From 185e7f554f7a1dfa1e2d6757f49396be0abae24d Mon Sep 17 00:00:00 2001 From: agibsonccc Date: Sat, 13 Mar 2021 11:00:11 +0900 Subject: [PATCH] Fix compilation issues from codegen --- contrib/codegen-tools/codegen/generate.sh | 2 +- contrib/codegen-tools/codegen/pom.xml | 2 +- .../ops/org/nd4j/codegen/ops/SDBaseOps.kt | 18 +++++- .../keras/e2e/KerasModelEndToEndTest.java | 1 - deeplearning4j/pom.xml | 1 - .../nd4j/autodiff/samediff/ops/SDBaseOps.java | 26 +++++++++ .../java/org/nd4j/ir/TensorNamespace.java | 12 ---- .../nd4j/linalg/api/ops/impl/shape/Split.java | 5 ++ .../org/nd4j/linalg/factory/ops/NDBase.java | 12 ++++ .../java/org/nd4j/nativeblas/Nd4jCpu.java | 55 +++++++++++++++++++ .../opvalidation/TransformOpValidation.java | 2 +- .../processing/GroupConvPreProcessingRule.kt | 2 +- 12 files changed, 117 insertions(+), 21 deletions(-) diff --git a/contrib/codegen-tools/codegen/generate.sh b/contrib/codegen-tools/codegen/generate.sh index a398af633..e254ba92c 100644 --- a/contrib/codegen-tools/codegen/generate.sh +++ b/contrib/codegen-tools/codegen/generate.sh @@ -7,5 +7,5 @@ if test "$#" -eq 0; then echo "Usage example 2 (all namespaces): ./generate.sh all" else mvn clean package -DskipTests - java -cp target/codegen-1.0.0-SNAPSHOT-shaded.jar org.nd4j.codegen.cli.CLI -dir ../../ -namespaces "$@" + java -cp target/codegen-1.0.0-SNAPSHOT-shaded.jar org.nd4j.codegen.cli.CLI -dir ../../../ -namespaces "$@" fi \ No newline at end of file diff --git a/contrib/codegen-tools/codegen/pom.xml b/contrib/codegen-tools/codegen/pom.xml index 5272b37d9..cbd00a825 100644 --- a/contrib/codegen-tools/codegen/pom.xml +++ b/contrib/codegen-tools/codegen/pom.xml @@ -102,7 +102,7 @@ org.nd4j nd4j-api - 1.0.0-SNAPSHOT + ${project.version} diff --git a/contrib/codegen-tools/codegen/src/main/ops/org/nd4j/codegen/ops/SDBaseOps.kt b/contrib/codegen-tools/codegen/src/main/ops/org/nd4j/codegen/ops/SDBaseOps.kt index 28a4d6887..862b951e3 100644 --- a/contrib/codegen-tools/codegen/src/main/ops/org/nd4j/codegen/ops/SDBaseOps.kt +++ b/contrib/codegen-tools/codegen/src/main/ops/org/nd4j/codegen/ops/SDBaseOps.kt @@ -29,7 +29,6 @@ import org.nd4j.codegen.api.doc.DocScope import org.nd4j.codegen.dsl.* import org.nd4j.codegen.api.DataType.* import org.nd4j.codegen.mixins.* -import org.nd4j.linalg.api.buffer.DataType import java.lang.Boolean.FALSE fun SDBaseOps() = Namespace("BaseOps"){ @@ -594,7 +593,7 @@ fun SDBaseOps() = Namespace("BaseOps"){ legacy = true Input(NUMERIC, "x") { description = "Input variable" } Arg(BOOL, "keepDims") { description = "If true: keep the dimensions that are reduced on (as size 1). False: remove the reduction dimensions" - ; defaultValue=FALSE } + ; defaultValue=FALSE } Arg(INT, "dimensions") { count = AtLeast(0); description = "Dimensions to reduce over. If dimensions are not specified, full array reduction is performed" } Output(NUMERIC, "output"){ description = "Reduced array of rank (input rank - num dimensions)" } Doc(Language.ANY, DocScope.ALL){ @@ -773,6 +772,19 @@ fun SDBaseOps() = Namespace("BaseOps"){ useMixin(keepDimsDoc) } + Op("split") { + javaPackage = "org.nd4j.linalg.api.ops.impl.shape" + javaOpClass = "Split" + Input(NUMERIC,"input") {description = "Input to split"} + Arg(INT, "numSplit") { description = "Number of splits" } + Arg(INT, "splitDim") { description = "The dimension to split on" } + Doc(Language.ANY, DocScope.ALL){ + """ + Split a value in to a list of ndarrays. + """.trimIndent() + } + } + Op("oneHot") { javaPackage = "org.nd4j.linalg.api.ops.impl.shape" Input(NUMERIC, "indices") { description = "Indices - value 0 to depth-1" } @@ -780,7 +792,7 @@ fun SDBaseOps() = Namespace("BaseOps"){ Arg(INT, "axis") { description = "" } Arg(NUMERIC, "on") { description = "" } Arg(NUMERIC, "off") { description = "" } - Arg(DATA_TYPE, "dataType") { description = "Output data type"; defaultValue = DataType.FLOAT } + Arg(DATA_TYPE, "dataType") { description = "Output data type"; defaultValue = org.nd4j.linalg.api.buffer.DataType.FLOAT } Output(NUMERIC, "output"){ description = "Output variable" } Doc(Language.ANY, DocScope.ALL){ diff --git a/deeplearning4j/deeplearning4j-modelimport/src/test/java/org/deeplearning4j/nn/modelimport/keras/e2e/KerasModelEndToEndTest.java b/deeplearning4j/deeplearning4j-modelimport/src/test/java/org/deeplearning4j/nn/modelimport/keras/e2e/KerasModelEndToEndTest.java index f6c33919c..ab83c7f7f 100644 --- a/deeplearning4j/deeplearning4j-modelimport/src/test/java/org/deeplearning4j/nn/modelimport/keras/e2e/KerasModelEndToEndTest.java +++ b/deeplearning4j/deeplearning4j-modelimport/src/test/java/org/deeplearning4j/nn/modelimport/keras/e2e/KerasModelEndToEndTest.java @@ -77,7 +77,6 @@ import static org.junit.Assert.