mxnet
Namespaces | Classes | Typedefs | Enumerations | Functions
mxnet Namespace Reference

namespace of mxnet More...

Namespaces

 common
 
 cpp
 
 csr
 
 engine
 namespace of engine internal types.
 
 op
 namespace of arguments
 
 rowsparse
 
 rtc
 

Classes

struct  Context
 Context information about the execution environment. More...
 
struct  DataBatch
 DataBatch of NDArray, returned by Iterator. More...
 
struct  DataInst
 a single data instance More...
 
struct  DataIteratorReg
 Registry entry for DataIterator factory functions. More...
 
class  Engine
 Dependency engine that schedules operations. More...
 
class  Executor
 Executor of a computation graph. Executor can be created by Binding a symbol. More...
 
class  IIterator
 iterator type More...
 
class  Imperative
 runtime functions for NDArray More...
 
class  KVStore
 distributed key-value store More...
 
class  NDArray
 ndarray interface More...
 
struct  NDArrayFunctionReg
 Registry entry for NDArrayFunction. More...
 
struct  OpContext
 All the possible information needed by Operator.Forward and Backward This is the superset of RunContext. We use this data structure to bookkeep everything needed by Forward and Backward. More...
 
class  Operator
 Operator interface. Operator defines basic operation unit of optimized computation graph in mxnet. This interface relies on pre-allocated memory in TBlob, the caller need to set the memory region in TBlob correctly before calling Forward and Backward. More...
 
class  OperatorProperty
 OperatorProperty is a object that stores all information about Operator. It also contains method to generate context(device) specific operators. More...
 
struct  OperatorPropertyReg
 Registry entry for OperatorProperty factory functions. More...
 
class  OpStatePtr
 Operator state. This is a pointer type, its content is mutable even if OpStatePtr is const. More...
 
struct  Resource
 Resources used by mxnet operations. A resource is something special other than NDArray, but will still participate. More...
 
class  ResourceManager
 Global resource manager. More...
 
struct  ResourceRequest
 The resources that can be requested by Operator. More...
 
struct  RunContext
 execution time context. The information needed in runtime for actual execution. More...
 
class  Storage
 Storage manager across multiple devices. More...
 
class  TBlob
 tensor blob class that can be used to hold tensor of any dimension, any device and any data type, This is a weak type that can be used to transfer data through interface TBlob itself do not involve any arithmentic operations, but it can be converted to tensor of fixed dimension for further operations More...
 

Typedefs

typedef mshadow::cpu cpu
 mxnet cpu More...
 
typedef mshadow::gpu gpu
 mxnet gpu More...
 
typedef mshadow::index_t index_t
 index type usually use unsigned More...
 
typedef mshadow::default_real_t real_t
 data type that will be used to store ndarray More...
 
using TShape = nnvm::TShape
 Shape data structure used to record shape information. More...
 
using Op = nnvm::Op
 operator structure from NNVM More...
 
using StorageTypeVector = std::vector< int >
 The result holder of storage type of each NodeEntry in the graph. More...
 
using DispatchModeVector = std::vector< DispatchMode >
 The result holder of dispatch mode of each Node in the graph. More...
 
using CachedOpPtr = std::shared_ptr< Imperative::CachedOp >
 
typedef std::function< IIterator< DataBatch > *()> DataIteratorFactory
 typedef the factory function of data iterator More...
 
typedef std::function< void(NDArray **used_vars, real_t *scalars, NDArray **mutate_vars, int num_params, char **param_keys, char **param_vals)> NDArrayAPIFunction
 definition of NDArray function More...
 
using FCreateOpState = std::function< OpStatePtr(const NodeAttrs &attrs, Context ctx, const std::vector< TShape > &in_shape, const std::vector< int > &in_type)>
 Create a Layer style, forward/backward operator. This is easy to write code that contains state. OpStatePtr is a pointer type, it's content is mutable even if OpStatePtr is constant. More...
 
using FExecType = std::function< ExecType(const NodeAttrs &attrs)>
 Execution mode of this operator. More...
 
using FStatefulCompute = std::function< void(const OpStatePtr &state, const OpContext &ctx, const std::vector< TBlob > &inputs, const std::vector< OpReqType > &req, const std::vector< TBlob > &outputs)>
 Resiger a compute function for stateful operator. OpStatePtr is a pointer type, it's content is mutable even if OpStatePtr is constant. More...
 
