【新增】MNN依赖库

This commit is contained in:
wdp
2025-01-18 21:11:17 +08:00
parent 8de1dfb3aa
commit 4b43617901
17 changed files with 4182 additions and 0 deletions

View File

@@ -0,0 +1,72 @@
//
// Executor.hpp
// MNN
//
// Created by MNN on 2019/07/25.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifndef Executor_hpp
#define Executor_hpp
#include <MNN/ErrorCode.hpp>
#include <MNN/expr/Expr.hpp>
#include <MNN/Tensor.hpp>
#include <MNN/Interpreter.hpp>
#include <vector>
#include <mutex>
#include <set>
#include <MNN/MNNForwardType.h>
namespace MNN {
class Backend;
class Execution;
class Runtime;
struct Op;
namespace Express {
class MNN_PUBLIC Executor {
public:
class ComputeCache;
struct Unit;
static void setShapeDirty(ComputeCache* cache);
static void setContentDirty(ComputeCache* cache);
static void* mapOutput(ComputeCache* cache, int offset, Tensor* dest);
struct Requirement {
std::vector<bool> contentNeedContent;
std::vector<bool> shapeNeedContent;
};
~Executor();
Requirement getRequirement(Expr* expr) const;
ErrorCode computeInfo(Expr* expr);
void makeCache(const std::vector<EXPRP>& expr, bool forceCPU = false);
ErrorCode runCache(std::shared_ptr<ComputeCache> cache);
void setGlobalExecutorConfig(MNNForwardType type, const BackendConfig& config, int numberThread);
enum GCFlag {
FULL,
PART
};
void gc(GCFlag flag = FULL);
static std::shared_ptr<Executor> getGlobalExecutor();
static std::shared_ptr<Executor> newExecutor(MNNForwardType type,
const BackendConfig& config,
int numberThread);
void resetProfile();
void dumpProfile();
void addOpCostTime(int op, float costTime);
void addOpCostTime(const std::string& type, float costTime);
void addOpFlops(const std::string& type, float flops);
class Profiler;
static RuntimeInfo getRuntime();
private:
void _makeCache(const std::vector<EXPRP>& outputs, bool forceCPU);
void _create(const std::vector<EXPRP>& outputs, std::set<std::shared_ptr<Executor::ComputeCache>>&& inputCaches, std::set<std::shared_ptr<Expr::Inside>>&& inputNode, bool forceCPU);
void _visit(EXPRP expr, std::set<std::shared_ptr<Executor::ComputeCache>>& inputCaches, std::set<std::shared_ptr<Expr::Inside>>& inputNode);
Executor(std::shared_ptr<Runtime> backend, MNNForwardType type);
std::pair<std::shared_ptr<Runtime>, MNNForwardType> mRuntime;
std::pair<std::shared_ptr<Runtime>, MNNForwardType> mBackupRuntime;
std::mutex mMutex;
std::shared_ptr<Profiler> mProfiler;
};
} // namespace Express
} // namespace MNN
#endif

View File

@@ -0,0 +1,264 @@
//
// Expr.hpp
// MNN
//
// Created by MNN on 2019/06/10.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifndef Expr_hpp
#define Expr_hpp
#include <functional>
#include <list>
#include <string>
#include <vector>
#include <map>
#include <memory>
#include <MNN/HalideRuntime.h>
#include <MNN/MNNDefine.h>
namespace MNN {
struct OpT;
struct Op;
struct NetT;
namespace Express {
class Variable;
class Expr;
class Executor;
typedef std::shared_ptr<Expr> EXPRP;
typedef std::weak_ptr<Expr> WeakEXPRP;
typedef std::vector<int> INTS;
enum Dimensionformat { NHWC, NC4HW4, NCHW };