*; * @author dave@skymind.io, Max Pumperla */ @Slf4j -@Ignore public class KerasModelEndToEndTest extends BaseDL4JTest { private static final String GROUP_ATTR_INPUTS = "inputs"; private static final String GROUP_ATTR_OUTPUTS = "outputs"; diff --git a/deeplearning4j/pom.xml b/deeplearning4j/pom.xml index 7b112e23f..a9687116e 100644 --- a/deeplearning4j/pom.xml +++ b/deeplearning4j/pom.xml @@ -197,7 +197,6 @@ deeplearning4j-modelimport deeplearning4j-modelexport-solr deeplearning4j-zoo - deeplearning4j-nearestneighbors-parent diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDBaseOps.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDBaseOps.java index 8190c4849..2838f66f3 100644 --- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDBaseOps.java +++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/autodiff/samediff/ops/SDBaseOps.java @@ -3799,6 +3799,32 @@ public class SDBaseOps { return sd.updateVariableNameAndReference(out, name); } + /** + * Split a value in to a list of ndarrays.
+ * + * @param input Input to split (NUMERIC type) + * @param numSplit Number of splits + * @param splitDim The dimension to split on + */ + public SDVariable[] split(SDVariable input, int numSplit, int splitDim) { + SDValidation.validateNumerical("split", "input", input); + return new org.nd4j.linalg.api.ops.impl.shape.Split(sd,input, numSplit, splitDim).outputVariables(); + } + + /** + * Split a value in to a list of ndarrays.
+ * + * @param names names May be null. Arrays of names for the output variables. + * @param input Input to split (NUMERIC type) + * @param numSplit Number of splits + * @param splitDim The dimension to split on + */ + public SDVariable[] split(String[] names, SDVariable input, int numSplit, int splitDim) { + SDValidation.validateNumerical("split", "input", input); + SDVariable[] out = new org.nd4j.linalg.api.ops.impl.shape.Split(sd,input, numSplit, splitDim).outputVariables(); + return sd.updateVariableNamesAndReferences(out, names); + } + /** * Squared L2 norm: see norm2(String, SDVariable, boolean, int...)
* diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/ir/TensorNamespace.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/ir/TensorNamespace.java index a0899ddbe..434bda3a8 100644 --- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/ir/TensorNamespace.java +++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/ir/TensorNamespace.java @@ -5411,14 +5411,12 @@ public final class TensorNamespace { * Serializations can either use one of the fields above, or use this * raw bytes field. The only exception is the string case, where one is * required to store the content in the repeated bytes string_data field. - * * When this raw_data field is used to store tensor value, elements MUST * be stored in as fixed-width, little-endian order. * Floating-point data types MUST be stored in IEEE 754 format. * Complex64 elements must be written as two consecutive FLOAT values, real component first. * Complex128 elements must be written as two consecutive DOUBLE values, real component first. * Boolean type MUST be written one byte per tensor element (00000001 for true, 00000000 for false). - * * Note: the advantage of specific field rather than the raw_data field is * that in some cases (e.g. int data), protobuf does a better packing via * variable length storage, and may lead to smaller binary footprint. @@ -5657,7 +5655,6 @@ public final class TensorNamespace { /** * 
    * Tensors
-   *
    * A serialized tensor value.
