Add new clion rules, fix batch norml

libnd4j/blas/CMakeLists.txt

@@ -373,7 +373,11 @@ elseif(SD_CPU)
     foreach (_variableName ${_variableNames})
         message(STATUS "${_variableName}=${${_variableName}}")
     endforeach()
+
+    #This breaks the build. Normally you want to run tests anyways.
+    if(NOT "$ENV{CLION_IDE}")
         target_link_libraries(${SD_LIBRARY_NAME} ${MKLDNN}  ${MKLDNN_LIBRARIES} ${ARMCOMPUTE_LIBRARIES} ${OPENBLAS_LIBRARIES} ${BLAS_LIBRARIES} ${CPU_FEATURES})
+    endif()
 
     if ("${SD_ALL_OPS}" AND "${SD_BUILD_MINIFIER}")
         message(STATUS "Building minifier...")

libnd4j/include/ops/declarable/generic/nn/fusedBatchNorm.cpp

@@ -61,6 +61,9 @@ CUSTOM_OP_IMPL(fused_batch_norm, 3, 3, false, 0, 2) {
                 iW = x->sizeAt(2);
             }
 
+            auto xCast = x->cast(sd::DataType::FLOAT32);
+
+
             REQUIRE_TRUE(scale->rankOf() == 1  && scale->sizeAt(0)  == iD, 0, "CUSTOM_OP fused_batch_norm: wrong shape of input scale array, expected is [%i], but got %s instead", iD, ShapeUtils::shapeAsString(scale).c_str());
             REQUIRE_TRUE(offset->rankOf() == 1 && offset->sizeAt(0) == iD, 0, "CUSTOM_OP fused_batch_norm: wrong shape of input offset array, expected is [%i], but got %s instead", iD, ShapeUtils::shapeAsString(offset).c_str());
 
@@ -74,36 +77,38 @@ CUSTOM_OP_IMPL(fused_batch_norm, 3, 3, false, 0, 2) {
             else {
                 //REQUIRE_TRUE(block.width() == 3, 0, "CUSTOM_OP fused_batch_norm: when isTraining=true then number of input arrays must be equal to 3, but got %i instead !", block.width());
                 std::vector<Nd4jLong> shape = {iD};
-        mean = NDArrayFactory::create_(scale->ordering(), shape, scale->dataType(), block.launchContext());
+                mean = NDArrayFactory::create_(scale->ordering(), shape, sd::DataType::FLOAT32, block.launchContext());
-        variance = NDArrayFactory::create_(scale->ordering(), shape, scale->dataType(), block.launchContext());
+                variance = NDArrayFactory::create_(scale->ordering(), shape, sd::DataType::FLOAT32, block.launchContext());
             }
 
-    // FIXME: double?
+
-    double epsilon;
+            float epsilon;
-    if(block.getTArguments()->size() > 0)
+            if(block.getTArguments()->size() > 0) {
-        epsilon = T_ARG(0) > 1.001e-5 ? T_ARG(0) : 1.001e-5;
+                epsilon = (float) (T_ARG(0) > 1.001e-5 ? T_ARG(0) : 1.001e-5);
-    else
+            }
-        epsilon = 0.001;
+            else {
+                epsilon = 0.001f;
+            }
 
             const int restSize = x->lengthOf() / iD;
-    auto xAffected = NDArrayFactory::create(x->ordering(), {restSize, iD}, mean->dataType(), block.launchContext());
+
-    xAffected.assign(x);
+            auto xAffected = NDArrayFactory::create(x->ordering(), {restSize, iD}, sd::DataType::FLOAT32, block.launchContext());
+            xAffected.assign(xCast);
 
             const int restSizeMinusOne = (restSize > 1) ? (restSize - 1) : 1;
-    // FIXME: float?
+            const float restSizeInv = 1.0f / restSize;
-    const double restSizeInv = 1.0 / restSize;
+            const float restSizeAdjust = (float)restSize / restSizeMinusOne;
-    const double restSizeAdjust = (double)restSize / restSizeMinusOne;
 
             if(isTraining) {
                 auto sum = xAffected.reduceAlongDimension(reduce::Sum, {0});
                 sum *=  restSizeInv;
                 mean->assign(sum);
                 *batchMean = *mean;
-        //delete sum;
             }
             else
                 *batchMean = 0.;
 