class MNN_PUBLIC VARP {
public:
VARP() {
// Do nothing
}
VARP(std::shared_ptr<Variable> c) {
mContent = std::move(c);
}
VARP(Variable* c) {
mContent.reset(c);
}
Variable* get() const {
return mContent.get();
}
~ VARP() {
// Do nothing
}
VARP(const VARP& var) {
mContent = var.mContent;
}
VARP(VARP&& var) {
mContent = std::move(var.mContent);
}
VARP operator+(VARP var) const;
VARP operator-(VARP var) const;
VARP operator*(VARP var) const;
VARP operator/(VARP var) const;
VARP mean(INTS dims) const;
VARP sum(INTS dims) const;
bool operator==(const VARP& var) const {
return var.mContent == mContent;
}
bool operator<(const VARP& var) const {
return mContent < var.mContent;
}
bool operator<=(const VARP& var) const {
return mContent <= var.mContent;
}
VARP& operator=(const VARP& var) {
mContent = var.mContent;
return *this;
}
VARP& operator=(Variable* var) {
mContent.reset(var);
return *this;
}
Variable* operator->() const {
return mContent.get();
}
enum InputType {
INPUT = 0,
CONSTANT = 1,
TRAINABLE = 2,
};
bool fix(InputType type) const;
private:
friend class Variable;
std::shared_ptr<Variable> mContent;
};
inline bool operator==(Variable* src, VARP dst) {
return src == dst.get();
}
inline bool operator!=(Variable* src, VARP dst) {
return src != dst.get();
}
// inline bool operator<(VARP src, VARP dst) {
// return src.get() < dst.get();
// }
typedef std::vector<VARP> VARPS;
class MNN_PUBLIC Variable {
public:
struct Info {
Dimensionformat order = NHWC;
INTS dim;
halide_type_t type;
int size;
void syncSize();
};
const std::string& name() const;
void setName(const std::string& name);
std::pair<EXPRP, int> expr() const {
return std::make_pair(mFrom, mFromIndex);
}
// If compute info error, return nullptr
const Info* getInfo();
bool resize(INTS dims);
template <typename T>
const T* readMap() {
return (const T*)readInternal();
}
template <typename T>
T* writeMap() {
return (T*)writeInternal();
}
//Depecerate
void unMap();
bool input(VARP src);
static void replace(VARP dst, VARP src);
static VARP create(EXPRP expr, int index = 0);
static std::vector<VARP> load(const char* fileName);
static std::map<std::string, VARP> loadMap(const char* fileName);
static std::vector<VARP> load(const uint8_t* buffer, size_t length);
static std::map<std::string, VARP> loadMap(const uint8_t* buffer, size_t length);
static std::pair<std::map<std::string, VARP>, std::map<std::string, VARP>> getInputAndOutput(const std::map<std::string, VARP>& allVariable);
static std::vector<VARP> mapToSequence(const std::map<std::string, VARP>& source);
static std::vector<EXPRP> getExecuteOrder(const std::vector<VARP>& output);
static void save(const std::vector<VARP>& vars, const char* fileName);
static void save(const std::vector<VARP>& vars, NetT* dest);
// Pack a few Variable to compute in one pipeline
static void prepareCompute(const std::vector<VARP>& vars, bool forceCPU = false);
size_t linkNumber() const;
const std::vector<WeakEXPRP>& toExprs() const;
void setExpr(EXPRP expr, int index) {
mFrom = expr;
mFromIndex = index;
}
private:
Variable(EXPRP expr, int index) {
