wehub-resource-sync
101 lines
3.1 KiB
C++
101 lines
3.1 KiB
C++
//
|
|
// SliceTorch.cpp
|
|
// MNNConverter
|
|
//
|
|
// Created by MNN on 2021/05/10.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include <stdio.h>
|
|
#include "torchOpConverter.hpp"
|
|
|
|
DECLARE_OP_CONVERTER(SliceTorch);
|
|
|
|
MNN::OpType SliceTorch::opType() {
|
|
return MNN::OpType_Slice;
|
|
}
|
|
MNN::OpParameter SliceTorch::type() {
|
|
return MNN::OpParameter_Slice;
|
|
}
|
|
std::vector<int> SliceTorch::inputTensorIdx() {
|
|
return {0};
|
|
}
|
|
|
|
void SliceTorch::run(MNN::OpT* dstOp, const torch::jit::Node* node, TorchScope* scope) {
|
|
auto param = new MNN::SliceT;
|
|
const std::string opType = getRealOpType(node);
|
|
if (opType == "chunk") {
|
|
param->axis = getValue<int64_t>(node->input(2));
|
|
param->sourceType = MNN::NetSource_TENSORFLOW;
|
|
} else if (opType == "ListUnpack") {
|
|
param->axis = 0;
|
|
param->sourceType = MNN::NetSource_TENSORFLOW;
|
|
}
|
|
dstOp->main.value = param;
|
|
}
|
|
|
|
REGISTER_CONVERTER(SliceTorch, chunk);
|
|
REGISTER_CONVERTER(SliceTorch, ListUnpack);
|
|
|
|
DECLARE_OP_CONVERTER(StridedSliceTorch);
|
|
|
|
MNN::OpType StridedSliceTorch::opType() {
|
|
return MNN::OpType_Extra;
|
|
}
|
|
MNN::OpParameter StridedSliceTorch::type() {
|
|
return MNN::OpParameter_Extra;
|
|
}
|
|
std::vector<int> StridedSliceTorch::inputTensorIdx() {
|
|
return {-1};
|
|
}
|
|
|
|
void StridedSliceTorch::run(MNN::OpT* dstOp, const torch::jit::Node* node, TorchScope* scope) {
|
|
auto extra = new MNN::ExtraT;
|
|
dstOp->main.value = extra;
|
|
extra->engine = "Torch";
|
|
extra->type = getRealOpType(node);
|
|
if (node->inputs().size() > 1) {
|
|
auto dim = node->input(1);
|
|
if (toIValue(dim)) {
|
|
std::unique_ptr<MNN::AttributeT> dimAttr(new MNN::AttributeT);
|
|
dimAttr->key = "dim";
|
|
dimAttr->i = getValue<int64_t>(dim);
|
|
extra->attr.emplace_back(std::move(dimAttr));
|
|
}
|
|
}
|
|
if (node->inputs().size() > 2) {
|
|
auto start = node->input(2);
|
|
if (toIValue(start)) {
|
|
std::unique_ptr<MNN::AttributeT> startAttr(new MNN::AttributeT);
|
|
startAttr->key = "start";
|
|
startAttr->i = getValue<int64_t>(start);
|
|
extra->attr.emplace_back(std::move(startAttr));
|
|
}
|
|
}
|
|
if (node->inputs().size() > 3) {
|
|
auto end = node->input(3);
|
|
if (toIValue(end)) {
|
|
std::unique_ptr<MNN::AttributeT> endAttr(new MNN::AttributeT);
|
|
endAttr->key = "end";
|
|
endAttr->i = std::min(getValue<int64_t>(end), static_cast<int64_t>(std::numeric_limits<int>::max()));
|
|
extra->attr.emplace_back(std::move(endAttr));
|
|
}
|
|
}
|
|
if (node->inputs().size() > 4) {
|
|
auto stride = node->input(4);
|
|
if (toIValue(stride)) {
|
|
std::unique_ptr<MNN::AttributeT> strideAttr(new MNN::AttributeT);
|
|
strideAttr->key = "stride";
|
|
strideAttr->i = getValue<int64_t>(stride);
|
|
extra->attr.emplace_back(std::move(strideAttr));
|
|
}
|
|
} else {
|
|
std::unique_ptr<MNN::AttributeT> strideAttr(new MNN::AttributeT);
|
|
strideAttr->key = "stride";
|
|
strideAttr->i = 1;
|
|
extra->attr.emplace_back(std::move(strideAttr));
|
|
}
|
|
}
|
|
|
|
REGISTER_CONVERTER(StridedSliceTorch, slice);
|