a] with a=3.0/(fanOut)
+ * If a scale is provided, a = 3.0 * scale / fanOut
*
* @author Adam Gibson
*/
-@EqualsAndHashCode
+@Data
+@NoArgsConstructor
public class WeightInitVarScalingUniformFanOut implements IWeightInit {
+ private Double scale;
+
+ public WeightInitVarScalingUniformFanOut(Double scale){
+ this.scale = scale;
+ }
+
@Override
public INDArray init(double fanIn, double fanOut, long[] shape, char order, INDArray paramView) {
double scalingFanOut = 3.0 / Math.sqrt(fanOut);
+ if(scale != null)
+ scalingFanOut *= scale;
Nd4j.rand(paramView, Nd4j.getDistributions().createUniform(-scalingFanOut, scalingFanOut));
return paramView.reshape(order, shape);
}
diff --git a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-aws/pom.xml b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-aws/pom.xml
index 7c9967ef8..94f66b405 100644
--- a/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-aws/pom.xml
+++ b/deeplearning4j/deeplearning4j-scaleout/deeplearning4j-aws/pom.xml
@@ -44,184 +44,48 @@
- * See: http://arxiv.org/abs/1511.07289
+ * See: https://arxiv.org/abs/1511.07289
*
* @param x Input variable
* @return Output variable
@@ -154,7 +154,7 @@ public class SDNN extends SDOps {
* out = a * (exp(x) - 1) if x <= 0
* with constant a = 1.0
*
- * See: http://arxiv.org/abs/1511.07289
+ * See: https://arxiv.org/abs/1511.07289
*
* @param name Output variable name
* @param x Input variable
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationRReLU.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationRReLU.java
index 478305e76..9221d601c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationRReLU.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/activations/impl/ActivationRReLU.java
@@ -34,7 +34,7 @@ import org.nd4j.shade.jackson.annotation.JsonIgnoreProperties;
* alpha is drawn from uniform(l,u) during training and is set to l+u/2 during test
* l and u default to 1/8 and 1/3 respectively
*
- *
+ *
* Empirical Evaluation of Rectified Activations in Convolutional Network
*/
@EqualsAndHashCode(callSuper = false)
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/OpContext.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/OpContext.java
index cd74a60a0..e66d52f91 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/OpContext.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/OpContext.java
@@ -128,4 +128,12 @@ public interface OpContext extends AutoCloseable {
* @param reallyInplace
*/
void markInplace(boolean reallyInplace);
+
+ /**
+ * This method allows to enable/disable use of platform helpers within ops. I.e. mkldnn or cuDNN.
+ * PLEASE NOTE: default value is True
+ *
+ * @param reallyAllow
+ */
+ void allowHelpers(boolean reallyAllow);
}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/custom/BitCast.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/custom/BitCast.java
index ee0adfb94..43bff11e6 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/custom/BitCast.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/custom/BitCast.java
@@ -1,25 +1,54 @@
package org.nd4j.linalg.api.ops.custom;
+import lombok.val;
import org.nd4j.autodiff.samediff.SDVariable;
import org.nd4j.autodiff.samediff.SameDiff;
+import org.nd4j.imports.graphmapper.tf.TFGraphMapper;
import org.nd4j.linalg.api.buffer.DataType;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.api.ops.DynamicCustomOp;
+import org.nd4j.linalg.api.shape.options.ArrayOptionsHelper;
import org.nd4j.linalg.factory.Nd4j;
+import org.tensorflow.framework.AttrValue;
+import org.tensorflow.framework.GraphDef;
+import org.tensorflow.framework.NodeDef;
+
+import java.util.Map;
public class BitCast extends DynamicCustomOp {
public BitCast() {}
+ public BitCast(INDArray in, DataType dataType, INDArray out) {
+ this(in, dataType.toInt(), out);
+ }
+
public BitCast(INDArray in, int dataType, INDArray out) {
inputArguments.add(in);
outputArguments.add(out);
iArguments.add(Long.valueOf(dataType));
}
+ public BitCast(INDArray in, DataType dataType) {
+ this(in, dataType.toInt());
+ }
+
+ public BitCast(INDArray in, int dataType) {
+ inputArguments.add(in);
+ iArguments.add(Long.valueOf(dataType));
+ }
+
public BitCast(SameDiff sameDiff, SDVariable in, SDVariable dataType) {
super("", sameDiff, new SDVariable[]{in, dataType});
}
+ @Override
+ public void initFromTensorFlow(NodeDef nodeDef, SameDiff initWith, Map
* Leaky ReLU may avoid zero gradient "dying ReLU" problem by having non-zero
* gradient below 0.