mFrom = expr;
mFromIndex = index;
}
void* readInternal(bool forShape = false);
void* writeInternal(bool inform=true);
void informDirty();
friend class Expr;
EXPRP mFrom;
int mFromIndex;
};
class MNN_PUBLIC Expr {
public:
struct Inside;
static EXPRP create(Variable::Info&& info, const void* ptr, VARP::InputType type, bool copy = true);
static EXPRP create(const OpT* op, std::vector<VARP> inputs, int outputSize = 1);
static EXPRP create(std::pair<std::shared_ptr<char>, int> extra, std::vector<VARP>&& inputs, int outputSize = 1);
static EXPRP create(std::unique_ptr<OpT>&& op, std::vector<VARP> inputs, int outputSize = 1) {
return create(op.get(), inputs, outputSize);
}
void setName(const std::string& name);
const Op* get() const {
return mOp;
}
const std::vector<VARP>& inputs() const {
return mInputs;
}
int outputSize() const {
return (int)mOutputNames.size();
}
static void replace(EXPRP oldExpr, EXPRP newExpr);
bool requireInfo();
void visitOutputs(const std::function<bool(EXPRP, int)>& visit);
static void visit(EXPRP expr, const std::function<bool(EXPRP)>& before, const std::function<bool(EXPRP)>& after);
const std::vector<WeakEXPRP>& outputs() const {
return mTo;
}
~Expr();
bool visited() const {
return mVisited;
}
void setVisited(bool visited) {
mVisited = visited;
}
const std::string& name() const {
return mName;
}
const std::string& outputName(int index) {
return mOutputNames[index];
}
VARP::InputType inputType() const {return mType;}
Variable::Info* outputInfo(int index) const;
std::pair<std::shared_ptr<char>, int> extra() const {
return std::make_pair(mExtraBuffer, mOpBufferSize);
}
bool setInfoDirty();
std::shared_ptr<Inside> inside() const {
return mInside;
}
bool valid() const {
return mValid;
}
void setEntry(const std::vector<VARP>& entries) {
mEntries = entries;
}
const std::vector<VARP>& getEntry() const {
return mEntries;
}
private:
static void _addLinkForInputs(EXPRP expr);
Expr(int outputSize);
friend class Variable;
friend class VARP;
VARP::InputType mType;
const Op* mOp;
std::vector<VARP> mInputs;
std::vector<std::string> mOutputNames;
bool mValid = true;
std::shared_ptr<char> mExtraBuffer;
int mOpBufferSize = 0;
std::string mName;
std::shared_ptr<Inside> mInside = nullptr;
bool mVisited = false;
std::vector<WeakEXPRP> mTo;
// Only the enter input has entries, and it helps to get info for enter
// input expression.
std::vector<VARP> mEntries;
};
} // namespace Express
} // namespace MNN
#endif /* Expr_hpp */

View File

@@ -0,0 +1,16 @@
//
// ExprCreator.hpp
// MNN
//
// Created by MNN on 2019/06/27.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifndef ExprCreator_hpp
#define ExprCreator_hpp
#include <MNN/expr/Expr.hpp>
#include <MNN/expr/MathOp.hpp>
#include <MNN/expr/NeuralNetWorkOp.hpp>
#endif

View File

@@ -0,0 +1,128 @@
//
// MathOp.hpp
// MNN
//
// Created by MNN on 2019/06/27.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifndef MathOp_HPP
#define MathOp_HPP
namespace MNN {
namespace Express {
//BinaryOPs
MNN_PUBLIC VARP _Add(VARP x, VARP y);
MNN_PUBLIC VARP _Subtract(VARP x, VARP y);
