Files
2026-07-13 13:33:03 +08:00

98 lines
3.4 KiB
C++

//
// ReductionTf.cpp
// MNNConverter
//
// Created by MNN on 2019/01/31.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include <string.h>
#include "TfUtils.hpp"
#include "tfOpConverter.hpp"
#include "graph.pb.h"
DECLARE_OP_CONVERTER(ReductionTf);
MNN::OpType ReductionTf::opType() {
return MNN::OpType_Reduction;
}
MNN::OpParameter ReductionTf::type() {
return MNN::OpParameter_ReductionParam;
}
void ReductionTf::run(MNN::OpT *dstOp, TmpNode *srcNode) {
auto reductionParam = new MNN::ReductionParamT;
// reduction parameter
tensorflow::AttrValue value;
reductionParam->dType = MNN::DataType_DT_FLOAT;
if (find_attr_value(srcNode->tfNode, "T", value)) {
reductionParam->dType = (MNN::DataType)value.type();
}
reductionParam->keepDims = false;
if (find_attr_value(srcNode->tfNode, "keep_dims", value)) {
reductionParam->keepDims = value.b();
}
#ifdef TF_CONVERT_ORIGIN
TmpNode *constNode = tempGraph->_getTmpNode(srcNode->inEdges[1]);
if (constNode->opType == "Const") {
if (find_attr_value(constNode->tfNode, "value", value)) {
const tensorflow::TensorProto &reductionIndices = value.tensor();
const tensorflow::TensorShapeProto &reductionIndicesShape = reductionIndices.tensor_shape();
int dimSize = 1;
if (reductionIndicesShape.dim_size() > 0) {
dimSize = reductionIndicesShape.dim(0).size();
}
reductionParam->dim.resize(dimSize);
if (reductionIndices.int_val_size() > 0) {
for (int i = 0; i < dimSize; ++i) {
reductionParam->dim[i] = reductionIndices.int_val(i);
}
} else {
DCHECK((MNN::DataType)reductionIndices.dtype() == MNN::DataType_DT_INT32);
DCHECK(reductionIndices.tensor_content().size() > 0);
const int *dimData = (int *)reductionIndices.tensor_content().c_str();
for (int i = 0; i < dimSize; i++) {
reductionParam->dim[i] = dimData[i];
}
}
}
}
#endif
// reduction operation
if (srcNode->opType == "Mean") {
reductionParam->operation = MNN::ReductionType_MEAN;
} else if (srcNode->opType == "Max") {
reductionParam->operation = MNN::ReductionType_MAXIMUM;
} else if (srcNode->opType == "Min") {
reductionParam->operation = MNN::ReductionType_MINIMUM;
} else if (srcNode->opType == "Sum") {
reductionParam->operation = MNN::ReductionType_SUM;
} else if (srcNode->opType == "Any") {
reductionParam->operation = MNN::ReductionType_ANY;
} else if (srcNode->opType == "All") {
reductionParam->operation = MNN::ReductionType_ALL;
} else if (srcNode->opType == "Prod") {
reductionParam->operation = MNN::ReductionType_PROD;
} else {
DLOG(ERROR) << "MNN Converter Not "
"Supported!!! ===> "
<< srcNode->opType;
}
reductionParam->coeff = 0.0f; // defalut
dstOp->main.value = reductionParam;
}
REGISTER_CONVERTER(ReductionTf, Mean);
REGISTER_CONVERTER(ReductionTf, Max);
REGISTER_CONVERTER(ReductionTf, Min);
REGISTER_CONVERTER(ReductionTf, Any);
REGISTER_CONVERTER(ReductionTf, All);
REGISTER_CONVERTER(ReductionTf, Sum);
REGISTER_CONVERTER(ReductionTf, Prod);