alibaba--mnn/tools/converter/source/tensorflow/tfOpConverter.cpp

//
//  tfOpConverter.cpp
//  MNNConverter
//
//  Created by MNN on 2019/01/31.
//  Copyright © 2018, Alibaba Group Holding Limited
//

#include "tfOpConverter.hpp"
#include "OpCount.hpp"
#include <stdlib.h>
using namespace MNN;
#define FUNCTION(dstType, srcType, contentType)\
static void _##dstType##srcType##_##contentType(BlobT* dst, const ::tensorflow::TensorProto& tensor, int dataSize) {\
dst->dstType.resize(dataSize);\
if (tensor.srcType##_size() == 1) {\
for (int i=0; i<dataSize; ++i) {\
dst->dstType[i] = tensor.srcType(0);\
}\
return;\
}\
if (tensor.srcType().empty()) {\
contentType* source = (contentType*)tensor.tensor_content().data();\
for (int i=0; i<dataSize; ++i) {\
dst->dstType[i] = source[i];\
}\
return;\
}\
for (int i=0; i<dataSize; ++i) {\
dst->dstType[i] = tensor.srcType(i);\
}\
}\

FUNCTION(float32s, double_val, double);
FUNCTION(float32s, float_val, float);
FUNCTION(int32s, int_val, int32_t);
FUNCTION(int32s, int64_val, int64_t);
FUNCTION(uint8s, int64_val, uint8_t);
FUNCTION(int8s, int64_val, int8_t);
FUNCTION(int32s, bool_val, uint8_t);
FUNCTION(strings, string_val, uint8_t);


typedef void(*proc)(BlobT* dst, const ::tensorflow::TensorProto& tensor, int dataSize);
void tfOpConverter::convertTensorToBlob(MNN::BlobT* parameter, const ::tensorflow::TensorProto& tensor) {
    parameter->dataFormat = MNN::MNN_DATA_FORMAT_NHWC;
    MNN::DataType dataType = MNN::DataType_DT_INVALID;
    dataType               = (MNN::DataType)tensor.dtype();
    
    //origin type in tensorflow, mnn's data type, tensor_content's data type
    std::map<MNN::DataType, std::pair<MNN::DataType, proc> > supporting {
        {DataType_DT_DOUBLE, {DataType_DT_FLOAT, _float32sdouble_val_double}},
        {DataType_DT_FLOAT, {DataType_DT_FLOAT, _float32sfloat_val_float}},
        {DataType_DT_INT32, {DataType_DT_INT32, _int32sint_val_int32_t}},
        {DataType_DT_INT64, {DataType_DT_INT32, _int32sint64_val_int64_t}},
        {DataType_DT_INT8, {DataType_DT_INT8, _int8sint64_val_int8_t}},
        {DataType_DT_UINT8, {DataType_DT_UINT8, _uint8sint64_val_uint8_t}},
        {DataType_DT_BOOL, {DataType_DT_INT32, _int32sbool_val_uint8_t}},
        {DataType_DT_STRING, {DataType_DT_STRING, _stringsstring_val_uint8_t}},
    };
    bool isSupport = supporting.find(dataType) != supporting.end();
    CHECK(isSupport) << "Const Data Type Not Supported!!!==> " << dataType;
    CHECK(dataType <= MNN::DataType_MAX) << "Const Data Type Not Supported!!!==> " << dataType;
    auto convert = supporting[dataType];
    parameter->dataType = convert.first;
    size_t dimSize = tensor.tensor_shape().dim_size();
    parameter->dims.resize(dimSize);
    size_t dataSize = 1;
    for (int i = 0; i < dimSize; i++) {
        dataSize           = dataSize * tensor.tensor_shape().dim(i).size();
        parameter->dims[i] = tensor.tensor_shape().dim(i).size();
    }
    convert.second(parameter, tensor, dataSize);
}