MNN_PUBLIC VARP _Multiply(VARP x, VARP y);
MNN_PUBLIC VARP _Divide(VARP x, VARP y);
MNN_PUBLIC VARP _Pow(VARP x, VARP y);
MNN_PUBLIC VARP _Minimum(VARP x, VARP y);
MNN_PUBLIC VARP _Maximum(VARP x, VARP y);
MNN_PUBLIC VARP _BiasAdd(VARP value, VARP bias);
MNN_PUBLIC VARP _Greater(VARP x, VARP y);
MNN_PUBLIC VARP _GreaterEqual(VARP x, VARP y);
MNN_PUBLIC VARP _Less(VARP x, VARP y);
MNN_PUBLIC VARP _FloorDiv(VARP x, VARP y);
MNN_PUBLIC VARP _SquaredDifference(VARP x, VARP y);
MNN_PUBLIC VARP _Equal(VARP x, VARP y);
MNN_PUBLIC VARP _LessEqual(VARP x, VARP y);
MNN_PUBLIC VARP _FloorMod(VARP x, VARP y);
MNN_PUBLIC VARP _Atan2(VARP x, VARP y);
MNN_PUBLIC VARP _LogicalOr(VARP x, VARP y);
MNN_PUBLIC VARP _NotEqual(VARP x, VARP y);
//UnaryOPs
MNN_PUBLIC VARP _Sign(VARP a);
MNN_PUBLIC VARP _Abs(VARP x);
MNN_PUBLIC VARP _Negative(VARP x);
MNN_PUBLIC VARP _Floor(VARP x);
MNN_PUBLIC VARP _Round(VARP x);
MNN_PUBLIC VARP _Ceil(VARP x);
MNN_PUBLIC VARP _Square(VARP x);
MNN_PUBLIC VARP _Sqrt(VARP x);
MNN_PUBLIC VARP _Rsqrt(VARP x);
MNN_PUBLIC VARP _Exp(VARP x);
MNN_PUBLIC VARP _Log(VARP x);
MNN_PUBLIC VARP _Sin(VARP x);
MNN_PUBLIC VARP _Sinh(VARP x);
MNN_PUBLIC VARP _Cos(VARP x);
MNN_PUBLIC VARP _Cosh(VARP x);
MNN_PUBLIC VARP _Tan(VARP x);
MNN_PUBLIC VARP _Asin(VARP x);
MNN_PUBLIC VARP _Asinh(VARP x);
MNN_PUBLIC VARP _Acos(VARP x);
MNN_PUBLIC VARP _Acosh(VARP x);
MNN_PUBLIC VARP _Atan(VARP x);
MNN_PUBLIC VARP _Atanh(VARP x);
MNN_PUBLIC VARP _Reciprocal(VARP x);
MNN_PUBLIC VARP _Log1p(VARP x);
//Only one but not in UnaryOPs
MNN_PUBLIC VARP _Tanh(VARP x);
MNN_PUBLIC VARP _Sigmoid(VARP x);
MNN_PUBLIC VARP _Erf(VARP x);
MNN_PUBLIC VARP _Erfc(VARP x);
MNN_PUBLIC VARP _Erfinv(VARP x);
MNN_PUBLIC VARP _Expm1(VARP x);
//ReduceOPs
MNN_PUBLIC VARP _ReduceSum(VARP input_variable, INTS axis = {}, bool keepDims = false);
MNN_PUBLIC VARP _ReduceMean(VARP input_variable, INTS axis = {}, bool keepDims = false);
MNN_PUBLIC VARP _ReduceMax(VARP input_variable, INTS axis = {}, bool keepDims = false);
MNN_PUBLIC VARP _ReduceMin(VARP input_variable, INTS axis = {}, bool keepDims = false);
MNN_PUBLIC VARP _ReduceProd(VARP input_variable, INTS axis = {}, bool keepDims = false);
MNN_PUBLIC VARP _ReduceAny(VARP input_variable, INTS axis = {}, bool keepDims = false);
MNN_PUBLIC VARP _ReduceAll(VARP input_variable, INTS axis = {}, bool keepDims = false);
MNN_PUBLIC VARP _ReduceSumMutable(VARP input_variable, VARP axis, bool keepDims = false);
MNN_PUBLIC VARP _ReduceMeanMutable(VARP input_variable, VARP axis, bool keepDims = false);
MNN_PUBLIC VARP _ReduceMaxMutable(VARP input_variable, VARP axis, bool keepDims = false);
MNN_PUBLIC VARP _ReduceMinMutable(VARP input_variable, VARP axis, bool keepDims = false);
MNN_PUBLIC VARP _ReduceProdMutable(VARP input_variable, VARP axis, bool keepDims = false);
MNN_PUBLIC VARP _ReduceAnyMutable(VARP input_variable, VARP axis, bool keepDims = false);
MNN_PUBLIC VARP _ReduceAllMutable(VARP input_variable, VARP axis, bool keepDims = false);
//EltwiseOPs
MNN_PUBLIC VARP _Prod(VARP a, VARP b, std::vector<float> coeff);