- * See for example http://arxiv.org/abs/1505.00853 for a comparison of
+ * See for example https://arxiv.org/abs/1505.00853 for a comparison of
* ReLU variants.
*
* @author Alex Black
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/MaxOut.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/MaxOut.java
index 05993cd7f..939ed854b 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/MaxOut.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/MaxOut.java
@@ -33,7 +33,7 @@ import java.util.List;
/**
* Max out activation:
- * http://arxiv.org/pdf/1302.4389.pdf
+ * https://arxiv.org/pdf/1302.4389.pdf
*
* @author Adam Gibson
*/
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/strict/ELU.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/strict/ELU.java
index 6923639fd..c4fc245b7 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/strict/ELU.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/transforms/strict/ELU.java
@@ -32,7 +32,7 @@ import java.util.List;
* Introduced in paper:
* Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)
* Djork-Arné Clevert, Thomas Unterthiner, Sepp Hochreiter (2015)
- * http://arxiv.org/abs/1511.07289
+ * https://arxiv.org/abs/1511.07289
*
* @author Alex Black
*/
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/custom/DistributionUniform.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/custom/DistributionUniform.java
index 0744533ba..ecc76a1b2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/custom/DistributionUniform.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/random/custom/DistributionUniform.java
@@ -74,6 +74,7 @@ public class DistributionUniform extends DynamicCustomOp {
AttrValue v = attributesForNode.get("dtype");
dataType = TFGraphMapper.convertType(v.getType());
addIArgument(dataType.toInt());
+ addTArgument(0.0, 1.0); //TF version is hardcoded 0 to 1
}
protected void addArgs() {
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/shape/Shape.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/shape/Shape.java
index 51711b3d2..44298ffa2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/shape/Shape.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/shape/Shape.java
@@ -3607,6 +3607,13 @@ public class Shape {
return ArrayUtil.prodLong(shape);
}
+ public static long lengthOf(int[] shape) {
+ if (shape.length == 0)
+ return 1L;
+ else
+ return ArrayUtil.prodLong(shape);
+ }
+
/**
* Calculate the length of the buffer required to store the given shape with the given strides
*
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/iterator/SamplingDataSetIterator.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/iterator/SamplingDataSetIterator.java
index c33b37565..cc6fba068 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/iterator/SamplingDataSetIterator.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/dataset/api/iterator/SamplingDataSetIterator.java
@@ -28,11 +28,6 @@ import java.util.List;
* @author Adam Gibson
*/
public class SamplingDataSetIterator implements DataSetIterator {
-
- /**
- *
- */
- private static final long serialVersionUID = -2700563801361726914L;
private DataSet sampleFrom;
private int batchSize;
private int totalNumberSamples;
@@ -145,6 +140,4 @@ public class SamplingDataSetIterator implements DataSetIterator {
numTimesSampled++;
return ret;
}
-
-
}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
index 25960a8a8..c95dc5ef2 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/Nd4j.java
@@ -1164,26 +1164,15 @@ public class Nd4j {
* @param type the opType to create
* @return the created buffer
*/
- public static DataBuffer createBuffer(int[] shape, DataType type) {
- long length = ArrayUtil.prodLong(shape);
-
- if (type == DataType.INT)
- return Nd4j.getMemoryManager().getCurrentWorkspace() == null ? DATA_BUFFER_FACTORY_INSTANCE.createInt(length, true) : DATA_BUFFER_FACTORY_INSTANCE.createInt(length, true, Nd4j.getMemoryManager().getCurrentWorkspace());
- else if (type == DataType.LONG)
- return Nd4j.getMemoryManager().getCurrentWorkspace() == null ? DATA_BUFFER_FACTORY_INSTANCE.createLong(length, true) : DATA_BUFFER_FACTORY_INSTANCE.createLong(length, true, Nd4j.getMemoryManager().getCurrentWorkspace());
- else if (type == DataType.HALF)
- return Nd4j.getMemoryManager().getCurrentWorkspace() == null ? DATA_BUFFER_FACTORY_INSTANCE.createHalf(length, true) : DATA_BUFFER_FACTORY_INSTANCE.createHalf(length, true, Nd4j.getMemoryManager().getCurrentWorkspace());
- else if (type == DataType.DOUBLE)
- return Nd4j.getMemoryManager().getCurrentWorkspace() == null ? DATA_BUFFER_FACTORY_INSTANCE.createDouble(length, true) : DATA_BUFFER_FACTORY_INSTANCE.createDouble(length, true, Nd4j.getMemoryManager().getCurrentWorkspace());
- else
- return Nd4j.getMemoryManager().getCurrentWorkspace() == null ? DATA_BUFFER_FACTORY_INSTANCE.createFloat(length, true) : DATA_BUFFER_FACTORY_INSTANCE.createFloat(length, true, Nd4j.getMemoryManager().getCurrentWorkspace());
+ public static DataBuffer createBuffer(@NonNull int[] shape, @NonNull DataType type) {
+ return createBuffer(ArrayUtil.toLongArray(shape), type);
}
/**
* See {@link #createBuffer(int[], DataType)}
*/
- public static DataBuffer createBuffer(long[] shape, DataType type) {
- long length = ArrayUtil.prodLong(shape);
+ public static DataBuffer createBuffer(@NonNull long[] shape, @NonNull DataType type) {
+ long length = Shape.lengthOf(shape);
switch (type) {
case BOOL:
@@ -1229,14 +1218,14 @@ public class Nd4j {
* @return the created buffer.
