Brutex
/
cavis
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add new clion rules, fix batch norml

master
agibsonccc 2021-02-09 07:44:23 +09:00
parent 968eaad2dd
commit 5bd386a4f9
8 changed files with 232 additions and 186 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
libnd4j/blas/CMakeLists.txt
View File

@ -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...")

48
libnd4j/include/ops/declarable/generic/nn/fusedBatchNorm.cpp
View File

@ -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());
variance = 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, sd::DataType::FLOAT32, block.launchContext());
}
// FIXME: double?
double epsilon;
if(block.getTArguments()->size() > 0)
epsilon = T_ARG(0) > 1.001e-5 ? T_ARG(0) : 1.001e-5;
else
epsilon = 0.001;
float epsilon;
if(block.getTArguments()->size() > 0) {
epsilon = (float) (T_ARG(0) > 1.001e-5 ? T_ARG(0) : 1.001e-5);
}
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 double restSizeInv = 1.0 / restSize;
const double restSizeAdjust = (double)restSize / restSizeMinusOne;
const float restSizeInv = 1.0f / restSize;
const float restSizeAdjust = (float)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;
//delete sum;
auto varOutput = (*variance) * restSizeAdjust;
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));
}
}
}

3
nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/api/ops/custom/FusedBatchNorm.java
View File

@ -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);

9
nd4j/nd4j-backends/nd4j-tests/src/test/java/org/nd4j/imports/TFGraphs/TFGraphTestAllSameDiff.java
View File

@ -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",
"max_pool_with_argmax/int64_int64_padding_SAME"
// "fused_batch_norm/float16_nhwc",
"fused_batch_norm/float32_nhwc"
// , "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/.*",

6
nd4j/samediff-import/samediff-import-tensorflow/src/main/kotlin/org/nd4j/samediff/frameworkimport/tensorflow/definitions/TensorflowOpDeclarations.kt
View File

@ -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))
)

18
nd4j/samediff-import/samediff-import-tensorflow/src/main/resources/tensorflow-mapping-ruleset.pbtxt
View File

@ -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"
}

10
nd4j/samediff-import/samediff-import-tensorflow/src/test/kotlin/org/nd4j/samediff/frameworkimport/tensorflow/TestTensorflowIR.kt
View File

@ -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)

18
nd4j/samediff-import/samediff-import-tensorflow/tensorflow-processes.pbtxt
View File

@ -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"
}