MNN_PUBLIC VARP _Sum(VARP a, VARP b, std::vector<float> coeff);
MNN_PUBLIC VARP _Max(VARP a, VARP b, std::vector<float> coeff);
MNN_PUBLIC VARP _Sub(VARP a, VARP b, std::vector<float> coeff);
MNN_PUBLIC VARP _EltwiseProdInt8(VARP x, VARP y,
std::vector<int8_t> x_weight, std::vector<int32_t> x_bias, std::vector<float> x_scale, std::vector<float> x_tensorScale,
std::vector<int8_t> y_weight, std::vector<int32_t> y_bias, std::vector<float> y_scale, std::vector<float> y_tensorScale,
std::vector<int8_t> output_weight, std::vector<int32_t> output_bias, std::vector<float> output_scale, std::vector<float> output_tensorScale);
MNN_PUBLIC VARP _EltwiseSumInt8(VARP x, VARP y,
std::vector<int8_t> x_weight, std::vector<int32_t> x_bias, std::vector<float> x_scale, std::vector<float> x_tensorScale,
std::vector<int8_t> y_weight, std::vector<int32_t> y_bias, std::vector<float> y_scale, std::vector<float> y_tensorScale,
std::vector<int8_t> output_weight, std::vector<int32_t> output_bias, std::vector<float> output_scale, std::vector<float> output_tensorScale);
MNN_PUBLIC VARP _EltwiseSubInt8(VARP x, VARP y,
std::vector<int8_t> x_weight, std::vector<int32_t> x_bias, std::vector<float> x_scale, std::vector<float> x_tensorScale,
std::vector<int8_t> y_weight, std::vector<int32_t> y_bias, std::vector<float> y_scale, std::vector<float> y_tensorScale,
std::vector<int8_t> output_weight, std::vector<int32_t> output_bias, std::vector<float> output_scale, std::vector<float> output_tensorScale);
MNN_PUBLIC VARP _EltwiseMaxInt8(VARP x, VARP y,
std::vector<int8_t> x_weight, std::vector<int32_t> x_bias, std::vector<float> x_scale, std::vector<float> x_tensorScale,
std::vector<int8_t> y_weight, std::vector<int32_t> y_bias, std::vector<float> y_scale, std::vector<float> y_tensorScale,
std::vector<int8_t> output_weight, std::vector<int32_t> output_bias, std::vector<float> output_scale, std::vector<float> output_tensorScale);
//OtherOPs
template<typename T>
VARP _Cast(VARP x) {
return _Cast(x, halide_type_of<T>());
}
MNN_PUBLIC VARP _Cast(VARP x, halide_type_t dtype);
MNN_PUBLIC VARP _MatMul(VARP a, VARP b, bool tranposeA = false, bool tranposeB = false);
MNN_PUBLIC VARP _Normalize(VARP x, int32_t acrossSpatial, int32_t channelShared, float eps, std::vector<float> scale);
MNN_PUBLIC VARP _ArgMax(VARP input, int axis = 0);
MNN_PUBLIC VARP _ArgMin(VARP input, int axis = 0);
MNN_PUBLIC VARP _BatchMatMul(VARP x, VARP y, bool adj_x = false, bool adj_y = false);
MNN_PUBLIC VARP _UnravelIndex(VARP indices, VARP dims);
MNN_PUBLIC VARP _ScatterNd(VARP indices, VARP updates, VARP shape);
MNN_PUBLIC VARP _OneHot(VARP indices, VARP depth, VARP onValue, VARP offValue, int axis = -1);
MNN_PUBLIC VARP _BroadcastTo(VARP a, VARP shape);
MNN_PUBLIC VARP _LinSpace(VARP start, VARP stop, VARP num);
}; // namespace Express
}; // namespace MNN
#endif /* MathOp_HPP */

View File

@@ -0,0 +1,142 @@
//
// NeuralNetWorkOp.hpp
// MNN
//
// Created by MNN on 2019/06/27.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifndef NeuralNetWorkOp_HPP
#define NeuralNetWorkOp_HPP
namespace MNN {
namespace Express {