+            auto xCentered = xAffected - *mean;
             xAffected -= *mean;
 
             if(isTraining) {
@@ -112,13 +117,17 @@ CUSTOM_OP_IMPL(fused_batch_norm, 3, 3, false, 0, 2) {
                 auto sum = xAffected.reduceAlongDimension(reduce::Sum, {0});
                 sum *=  restSizeInv;
                 variance->assign(sum);
-        *batchVar  = (*variance) * restSizeAdjust;
+                auto varOutput   = (*variance) * restSizeAdjust;
-        //delete sum;
+                batchVar->assign(varOutput);
             }
             else
                 *batchVar  = 0.;
-    xAffected *= (*variance + epsilon).transform(transform::RSqrt) * (*scale) + (*offset);
+
-    y->assign( xAffected );
+            auto scaledVariance =  ((*variance + epsilon).transform(transform::RSqrt) * (*scale)).cast(xAffected.dataType());
+            auto xScaled1 = xCentered * scaledVariance;
+            auto xShifted1 = xScaled1 + *offset;
+
+            y->assign(xShifted1);
 
             if(isTraining) {
                 delete mean;
@@ -148,9 +157,6 @@ DECLARE_SHAPE_FN(fused_batch_norm) {
             return SHAPELIST(CONSTANT(outShapeInfo), CONSTANT(batchMeanShapeInfo), CONSTANT(batchVarShapeInfo));
         }
         
-
-
-
     }
 }
 

nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/custom/FusedBatchNorm.java

@@ -90,6 +90,9 @@ public class FusedBatchNorm extends DynamicCustomOp {
     public List<DataType> calculateOutputDataTypes(List<DataType> inputDataTypes) {
         int n = args().length;
         Preconditions.checkState(inputDataTypes != null && inputDataTypes.size() == n, "Expected %s input data types for %s, got %s", n, getClass(), inputDataTypes);
+        if(!dArguments.isEmpty()) {
+            return Arrays.asList(dArguments.get(0),dArguments.get(0),dArguments.get(0));
+        }
         return Arrays.asList(outputDataType == null ? DataType.FLOAT : outputDataType,
                 outputDataType == null ? DataType.FLOAT : outputDataType,
                 outputDataType == null ? DataType.FLOAT : outputDataType);

nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllSameDiff.java

@@ -69,10 +69,8 @@ public class TFGraphTestAllSameDiff {   //Note: Can't extend BaseNd4jTest here a
      * the status of the test failing. No tests will run.
      */
     public final static List<String> EXECUTE_ONLY_MODELS = Arrays.asList(
-               "max_pool_with_argmax/int32_int64_padding_SAME",
+                  "fused_batch_norm/float32_nhwc"
-            //      "fused_batch_norm/float32_nhwc",
+           // , "fused_batch_norm/float16_nhwc"
-                 "max_pool_with_argmax/int64_int64_padding_SAME"
-            //      "fused_batch_norm/float16_nhwc",
 
     );
 
@@ -86,9 +84,6 @@ public class TFGraphTestAllSameDiff {   //Note: Can't extend BaseNd4jTest here a
             // Still failing 2020/04/27 java.lang.IllegalStateException: Could not find class for TF Ops: TruncateMod
             "truncatemod/.*",
 