*/
public static DataBuffer createBufferDetached(int[] shape, DataType type) {
- return createBufferDetachedImpl( ArrayUtil.prodLong(shape), type);
+ return createBufferDetachedImpl( Shape.lengthOf(shape), type);
}
/**
* See {@link #createBufferDetached(int[], DataType)}
*/
public static DataBuffer createBufferDetached(long[] shape, DataType type) {
- return createBufferDetachedImpl( ArrayUtil.prodLong(shape), type);
+ return createBufferDetachedImpl( Shape.lengthOf(shape), type);
}
// used by createBufferDetached(long[] DataType) and createBufferDetached(int[] , DataType)
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/AdaMaxUpdater.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/AdaMaxUpdater.java
index c398dad72..20a908f1e 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/AdaMaxUpdater.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/AdaMaxUpdater.java
@@ -32,7 +32,7 @@ import java.util.Map;
/**
* The AdaMax updater, a variant of Adam.
- * http://arxiv.org/abs/1412.6980
+ * https://arxiv.org/abs/1412.6980
*
* @author Justin Long
*/
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/AdamUpdater.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/AdamUpdater.java
index 8d7709873..e68af09f7 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/AdamUpdater.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/AdamUpdater.java
@@ -30,7 +30,7 @@ import java.util.Map;
/**
* The Adam updater.
- * http://arxiv.org/abs/1412.6980
+ * https://arxiv.org/abs/1412.6980
*
* @author Adam Gibson
*/
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/config/AdaMax.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/config/AdaMax.java
index 00956589a..848bb3408 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/config/AdaMax.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/config/AdaMax.java
@@ -28,7 +28,7 @@ import java.util.Map;
/**
* The AdaMax updater, a variant of Adam.
- * http://arxiv.org/abs/1412.6980
+ * https://arxiv.org/abs/1412.6980
*
* @author Justin Long
*/
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/config/Adam.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/config/Adam.java
index 22ebe06f3..6901af59c 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/config/Adam.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/learning/config/Adam.java
@@ -29,7 +29,7 @@ import java.util.Map;
/**
* The Adam updater.