using FStatefulComputeEx = std::function< void(const OpStatePtr &state, const OpContext &ctx, const std::vector< NDArray > &inputs, const std::vector< OpReqType > &req, const std::vector< NDArray > &outputs)>
 Resiger a compute function for stateful operator using NDArray interface. OpStatePtr is a pointer type, it's content is mutable even if OpStatePtr is constant. More...
 
using FResourceRequest = std::function< std::vector< ResourceRequest >(const NodeAttrs &n)>
 The resource request from the operator. More...
 
using FNDArrayFunction = std::function< void(const nnvm::NodeAttrs &attrs, const std::vector< NDArray > &inputs, std::vector< NDArray > *outputs)>
 Register an operator called as a NDArray function. More...
 
using FCompute = std::function< void(const nnvm::NodeAttrs &attrs, const OpContext &ctx, const std::vector< TBlob > &inputs, const std::vector< OpReqType > &req, const std::vector< TBlob > &outputs)>
 Resiger a compute function for simple stateless forward only operator. More...
 
using FComputeEx = std::function< void(const nnvm::NodeAttrs &attrs, const OpContext &ctx, const std::vector< NDArray > &inputs, const std::vector< OpReqType > &req, const std::vector< NDArray > &outputs)>
 Resiger an NDArray compute function for simple stateless forward only operator. More...
 
using FInferStorageType = std::function< bool(const NodeAttrs &attrs, const int dev_mask, DispatchMode *dispatch_mode, std::vector< int > *in_attrs, std::vector< int > *out_attrs)>
 Resiger a storage and dispatch mode inference function based on storage types of the inputs and outputs, and the dev_mask for the operator. More...
 
typedef std::function< OperatorProperty *()> OperatorPropertyFactory
 typedef the factory function of operator property More...
 

Enumerations

enum  FnProperty {
  FnProperty::kNormal, FnProperty::kCopyFromGPU, FnProperty::kCopyToGPU, FnProperty::kCPUPrioritized,
  FnProperty::kAsync, FnProperty::kDeleteVar
}
 Function property, used to hint what action is pushed to engine. More...
 
enum  NDArrayStorageType { kUndefinedStorage = -1, kDefaultStorage, kRowSparseStorage, kCSRStorage }
 
enum  NDArrayFunctionTypeMask { kNDArrayArgBeforeScalar = 1, kScalarArgBeforeNDArray = 1 << 1, kAcceptEmptyMutateTarget = 1 << 2 }
 mask information on how functions can be exposed More...
 
enum  OpReqType { kNullOp, kWriteTo, kWriteInplace, kAddTo }
 operation request type to Forward and Backward More...
 
enum  ExecType { ExecType::kSync, ExecType::kAsync, ExecType::kLocal, ExecType::kCrossDeviceCopy }
 the execution type of the operator More...
 
enum  DispatchMode {
  DispatchMode::kUndefined = -1, DispatchMode::kFCompute, DispatchMode::kFComputeEx, DispatchMode::kFComputeFallback,
  DispatchMode::kVariable
}
 the dispatch mode of the operator More...
 

Functions

size_t num_aux_data (NDArrayStorageType stype)
 
void CopyFromTo (const NDArray &from, NDArray *to, int priority=0)
 issue an copy operation from one NDArray to another the two ndarray can sit on different devices this operation will be scheduled by the engine More...
 
void CopyFromTo (const NDArray &from, const NDArray &to, int priority=0)
 issue an copy operation from one NDArray to another the two ndarray can sit on different devices this operation will be scheduled by the engine More...
 
void ElementwiseSum (const std::vector< NDArray > &source, NDArray *out, int priority=0)
 Perform elementwise sum over each data from source, store result into out. More...
 
NDArray operator+ (const NDArray &lhs, const NDArray &rhs)
 elementwise add More...
 
NDArray operator+ (const NDArray &lhs, const real_t &rhs)
 elementwise add More...
 
NDArray operator- (const NDArray &lhs, const NDArray &rhs)
 elementwise subtraction More...
 
NDArray operator- (const NDArray &lhs, const real_t &rhs)
 elementwise subtraction More...
 
NDArray operator* (const NDArray &lhs, const NDArray &rhs)
 elementwise multiplication More...
 
NDArray operator* (const NDArray &lhs, const real_t &rhs)
 elementwise multiplication More...
 
NDArray operator/ (const NDArray &lhs, const NDArray &rhs)
 elementwise division More...
 