-            //Still failing as of 2019/09/11 - https://github.com/deeplearning4j/deeplearning4j/issues/6464 - not sure if related to: https://github.com/deeplearning4j/deeplearning4j/issues/6447
-            "cnn2d_nn/nhwc_b1_k12_s12_d12_SAME",
-
             //2019/09/11 - No tensorflow op found for SparseTensorDenseAdd
             // 2020/04/27 java.lang.IllegalStateException: Could not find class for TF Ops: SparseTensorDenseAdd
             "confusion/.*",

nd4j/samediff-import/samediff-import-tensorflow/src/main/kotlin/org/nd4j/samediff/frameworkimport/tensorflow/definitions/TensorflowOpDeclarations.kt

@@ -958,7 +958,7 @@ val fusedBatchnormV1 = TensorflowMappingProcess(
                 "offset" to "offset","mean" to "mean","variance" to "variance"))),
         inputFrameworkOpName = "FusedBatchNorm",
         opMappingRegistry = tensorflowOpRegistry,
-        attributeMappingRules = listOf(valueMapping(mutableMapOf("epsilon" to "epsilon")),
+        attributeMappingRules = listOf(valueMapping(mutableMapOf("epsilon" to "epsilon","dtype" to "T")),
                 invertBooleanNumber(mutableMapOf("isTraining" to "is_training")),
                 stringEqualsRule(outputAttribute = "dataFormat",inputFrameworkAttributeName = "data_format",valueToTest = "NCHW",argumentIndex = 0))
 )
@@ -971,7 +971,7 @@ val fusedBatchnormV2 = TensorflowMappingProcess(
                 "offset" to "offset","mean" to "mean","variance" to "variance"))),
         inputFrameworkOpName = "FusedBatchNormV2",
         opMappingRegistry = tensorflowOpRegistry,
-        attributeMappingRules = listOf(valueMapping(mutableMapOf("epsilon" to "epsilon")),
+        attributeMappingRules = listOf(valueMapping(mutableMapOf("epsilon" to "epsilon","dtype" to "T")),
                 invertBooleanNumber(mutableMapOf("isTraining" to "is_training")),
                 stringEqualsRule(outputAttribute = "dataFormat",inputFrameworkAttributeName = "data_format",valueToTest = "NCHW",argumentIndex = 0))
 )
@@ -983,7 +983,7 @@ val fusedBatchnormV3 = TensorflowMappingProcess(
                 "offset" to "offset","mean" to "mean","variance" to "variance"))),
         inputFrameworkOpName = "FusedBatchNormV3",
         opMappingRegistry = tensorflowOpRegistry,
-        attributeMappingRules = listOf(valueMapping(mutableMapOf("epsilon" to "epsilon")),
+        attributeMappingRules = listOf(valueMapping(mutableMapOf("epsilon" to "epsilon","dtype" to "T")),
                 invertBooleanNumber(mutableMapOf("isTraining" to "is_training")),
                 stringEqualsRule(outputAttribute = "dataFormat",inputFrameworkAttributeName = "data_format",valueToTest = "NCHW",argumentIndex = 0))
 )

nd4j/samediff-import/samediff-import-tensorflow/src/main/resources/tensorflow-mapping-ruleset.pbtxt

@@ -8367,10 +8367,16 @@ mappings {
     functionName: "valuemapping"
     inputFloatName: "epsilon"
     outputDoubleName: "epsilon"
+    inputDataTypeName: "T"
+    outputDataTypeName: "dtype"
     inputToOutput {
       key: "epsilon"
       value: "epsilon"
     }
+    inputToOutput {
+      key: "dtype"
+      value: "T"
+    }
     ruleType: "attribute"
     inputFrameworkOpName: "FusedBatchNorm"
   }
@@ -12480,10 +12486,16 @@ mappings {
     functionName: "valuemapping"
     inputFloatName: "epsilon"
     outputDoubleName: "epsilon"
+    inputDataTypeName: "T"
+    outputDataTypeName: "dtype"
     inputToOutput {
       key: "epsilon"
       value: "epsilon"
     }
+    inputToOutput {
+      key: "dtype"
+      value: "T"
+    }
     ruleType: "attribute"
     inputFrameworkOpName: "FusedBatchNormV3"
   }
@@ -13056,10 +13068,16 @@ mappings {
     functionName: "valuemapping"
     inputFloatName: "epsilon"
     outputDoubleName: "epsilon"
+    inputDataTypeName: "T"
+    outputDataTypeName: "dtype"
     inputToOutput {
       key: "epsilon"
       value: "epsilon"
     }
+    inputToOutput {
+      key: "dtype"
+      value: "T"
+    }
     ruleType: "attribute"
     inputFrameworkOpName: "FusedBatchNormV2"
   }