- * http://arxiv.org/abs/1412.6980
+ * https://arxiv.org/abs/1412.6980
*
* @author Adam Gibson
*/
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/string/NDArrayStrings.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/string/NDArrayStrings.java
index c28f35151..f5c0d5c9e 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/string/NDArrayStrings.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/string/NDArrayStrings.java
@@ -46,6 +46,9 @@ public class NDArrayStrings {
public static final String EMPTY_ARRAY_STR = "[]";
+ private static final String[] OPEN_BRACKETS = new String[]{"", "[", "[[", "[[[", "[[[[", "[[[[[", "[[[[[[", "[[[[[[[", "[[[[[[[["};
+ private static final String[] CLOSE_BRACKETS = new String[]{"", "]", "]]", "]]]", "]]]]", "]]]]]", "]]]]]]", "]]]]]]]", "]]]]]]]]"};
+
/**
* The default number of elements for printing INDArrays (via NDArrayStrings or INDArray.toString)
*/
@@ -190,29 +193,29 @@ public class NDArrayStrings {
private String format(INDArray arr, int offset, boolean summarize) {
int rank = arr.rank();
- if (arr.isScalar()) {
+ if (arr.isScalar() || arr.length() == 1) {
+ int fRank = Math.min(rank, OPEN_BRACKETS.length-1);
if (arr.isR()) {
- //true scalar i.e shape = [] not legacy which is [1,1]
double arrElement = arr.getDouble(0);
if (!dontOverrideFormat && ((Math.abs(arrElement) < this.minToPrintWithoutSwitching && arrElement != 0) || (Math.abs(arrElement) >= this.maxToPrintWithoutSwitching))) {
//switch to scientific notation
String asString = localeIndifferentDecimalFormat(scientificFormat).format(arrElement);
//from E to small e
asString = asString.replace('E', 'e');
- return asString;
+ return OPEN_BRACKETS[fRank] + asString + CLOSE_BRACKETS[fRank];
} else {
- if (arr.getDouble(0) == 0) return "0";
- return decimalFormat.format(arr.getDouble(0));
+ if (arr.getDouble(0) == 0) return OPEN_BRACKETS[fRank] + "0" + CLOSE_BRACKETS[fRank];
+ return OPEN_BRACKETS[fRank] + decimalFormat.format(arr.getDouble(0)) + CLOSE_BRACKETS[fRank];
}
} else if (arr.isZ()) {
long arrElement = arr.getLong(0);
- return String.valueOf(arrElement);
+ return OPEN_BRACKETS[fRank] + arrElement + CLOSE_BRACKETS[fRank];
} else if (arr.isB()) {
long arrElement = arr.getLong(0);
- return arrElement == 0 ? "false" : "true";
+ return OPEN_BRACKETS[fRank] + (arrElement == 0 ? "false" : "true") + CLOSE_BRACKETS[fRank];
} else if (arr.dataType() == DataType.UTF8){
String s = arr.getString(0);
- return "\"" + s.replaceAll("\n","\\n") + "\"";
+ return OPEN_BRACKETS[fRank] + "\"" + s.replaceAll("\n","\\n") + "\"" + CLOSE_BRACKETS[fRank];
} else
throw new ND4JIllegalStateException();
} else if (rank == 1) {
@@ -246,9 +249,10 @@ public class NDArrayStrings {
//hack fix for slice issue with 'f' order
if (arr.ordering() == 'f' && arr.rank() > 2 && arr.size(arr.rank() - 1) == 1) {
sb.append(format(arr.dup('c').slice(i), offset, summarize));
- } else if(arr.rank() <= 1 || arr.length() == 1) {
- sb.append(format(Nd4j.scalar(arr.getDouble(0)),offset,summarize));
}
+// else if(arr.rank() <= 1 || arr.length() == 1) {
+// sb.append(format(Nd4j.scalar(arr.getDouble(0)),offset,summarize));
+// }
else {
sb.append(format(arr.slice(i), offset, summarize));
}
diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/nativeblas/NativeOps.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/nativeblas/NativeOps.java
index 8f621668b..d4a7b8f8b 100644
--- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/nativeblas/NativeOps.java
+++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-native-api/src/main/java/org/nd4j/nativeblas/NativeOps.java
@@ -1123,6 +1123,7 @@ public interface NativeOps {
void setGraphContextTArguments(OpaqueContext ptr, DoublePointer arguments, int numberOfArguments);
void setGraphContextIArguments(OpaqueContext ptr, LongPointer arguments, int numberOfArguments);
void setGraphContextBArguments(OpaqueContext ptr, BooleanPointer arguments, int numberOfArguments);
+ void ctxAllowHelpers(OpaqueContext ptr, boolean reallyAllow);
void deleteGraphContext(OpaqueContext ptr);
OpaqueRandomGenerator createRandomGenerator(long rootSeed, long nodeSeed);
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
index 8fe744b38..20f2b5f22 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaExecutioner.java
@@ -2226,7 +2226,7 @@ public class CudaExecutioner extends DefaultOpExecutioner {
cnt = 0;
for (val t: op.tArgs())
- tArgs.put(cnt++, (float) t);
+ tArgs.put(cnt++, t);