tfOpConverterSuit *tfOpConverterSuit::global = nullptr;
class DefaultTfOpConverter : public tfOpConverter {
public:
    virtual void run(MNN::OpT *dstOp, TmpNode *srcNode) override {
        dstOp->main.value                                                     = new MNN::ExtraT;
        dstOp->main.AsExtra()->engine                                         = "Tensorflow";
        dstOp->main.AsExtra()->type                                           = srcNode->opType;
        const google::protobuf::Map<std::string, tensorflow::AttrValue> &attr = srcNode->tfNode->attr();
        for (auto iter = attr.begin(); iter != attr.end(); iter++) {
            auto attrExtr = ConvertTfAttribute(iter->first/*attr key*/,
                                               iter->second/*attr*/);
            dstOp->main.AsExtra()->attr.emplace_back(std::move(attrExtr));
        }
    }
    virtual MNN::OpParameter type() override {
        return MNN::OpParameter_Extra;
    }
    virtual MNN::OpType opType() override {
        return MNN::OpType_Extra;
    }

private:
    std::unique_ptr<MNN::AttributeT> ConvertTfAttribute(
        const std::string& attr_name,
        const tensorflow::AttrValue& tf_attr) const;
};

std::unique_ptr<MNN::AttributeT> DefaultTfOpConverter::ConvertTfAttribute(
        const std::string& attr_name,
        const tensorflow::AttrValue& tf_attr) const {
    std::unique_ptr<MNN::AttributeT> attrExtr(new MNN::AttributeT);
    attrExtr->key  = attr_name;
    attrExtr->s    = tf_attr.s();
    attrExtr->f    = tf_attr.f();
    attrExtr->i    = (int)tf_attr.i();
    attrExtr->b    = tf_attr.b();
    if (tf_attr.has_tensor()) {
        attrExtr->tensor.reset(new BlobT);
        convertTensorToBlob(attrExtr->tensor.get(), tf_attr.tensor());
    }
    attrExtr->type = (MNN::DataType)tf_attr.type();
    if (tf_attr.has_list()) {
        auto &listValue = tf_attr.list();
        attrExtr->list.reset(new MNN::ListValueT);
        for (int j = 0; j < listValue.s_size(); ++j) {
            attrExtr->list->s.push_back(listValue.s(j));
        }
        for (int j = 0; j < listValue.b_size(); ++j) {
            attrExtr->list->b.push_back(listValue.b(j));
        }
        for (int j = 0; j < listValue.i_size(); ++j) {
            attrExtr->list->i.push_back(listValue.i(j));
        }
        for (int j = 0; j < listValue.f_size(); ++j) {
            attrExtr->list->f.push_back(listValue.f(j));
        }
        for (int j = 0; j < listValue.type_size(); ++j) {
            attrExtr->list->type.push_back((MNN::DataType)listValue.type(j));
        }
    }
    if (tf_attr.has_func()) {
        auto &func = tf_attr.func();
        attrExtr->func.reset(new MNN::NamedAttrListT);
        attrExtr->func->name = func.name();
        for (const auto& it : func.attr()) {
            auto func_attr = ConvertTfAttribute(it.first, it.second);
            attrExtr->func->attr.push_back(std::move(func_attr));
        }
    }
    return std::move(attrExtr);
}

tfOpConverter *tfOpConverterSuit::search(const std::string &name) {
    auto iter = mTests.find(name);
    if (iter == mTests.end()) {
        static DefaultTfOpConverter converter;
        return &converter;
    }
    return iter->second;
}

tfOpConverterSuit *tfOpConverterSuit::get() {
    if (global == nullptr)
        global = new tfOpConverterSuit;
    return global;
}

tfOpConverterSuit::~tfOpConverterSuit() {
    for (auto &iter : mTests) {
        delete iter.second;
    }
    mTests.clear();
}

void tfOpConverterSuit::insert(tfOpConverter *t, const char *name) {
    OpCount::get()->insertOp("TF", name);
    mTests.insert(std::make_pair(name, t));
}