nd4j/samediff-import/samediff-import-tensorflow/src/test/kotlin/org/nd4j/samediff/frameworkimport/tensorflow/TestTensorflowIR.kt

@@ -90,7 +90,9 @@ class TestTensorflowIR {
         //val inputMap = mapOf("image" to Nd4j.ones(1,128,128,4))
         val inputMap = emptyMap<String,INDArray>()
         val tensorflowIRGraph = TensorflowIRGraph(textGraph,tensorflowOps,tfImporter.registry)
-        val outputList = tensorflowIRGraph.nodeList().map { input -> input.nodeName() }.toSet()
+        val outputList = tensorflowIRGraph.nodeList().map { input -> input.nodeName() }.toMutableSet()
+        outputList.add("FusedBatchNormV3:1")
+        outputList.add("FusedBatchNormV3:2")
         val tfGraphRunner = TensorflowIRGraphRunner(tensorflowIRGraph, inputMap.keys.toList(), outputList.toList())
         val importedGraph = TFGraphMapper.importGraph(textGraph)
         val graph = tfImporter.importFromGraph(textGraph,inputMap)
@@ -104,7 +106,7 @@ class TestTensorflowIR {
         val names = tensorflowIRGraph.nodeList().map { input -> input.nodeName() }
         val skipValidation = setOf("parallel_stack/ExpandDims/dim")
         //assertEquals(output.keys,output2.keys)
-        val notEquals = HashSet<String>()
+    /*    val notEquals = HashSet<String>()
         names.forEach {
             val value = output[it]
             val value2 = output2[it]
@@ -115,9 +117,9 @@ class TestTensorflowIR {
                 val newVar = graph.variables[it]
                 notEquals.add(it)
             }
-        }
+        }*/
 
-        println(notEquals)
+        //println(notEquals)
 
         // assertEquals(output,output2)
         //assertEquals(tfOutput,output)

nd4j/samediff-import/samediff-import-tensorflow/tensorflow-processes.pbtxt

@@ -8367,10 +8367,16 @@ mappings {
     functionName: "valuemapping"
     inputFloatName: "epsilon"
     outputDoubleName: "epsilon"
+    inputDataTypeName: "T"
+    outputDataTypeName: "dtype"
     inputToOutput {
       key: "epsilon"
       value: "epsilon"
     }
+    inputToOutput {
+      key: "dtype"
+      value: "T"
+    }
     ruleType: "attribute"
     inputFrameworkOpName: "FusedBatchNorm"
   }
@@ -12480,10 +12486,16 @@ mappings {
     functionName: "valuemapping"
     inputFloatName: "epsilon"
     outputDoubleName: "epsilon"
+    inputDataTypeName: "T"
+    outputDataTypeName: "dtype"
     inputToOutput {
       key: "epsilon"
       value: "epsilon"
     }
+    inputToOutput {
+      key: "dtype"
+      value: "T"
+    }
     ruleType: "attribute"
     inputFrameworkOpName: "FusedBatchNormV3"
   }
@@ -13056,10 +13068,16 @@ mappings {
     functionName: "valuemapping"
     inputFloatName: "epsilon"
     outputDoubleName: "epsilon"
+    inputDataTypeName: "T"
+    outputDataTypeName: "dtype"
     inputToOutput {
       key: "epsilon"
       value: "epsilon"
     }
+    inputToOutput {
+      key: "dtype"
+      value: "T"
+    }
     ruleType: "attribute"
     inputFrameworkOpName: "FusedBatchNormV2"
   }