    *
* @@ -7013,14 +7010,12 @@ public final class TensorNamespace { * Serializations can either use one of the fields above, or use this * raw bytes field. The only exception is the string case, where one is * required to store the content in the repeated bytes string_data field. - * * When this raw_data field is used to store tensor value, elements MUST * be stored in as fixed-width, little-endian order. * Floating-point data types MUST be stored in IEEE 754 format. * Complex64 elements must be written as two consecutive FLOAT values, real component first. * Complex128 elements must be written as two consecutive DOUBLE values, real component first. * Boolean type MUST be written one byte per tensor element (00000001 for true, 00000000 for false). - * * Note: the advantage of specific field rather than the raw_data field is * that in some cases (e.g. int data), protobuf does a better packing via * variable length storage, and may lead to smaller binary footprint. @@ -7771,7 +7766,6 @@ public final class TensorNamespace { /** * 
      * Tensors
-     *
      * A serialized tensor value.
      *
* @@ -9086,14 +9080,12 @@ public final class TensorNamespace { * Serializations can either use one of the fields above, or use this * raw bytes field. The only exception is the string case, where one is * required to store the content in the repeated bytes string_data field. - * * When this raw_data field is used to store tensor value, elements MUST * be stored in as fixed-width, little-endian order. * Floating-point data types MUST be stored in IEEE 754 format. * Complex64 elements must be written as two consecutive FLOAT values, real component first. * Complex128 elements must be written as two consecutive DOUBLE values, real component first. * Boolean type MUST be written one byte per tensor element (00000001 for true, 00000000 for false). - * * Note: the advantage of specific field rather than the raw_data field is * that in some cases (e.g. int data), protobuf does a better packing via * variable length storage, and may lead to smaller binary footprint. @@ -9110,14 +9102,12 @@ public final class TensorNamespace { * Serializations can either use one of the fields above, or use this * raw bytes field. The only exception is the string case, where one is * required to store the content in the repeated bytes string_data field. - * * When this raw_data field is used to store tensor value, elements MUST * be stored in as fixed-width, little-endian order. * Floating-point data types MUST be stored in IEEE 754 format. * Complex64 elements must be written as two consecutive FLOAT values, real component first. * Complex128 elements must be written as two consecutive DOUBLE values, real component first. * Boolean type MUST be written one byte per tensor element (00000001 for true, 00000000 for false). - * * Note: the advantage of specific field rather than the raw_data field is * that in some cases (e.g. int data), protobuf does a better packing via * variable length storage, and may lead to smaller binary footprint. @@ -9140,14 +9130,12 @@ public final class TensorNamespace { * Serializations can either use one of the fields above, or use this * raw bytes field. The only exception is the string case, where one is * required to store the content in the repeated bytes string_data field. - * * When this raw_data field is used to store tensor value, elements MUST * be stored in as fixed-width, little-endian order. * Floating-point data types MUST be stored in IEEE 754 format. * Complex64 elements must be written as two consecutive FLOAT values, real component first. * Complex128 elements must be written as two consecutive DOUBLE values, real component first. * Boolean type MUST be written one byte per tensor element (00000001 for true, 00000000 for false). - * * Note: the advantage of specific field rather than the raw_data field is * that in some cases (e.g. int data), protobuf does a better packing via * variable length storage, and may lead to smaller binary footprint. diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Split.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Split.java index f6856c0e8..8325c074d 100644 --- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Split.java +++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/impl/shape/Split.java @@ -58,6 +58,11 @@ public class Split extends DynamicCustomOp { super(null, new INDArray[]{in}, wrapOrNull(out), null, (List)null); } + public Split(INDArray input, int numSplit, int splitDim) { + super(null,input,null,Collections.emptyList(),new int[0]); + addIArgument(numSplit,splitDim); + } + @Override public String opName() { diff --git a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/ops/NDBase.java b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/ops/NDBase.java index 1b2718e2e..039e9184a 100644 --- a/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/ops/NDBase.java +++ b/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/factory/ops/NDBase.java @@ -1800,6 +1800,18 @@ public class NDBase { return Nd4j.exec(new org.nd4j.linalg.api.ops.impl.shape.Slice(input, begin, size))[0]; } + /** + * Split a value in to a list of ndarrays.
+ * + * @param input Input to split (NUMERIC type) + * @param numSplit Number of splits + * @param splitDim The dimension to split on + */ + public INDArray[] split(INDArray input, int numSplit, int splitDim) { + NDValidation.validateNumerical("split", "input", input); + return Nd4j.exec(new org.nd4j.linalg.api.ops.impl.shape.Split(input, numSplit, splitDim)); + } + /** * Squared L2 norm: see norm2(String, SDVariable, boolean, int...)