enum PaddingMode {CAFFE, VALID, SAME};
enum PoolingMode {MAXPOOL, AVEPOOL};
enum PadValueMode {CONSTANT, REFLECT, SYMMETRIC};
MNN_PUBLIC VARP _Input(INTS shape = {}, Dimensionformat data_format = NC4HW4, halide_type_t dtype = halide_type_of<float>()) ;
MNN_PUBLIC VARP _Clone(VARP source, bool deepCopy = false);
MNN_PUBLIC VARP _Scalar(const void* ptr, halide_type_t type);
template <typename T>
VARP _Scalar(T value) {
return _Scalar(&value, halide_type_of<T>());
}
MNN_PUBLIC VARP _Const(float value, INTS shape = {}, Dimensionformat format = NHWC);
MNN_PUBLIC VARP _Const(const void* ptr, INTS shape = {}, Dimensionformat format = NHWC,
halide_type_t type = halide_type_of<float>());
MNN_PUBLIC VARP _TrainableParam(float value, INTS dims, Dimensionformat format);
MNN_PUBLIC VARP _TrainableParam(const void* ptr, INTS dims, Dimensionformat format,
halide_type_t type = halide_type_of<float>());
MNN_PUBLIC VARP _InnerProduct(std::vector<float>&& weight, std::vector<float>&& bias, VARP x, INTS outputShape);
MNN_PUBLIC VARP _Conv(VARP weight, VARP bias, VARP x, PaddingMode pad = VALID, INTS stride = {1, 1},
INTS dilate = {1, 1}, int group = 1, INTS pads = {0, 0});
MNN_PUBLIC VARP _Conv(float weight, float bias, VARP x, INTS channel, INTS kernelSize, PaddingMode pad = VALID,
INTS stride = {1, 1}, INTS dilate = {1, 1}, int group = 1);
MNN_PUBLIC VARP _Conv(std::vector<int8_t>&& weight, std::vector<float>&& bias, VARP x, INTS channel, INTS kernelSize,
PaddingMode pad = VALID, INTS stride = {1, 1}, INTS dilate = {1, 1}, int group = 1, INTS pads = {0, 0}, bool relu = false, bool relu6 = false, int nbits = 8);
MNN_PUBLIC VARP _Conv(std::vector<float>&& weight, std::vector<float>&& bias, VARP x, INTS channel, INTS kernelSize,
PaddingMode pad = VALID, INTS stride = {1, 1}, INTS dilate = {1, 1}, int group = 1, INTS pads = {0, 0}, bool relu = false, bool relu6 = false);
MNN_PUBLIC VARP _Deconv(VARP weight, VARP bias, VARP x, PaddingMode pad = VALID, INTS stride = {1, 1},
INTS dilate = {1, 1}, int group = 1, INTS pads = {0, 0});
MNN_PUBLIC VARP _Deconv(std::vector<float>&& weight, std::vector<float>&& bias, VARP x, INTS channel, INTS kernelSize,
PaddingMode pad, INTS stride = {1, 1}, INTS dilate = {1, 1}, int group = 1, INTS pads = {0, 0}, bool relu = false, bool relu6 = false);
MNN_PUBLIC VARP _MaxPool(VARP x, INTS kernel, INTS stride = {1, 1}, PaddingMode pad = VALID, INTS pads= {0, 0});
MNN_PUBLIC VARP _AvePool(VARP x, INTS kernel, INTS stride = {1, 1}, PaddingMode pad = VALID, INTS pads= {0, 0});
MNN_PUBLIC VARP _Reshape(VARP x, INTS shape, Dimensionformat original_format = NCHW);
MNN_PUBLIC VARP _Reshape(VARP x, VARP shape);
MNN_PUBLIC VARP _Scale(VARP x, int channels, std::vector<float>&& scales, std::vector<float>&& bias);
MNN_PUBLIC VARP _Relu(VARP x, float slope = 0.0f);
MNN_PUBLIC VARP _Relu6(VARP x, float minValue = 0.0f, float maxValue = 6.0f);
MNN_PUBLIC VARP _PRelu(VARP x, std::vector<float> &&slopes);
MNN_PUBLIC VARP _Softmax(VARP logits, int axis = -1);
MNN_PUBLIC VARP _Softplus(VARP features);
MNN_PUBLIC VARP _Softsign(VARP features);