OpaqueShapeList ptrptr = nativeOps.calculateOutputShapes2(null, hash, inputBuffers, inputShapes, op.inputArguments().length, tArgs, op.tArgs().length, iArgs, op.iArgs().length, bArgs, op.numBArguments());
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaOpContext.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaOpContext.java
index 32f1b0a10..b75f688fe 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaOpContext.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/linalg/jcublas/ops/executioner/CudaOpContext.java
@@ -136,4 +136,9 @@ public class CudaOpContext extends BaseOpContext implements OpContext {
public void markInplace(boolean reallyInplace) {
nativeOps.markGraphContextInplace(context, reallyInplace);
}
+
+ @Override
+ public void allowHelpers(boolean reallyAllow) {
+ nativeOps.ctxAllowHelpers(context, reallyAllow);
+ }
}
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java
index fecb64012..e8b5e15c9 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-cuda/src/main/java/org/nd4j/nativeblas/Nd4jCuda.java
@@ -1,4 +1,4 @@
-// Targeted by JavaCPP version 1.5.1-1: DO NOT EDIT THIS FILE
+// Targeted by JavaCPP version 1.5.2: DO NOT EDIT THIS FILE
package org.nd4j.nativeblas;
@@ -575,6 +575,8 @@ public class Nd4jCuda extends org.nd4j.nativeblas.Nd4jCudaHelper {
public native void setDebug(@Cast("bool") boolean reallyDebug);
public native void setProfiling(@Cast("bool") boolean reallyProfile);
public native void setLeaksDetector(@Cast("bool") boolean reallyDetect);
+ public native @Cast("bool") boolean helpersAllowed();
+ public native void allowHelpers(@Cast("bool") boolean reallyAllow);
public native int tadThreshold();
public native void setTadThreshold(int threshold);
@@ -585,6 +587,13 @@ public class Nd4jCuda extends org.nd4j.nativeblas.Nd4jCudaHelper {
public native int maxThreads();
public native void setMaxThreads(int max);
+ public native int maxMasterThreads();
+ public native void setMaxMasterThreads(int max);
+
+ public native void setMaxPrimaryMemory(@Cast("uint64_t") long maxBytes);
+ public native void setMaxSpecialyMemory(@Cast("uint64_t") long maxBytes);
+ public native void setMaxDeviceMemory(@Cast("uint64_t") long maxBytes);
+
public native @Cast("bool") boolean isUseMKLDNN();
public native void setUseMKLDNN(@Cast("bool") boolean useMKLDNN);
@@ -3087,6 +3096,7 @@ public native void deleteShapeBuffer(OpaqueConstantDataBuffer ptr);
public native OpaqueContext createGraphContext(int nodeId);
public native OpaqueRandomGenerator getGraphContextRandomGenerator(OpaqueContext ptr);
+public native void ctxAllowHelpers(OpaqueContext ptr, @Cast("bool") boolean reallyAllow);
public native void markGraphContextInplace(OpaqueContext ptr, @Cast("bool") boolean reallyInplace);
public native void setGraphContextCudaContext(OpaqueContext ptr, Pointer stream, Pointer reductionPointer, Pointer allocationPointer);
public native void setGraphContextInputArray(OpaqueContext ptr, int index, Pointer buffer, Pointer shapeInfo, Pointer specialBuffer, Pointer specialShapeInfo);
@@ -5454,6 +5464,10 @@ NDArray& NDArray::operator()(const Nd4jLong* idx) {
+
+
+
+
@@ -6740,7 +6754,20 @@ NDArray& NDArray::operator()(const Nd4jLong* idx) {
public native void setBArguments(@Cast("bool*") BooleanPointer arguments, int numberOfArguments);
public native void setBArguments(@Cast("bool*") boolean[] arguments, int numberOfArguments);
+ public native void setTArguments(@StdVector DoublePointer tArgs);
+ public native void setTArguments(@StdVector DoubleBuffer tArgs);
+ public native void setTArguments(@StdVector double[] tArgs);
+ public native void setIArguments(@Cast("Nd4jLong*") @StdVector LongPointer tArgs);
+ public native void setIArguments(@Cast("Nd4jLong*") @StdVector LongBuffer tArgs);
+ public native void setIArguments(@Cast("Nd4jLong*") @StdVector long[] tArgs);
+ public native void setBArguments(@Cast("bool*") @StdVector BooleanPointer tArgs);
+ public native void setBArguments(@Cast("bool*") @StdVector boolean[] tArgs);
+
public native void setCudaContext(@Cast("Nd4jPointer") Pointer cudaStream, @Cast("Nd4jPointer") Pointer reductionPointer, @Cast("Nd4jPointer") Pointer allocationPointer);
+
+
+ public native void allowHelpers(@Cast("bool") boolean reallyAllow);
+ public native @Cast("bool") boolean helpersAllowed();
}
diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/pom.xml b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/pom.xml
index c6017e3a7..33e54bd4a 100644
--- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/pom.xml
+++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/pom.xml
@@ -34,13 +34,6 @@