NDArray operator/ (const NDArray &lhs, const real_t &rhs)
 elementwise division More...
 
void RandomSeed (uint32_t seed)
 Seed the random number generator. More...
 
void SampleUniform (real_t begin, real_t end, NDArray *out)
 Sample uniform distribution for each elements of out. More...
 
void SampleGaussian (real_t mu, real_t sigma, NDArray *out)
 Sample gaussian distribution for each elements of out. More...
 
void SampleGamma (real_t alpha, real_t beta, NDArray *out)
 Sample gamma distribution for each elements of out. More...
 
void SampleExponential (real_t lambda, NDArray *out)
 Sample exponential distribution for each elements of out. More...
 
void SamplePoisson (real_t lambda, NDArray *out)
 Sample Poisson distribution for each elements of out. More...
 
void SampleNegBinomial (int32_t k, real_t p, NDArray *out)
 Sample negative binomial distribution for each elements of out. More...
 
void SampleGenNegBinomial (real_t mu, real_t alpha, NDArray *out)
 Sample generalized negative binomial distribution for each elements of out. More...
 

Detailed Description

namespace of mxnet

Macros/inlines to assist CLion to parse Cuda files (*.cu, *.cuh)

Typedef Documentation

using mxnet::CachedOpPtr = typedef std::shared_ptr<Imperative::CachedOp>
typedef mshadow::cpu mxnet::cpu

mxnet cpu

typedef std::function<IIterator<DataBatch> *()> mxnet::DataIteratorFactory

typedef the factory function of data iterator

using mxnet::DispatchModeVector = typedef std::vector<DispatchMode>

The result holder of dispatch mode of each Node in the graph.

  • *+ *
    Note
    Stored under graph.attrs["dispatch_mode"], provided by Pass "InferStorageType"
  • *
  • *
    + * Graph g = ApplyPass(src_graph, "InferStorageType");
    + * const DispatchModeVector& dispatch_modes = g.GetAttr<DispatchModeVector>("dispatch_mode");
    + * // get dispatch mode by entry node id
    + * int node_type = dispatch_modes[nid];
    + *
  • *
  • *
    See also
    FInferStorageType
using mxnet::FCompute = typedef std::function<void (const nnvm::NodeAttrs& attrs, const OpContext& ctx, const std::vector<TBlob>& inputs, const std::vector<OpReqType>& req, const std::vector<TBlob>& outputs)>

Resiger a compute function for simple stateless forward only operator.

Note
Register under "FCompute<cpu>" and "FCompute<gpu>"
using mxnet::FComputeEx = typedef std::function<void (const nnvm::NodeAttrs& attrs, const OpContext& ctx, const std::vector<NDArray>& inputs, const std::vector<OpReqType>& req, const std::vector<NDArray>& outputs)>

Resiger an NDArray compute function for simple stateless forward only operator.

Note
Register under "FComputeEx<xpu>" and "FComputeEx<xpu>" Dispatched only when inferred dispatch_mode is FDispatchComputeEx
using mxnet::FCreateOpState = typedef std::function<OpStatePtr (const NodeAttrs& attrs, Context ctx, const std::vector<TShape>& in_shape, const std::vector<int>& in_type)>

Create a Layer style, forward/backward operator. This is easy to write code that contains state. OpStatePtr is a pointer type, it's content is mutable even if OpStatePtr is constant.

This is not the only way to register an op execution function. More simpler or specialized operator form can be registered

Note
Register under "FCreateLayerOp"
using mxnet::FExecType = typedef std::function<ExecType (const NodeAttrs& attrs)>

Execution mode of this operator.

using mxnet::FInferStorageType = typedef std::function<bool (const NodeAttrs& attrs, const int dev_mask, DispatchMode* dispatch_mode, std::vector<int>* in_attrs, std::vector<int>* out_attrs)>

Resiger a storage and dispatch mode inference function based on storage types of the inputs and outputs, and the dev_mask for the operator.

Note
Register under "FInferStorageType"
using mxnet::FNDArrayFunction = typedef std::function<void (const nnvm::NodeAttrs& attrs, const std::vector<NDArray>& inputs, std::vector<NDArray>* outputs)>

Register an operator called as a NDArray function.

Note
Register under "FNDArrayFunction"
using mxnet::FResourceRequest = typedef std::function< std::vector<ResourceRequest> (const NodeAttrs& n)>

The resource request from the operator.