MNN_PUBLIC std::vector<VARP> _Split(VARP value, INTS size_splits, int axis = 0);
MNN_PUBLIC VARP _Slice(VARP x, VARP starts, VARP sizes);
MNN_PUBLIC VARP _StridedSlice(VARP input, VARP begin, VARP end, VARP strided,
int32_t beginMask, int32_t endMask, int32_t ellipsisMask,
int32_t newAxisMask, int32_t shrinkAxisMask);
MNN_PUBLIC VARP _Concat(VARPS values, int axis);
MNN_PUBLIC VARP _Convert(VARP input, Dimensionformat format);
MNN_PUBLIC VARP _Transpose(VARP x, INTS perm);
MNN_PUBLIC VARP _Transpose(VARP x, VARP perm);
MNN_PUBLIC VARP _ChannelShuffle(VARP x, int group);
MNN_PUBLIC VARP _ChangeInputFormat(VARP input, Dimensionformat format);
MNN_PUBLIC VARP _Conv2DBackPropFilter(VARP input, VARP inputGrad, INTS kernelSize, PaddingMode pad = VALID, INTS stride = {1, 1}, INTS dilate = {1, 1}, int group = 1, INTS pads = {0, 0});
MNN_PUBLIC VARP _PoolGrad(VARP originInput, VARP originOutput, VARP inputGrad, INTS kernel, INTS stride, PoolingMode type, PaddingMode pad = VALID, INTS pads= {0, 0});
// FIXME: move the api to Array Ops
MNN_PUBLIC VARP _ReverseSequence(VARP x, VARP y, int batchDim, int seqDim);
// FIXME: move the api to Image Ops
MNN_PUBLIC VARP _Crop(VARP images, VARP size, int axis, INTS offset);
MNN_PUBLIC VARP _Resize(VARP images, float xScale, float yScale);
MNN_PUBLIC VARP _Pad(VARP x, VARP paddings, PadValueMode mode = CONSTANT);
MNN_PUBLIC VARP _ExpandDims(VARP input, int axis);
MNN_PUBLIC VARP _ExpandDims(VARP input, VARP axis);
MNN_PUBLIC VARP _Shape(VARP input, bool nchw = false);
MNN_PUBLIC VARP _Stack(VARPS values, int axis=0);
enum InterpolationMethod {BILINEAR, NEAREST};
MNN_PUBLIC VARP _CropAndResize(VARP image, VARP boxes, VARP box_ind, VARP crop_size,
InterpolationMethod method, float extrapolation_value = 0.0);
MNN_PUBLIC VARP _Fill(VARP dims, VARP value);
MNN_PUBLIC VARP _Tile(VARP input, VARP multiples);
MNN_PUBLIC VARP _Gather(VARP params, VARP indices);
MNN_PUBLIC VARP _GatherV2(VARP params, VARP indices, VARP axis = nullptr);
MNN_PUBLIC VARP _Squeeze(VARP input, INTS axis = {});
MNN_PUBLIC VARP _Unsqueeze(VARP input, INTS axis = {});
MNN_PUBLIC VARP _BatchToSpaceND(VARP input, VARP block_shape, VARP crops);
MNN_PUBLIC VARP _GatherND(VARP params, VARP indices);
MNN_PUBLIC VARP _Selu(VARP features, float scale, float alpha);
MNN_PUBLIC VARP _Size(VARP input);
MNN_PUBLIC VARP _Elu(VARP features, float alpha=1.0);
MNN_PUBLIC VARP _Threshold(VARP features, float alpha=1.0);
MNN_PUBLIC VARP _MatrixBandPart(VARP input, VARP num_lower, VARP num_upper);
MNN_PUBLIC std::vector<VARP> _Moments(VARP x, INTS axis, VARP shift, bool keepDims);
MNN_PUBLIC VARP _SetDiff1D(VARP x, VARP y);
MNN_PUBLIC VARP _SpaceToDepth(VARP input, int block_size);
MNN_PUBLIC VARP _SpaceToBatchND(VARP input, VARP block_shape, VARP paddings);
MNN_PUBLIC VARP _ZerosLike(VARP input);
MNN_PUBLIC std::vector<VARP> _Unstack(VARP value, int axis=0);
MNN_PUBLIC VARP _Rank(VARP input);
MNN_PUBLIC VARP _Range(VARP start, VARP limit, VARP delta);
MNN_PUBLIC VARP _DepthToSpace(VARP input, int block_size);