* diff --git a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java index 1dff1b893..de65b60f9 100644 --- a/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java +++ b/nd4j/nd4j-backends/nd4j-backend-impls/nd4j-native/src/main/java/org/nd4j/nativeblas/Nd4jCpu.java @@ -25870,6 +25870,61 @@ public static final double TAD_THRESHOLD = TAD_THRESHOLD(); } // #endif + /** + * Implementation of CTC loss function + * + * Input arrays: + * 0: labels - labels NDArray {BATCH_LEN, MAX_TARGET_LEN}, type integer + * 1: logits - logits NDArray {BATCH_LEN, FRAME_LEN, CLASS_LEN }. log softmax of rnn output. It should include a blank label as well, type float + * 2: targetLabelLengths - Length of label sequence in labels NDArray {BATCH_LEN}, type integer + * 3: logitsLengths - Length of input sequence in logits NDArray {BATCH_LEN}, type integer + * + * + * Input integer arguments: + * 0: blank index - index of the blank label in logits + * + * Output array: + * 0: loss values, type float. NDArray {BATCH_LEN} negative log probabilities of loss + */ +// #if NOT_EXCLUDED(OP_ctc_loss) + @Namespace("sd::ops") public static class ctc_loss extends DeclarableCustomOp { + static { Loader.load(); } + /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */ + public ctc_loss(Pointer p) { super(p); } + /** Native array allocator. Access with {@link Pointer#position(long)}. */ + public ctc_loss(long size) { super((Pointer)null); allocateArray(size); } + private native void allocateArray(long size); + @Override public ctc_loss position(long position) { + return (ctc_loss)super.position(position); + } + @Override public ctc_loss getPointer(long i) { + return new ctc_loss((Pointer)this).position(position + i); + } + + public ctc_loss() { super((Pointer)null); allocate(); } + private native void allocate(); + public native ShapeList calculateOutputShape(ShapeList inputShape, @ByRef Context block); + } + @Namespace("sd::ops") public static class ctc_loss_grad extends DeclarableCustomOp { + static { Loader.load(); } + /** Pointer cast constructor. Invokes {@link Pointer#Pointer(Pointer)}. */ + public ctc_loss_grad(Pointer p) { super(p); } + /** Native array allocator. Access with {@link Pointer#position(long)}. */ + public ctc_loss_grad(long size) { super((Pointer)null); allocateArray(size); } + private native void allocateArray(long size); + @Override public ctc_loss_grad position(long position) { + return (ctc_loss_grad)super.position(position); + } + @Override public ctc_loss_grad getPointer(long i) { + return new ctc_loss_grad((Pointer)this).position(position + i); + } + + public ctc_loss_grad() { super((Pointer)null); allocate(); } + private native void allocate(); + public native ShapeList calculateOutputShape(ShapeList inputShape, @ByRef Context block); + } +// #endif + diff --git a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java index fc63e5621..fcb78a9ae 100644 --- a/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java +++ b/nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/autodiff/opvalidation/TransformOpValidation.java @@ -2123,7 +2123,7 @@ public class TransformOpValidation extends BaseOpValidation { //TODO: Methods failed ResizeLanczos5, ResizeMitchelcubic, ResizeArea for (ImageResizeMethod method : ImageResizeMethod.values()) { - if (method==ImageResizeMethod.ResizeLanczos5 || method==ImageResizeMethod.ResizeArea || method==ImageResizeMethod.ResizeMitchellcubic) + if (method==ImageResizeMethod.ResizeLanczos5 || method==ImageResizeMethod.ResizeArea || method == ImageResizeMethod.ResizeMitchelcubic) {continue;} log.info("Trying {}", method); diff --git a/nd4j/samediff-import/samediff-import-onnx/src/test/kotlin/org/nd4j/samediff/frameworkimport/onnx/processing/GroupConvPreProcessingRule.kt b/nd4j/samediff-import/samediff-import-onnx/src/test/kotlin/org/nd4j/samediff/frameworkimport/onnx/processing/GroupConvPreProcessingRule.kt index a087fb221..3732ca7d2 100644 --- a/nd4j/samediff-import/samediff-import-onnx/src/test/kotlin/org/nd4j/samediff/frameworkimport/onnx/processing/GroupConvPreProcessingRule.kt +++ b/nd4j/samediff-import/samediff-import-onnx/src/test/kotlin/org/nd4j/samediff/frameworkimport/onnx/processing/GroupConvPreProcessingRule.kt @@ -69,7 +69,7 @@ class GroupConvPreProcessingRule: PreImportHook { val listOfFunctions = ArrayList() val weights = sd.getVariable(op.inputsToOp[1]) //for onnx, this is the number of ops - val split = sd.split(op.name + "_split",weights,numSizeSplits.toInt(),1) + val split = sd.split(listOf(op.name + "_split").toTypedArray(),weights,numSizeSplits.toInt(),1) val resultMap = HashMap>() /** * NOTE: Need to look in to how to wire up inputs and outputs properly.