wehub-resource-sync
75 lines
3.0 KiB
C++
75 lines
3.0 KiB
C++
//
|
|
// InnerProduct.cpp
|
|
// MNNConverter
|
|
//
|
|
// Created by MNN on 2019/01/31.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "OpConverter.hpp"
|
|
#include "logkit.h"
|
|
|
|
class InnerProductCommon : public OpConverter {
|
|
public:
|
|
virtual void run(MNN::OpT* dstOp, const caffe::LayerParameter& parameters, const caffe::LayerParameter& weight) {
|
|
auto innerproduct = new MNN::InnerProductT;
|
|
dstOp->main.value = innerproduct;
|
|
auto& l = parameters;
|
|
const caffe::InnerProductParameter& par = l.inner_product_param();
|
|
innerproduct->outputCount = par.num_output();
|
|
innerproduct->axis = 1;
|
|
if (par.has_axis()) {
|
|
innerproduct->axis = par.axis();
|
|
}
|
|
innerproduct->transpose = false;
|
|
if (par.has_transpose()) {
|
|
innerproduct->transpose = par.transpose();
|
|
}
|
|
}
|
|
InnerProductCommon() {
|
|
}
|
|
virtual ~InnerProductCommon() {
|
|
}
|
|
virtual MNN::OpType opType() {
|
|
return MNN::OpType_InnerProduct;
|
|
}
|
|
virtual MNN::OpParameter type() {
|
|
return MNN::OpParameter_InnerProduct;
|
|
}
|
|
};
|
|
|
|
class InnerProduct : public InnerProductCommon {
|
|
virtual void run(MNN::OpT* dstOp, const caffe::LayerParameter& parameters, const caffe::LayerParameter& weight) {
|
|
InnerProductCommon::run(dstOp, parameters, weight);
|
|
auto innerproduct = dstOp->main.AsInnerProduct();
|
|
const caffe::InnerProductParameter& par = parameters.inner_product_param();
|
|
const caffe::LayerParameter* v0w = &weight;
|
|
DCHECK(v0w->blobs_size() >= 1) << "caffemodel error!";
|
|
innerproduct->biasTerm = par.bias_term();
|
|
innerproduct->bias.resize(par.num_output());
|
|
::memset(innerproduct->bias.data(), 0, innerproduct->bias.size() * sizeof(float));
|
|
if (par.bias_term()) {
|
|
::memcpy(innerproduct->bias.data(), v0w->blobs(1).data().data(), par.num_output() * sizeof(float));
|
|
}
|
|
const caffe::BlobProto& WeightBlob = v0w->blobs(0);
|
|
innerproduct->weightSize = WeightBlob.data_size();
|
|
innerproduct->weight.resize(innerproduct->weightSize);
|
|
if (innerproduct->transpose) {
|
|
const float* src = WeightBlob.data().data();
|
|
float *dst = innerproduct->weight.data();
|
|
int outputCount = innerproduct->outputCount;
|
|
int srcCount = innerproduct->weightSize / outputCount;
|
|
for (int i = 0; i < outputCount; i++) {
|
|
for (int j = 0; j < srcCount; j++) {
|
|
dst[i * srcCount + j] = src[i + j * outputCount];
|
|
}
|
|
}
|
|
innerproduct->transpose = false;
|
|
} else {
|
|
::memcpy(innerproduct->weight.data(), WeightBlob.data().data(), sizeof(float) * innerproduct->weightSize);
|
|
}
|
|
}
|
|
};
|
|
|
|
static OpConverterRegister<InnerProduct> a("InnerProduct");
|