MNN_PUBLIC VARP _PriorBox(VARP feature, VARP image,
std::vector<float> min_size, std::vector<float> max_size, std::vector<float>aspect_ratio,
bool flip, bool clip, std::vector<float>variance,
unsigned int img_h, unsigned int img_w, float step_h, float step_w, float offset = 0.5);
MNN_PUBLIC VARP _Permute(VARP input, INTS dims);
MNN_PUBLIC VARP _DetectionOutput(VARP location, VARP confidence, VARP priorbox,
unsigned int num_classes, bool share_location, int background_label_id,
float nms_threshhold, int nms_topk, int code_type,
bool variance_encoded_in_target,
int keep_top_k, float confidence_threshold, float visualize_threshold);
MNN_PUBLIC std::vector<VARP> _DetectionPostProcess(VARP encode_boxes, VARP class_predictions, VARP anchors,
int num_classes, int max_detections,
int max_class_per_detection, int detections_per_class,
float nms_threshold, float iou_threshold,
bool use_regular_nms, std::vector<float> centersize_encoding);
MNN_PUBLIC VARP _Interp(VARPS xs, float widthScale, float heightScale, int outputWidth, int outputHeight, int resizeType, bool alignCorners);
MNN_PUBLIC VARP _ZeroGrad(VARP x);
// Int8 Inference
MNN_PUBLIC VARP _Conv(std::vector<int8_t>&& weight, std::vector<int>&& bias, std::vector<float>&& scale, VARP x, INTS channel, INTS kernelSize,
PaddingMode pad, INTS stride, INTS dilate, int group, INTS pads, bool relu, int nbits = 8);
MNN_PUBLIC VARP _CosineSimilarity(VARP input0, VARP input1, VARP inputDim);
MNN_PUBLIC VARP _FloatToInt8(VARP x, VARP scale, char minValue, char maxValue);
MNN_PUBLIC VARP _Int8ToFloat(VARP x, VARP scale);
MNN_PUBLIC VARP _Select(VARP select, VARP input0, VARP input1);
} // namespace Express
} // namespace MNN
#endif /* NeuralNetWorkOp_HPP */

View File

@@ -0,0 +1,64 @@
//
// Optimizer.hpp
// MNN
//
// Created by MNN on 2019/08/20.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifndef Optimizer_hpp
#define Optimizer_hpp
#include <MNN/expr/Expr.hpp>
#include <MNN/MNNForwardType.h>
namespace MNN {
namespace Express {
class MNN_PUBLIC Optimizer {
public:
enum Device {
CPU = 0,
GPU = 1,
OTHER = 2,
AUTO = 3
};
struct Config {
Device device = CPU;
MNNForwardType forwardType = MNN_FORWARD_ALL;
int numThread = 4;
};
static std::shared_ptr<Optimizer> create(Config config);
struct Cost {
float compute; // MFlops
float memory; // MB
};
class Parameters {
public:
Parameters(int n);
virtual ~Parameters();
float* get() const {
return mValue;
}
int size() const {
return mSize;
}
private:
float* mValue;
int mSize;
};
virtual std::shared_ptr<Parameters> onGetParameters(const std::vector<VARP>& outputs) {
return nullptr;
}
//Given paramters and measure cost, the parameters must be the same as onGetParameters
virtual Cost onMeasure(const std::vector<VARP>& outputs, std::shared_ptr<Parameters> parameters = nullptr) = 0;
//Modify the output directly, the parameters must be the same as onGetParameters
virtual bool onExecute(const std::vector<VARP>& outputs, std::shared_ptr<Parameters> parameters = nullptr) = 0;
Optimizer() = default;
virtual ~Optimizer() = default;
};
} // namespace Express
} // namespace MNN
#endif