Note
Register under "FResourceRequest"
using mxnet::FStatefulCompute = typedef std::function<void (const OpStatePtr& state, const OpContext& ctx, const std::vector<TBlob>& inputs, const std::vector<OpReqType>& req, const std::vector<TBlob>& outputs)>

Resiger a compute function for stateful operator. OpStatePtr is a pointer type, it's content is mutable even if OpStatePtr is constant.

Note
Register under "FStatefulCompute<cpu>" and "FStatefulCompute<gpu>"
using mxnet::FStatefulComputeEx = typedef std::function<void (const OpStatePtr& state, const OpContext& ctx, const std::vector<NDArray>& inputs, const std::vector<OpReqType>& req, const std::vector<NDArray>& outputs)>

Resiger a compute function for stateful operator using NDArray interface. OpStatePtr is a pointer type, it's content is mutable even if OpStatePtr is constant.

Note
Register under "FStatefulComputeEx<cpu>" and "FStatefulComputeEx<gpu>"
typedef mshadow::gpu mxnet::gpu

mxnet gpu

typedef mshadow::index_t mxnet::index_t

index type usually use unsigned

typedef std::function<void (NDArray **used_vars, real_t *scalars, NDArray **mutate_vars, int num_params, char **param_keys, char **param_vals)> mxnet::NDArrayAPIFunction

definition of NDArray function

using mxnet::Op = typedef nnvm::Op

operator structure from NNVM

typedef the factory function of operator property

typedef mshadow::default_real_t mxnet::real_t

data type that will be used to store ndarray

using mxnet::StorageTypeVector = typedef std::vector<int>

The result holder of storage type of each NodeEntry in the graph.

Note
Stored under graph.attrs["storage_type"], provided by Pass "InferStorageType"
Graph g = ApplyPass(src_graph, "InferStorageType");
const StorageVector& stypes = g.GetAttr<StorageTypeVector>("storage_type");
// get storage type by entry id
int entry_type = stypes[g.indexed_graph().entry_id(my_entry)];
See also
FInferStorageType
using mxnet::TShape = typedef nnvm::TShape

Shape data structure used to record shape information.

Enumeration Type Documentation

enum mxnet::DispatchMode
strong

the dispatch mode of the operator

Enumerator
kUndefined 
kFCompute 
kFComputeEx 
kFComputeFallback 
kVariable 
enum mxnet::ExecType
strong

the execution type of the operator

Enumerator
kSync 

Forward/Backward are synchronize calls.

kAsync 

Forward/Backward are asynchronize, will call OpContext.async_on_complete when operation finishes.

kLocal 

Run this operator on the scheduling thread without pushing to engine.

kCrossDeviceCopy 

Cross device copy operation, this is a special operator That indicates copy across devices, the input and output can sit on different device. In current implementation, copy operator is specially handled by executor. This flag is used for special case treatment and future extension of different copy ops.

enum mxnet::FnProperty
strong

Function property, used to hint what action is pushed to engine.

Enumerator
kNormal 

Normal operation.

kCopyFromGPU 

Copy operation from GPU to other devices.

kCopyToGPU 

Copy operation from CPU to other devices.

kCPUPrioritized 

Prioritized sync operation on CPU.

kAsync 

Asynchronous function call.

kDeleteVar 

Delete variable call.

mask information on how functions can be exposed

Enumerator
kNDArrayArgBeforeScalar 

all the use_vars should go before scalar

kScalarArgBeforeNDArray 

all the scalar should go before use_vars

kAcceptEmptyMutateTarget 

whether this function allows the handles in the target to be empty NDArray that are not yet initialized, and will initialize them when the function is invoked.

most function should support this, except copy between different devices, which requires the NDArray to be pre-initialized with context

Enumerator
kUndefinedStorage 
kDefaultStorage 
kRowSparseStorage 
kCSRStorage 

operation request type to Forward and Backward

Enumerator
kNullOp 

no operation, do not write anything

kWriteTo 

write gradient to provided space

kWriteInplace 

perform an inplace write, Target shares memory with one of input arguments. This option only happen when

kAddTo 

add to the provided space

Function Documentation

void mxnet::CopyFromTo ( const NDArray from,
NDArray to,
int  priority = 0 
)

issue an copy operation from one NDArray to another the two ndarray can sit on different devices this operation will be scheduled by the engine

Parameters
fromthe ndarray we want to copy data from
tothe target ndarray
priorityPriority of the action.
Note
The function name explicitly marks the order of from and to due to different possible convention carried by copy function.
void mxnet::CopyFromTo ( const NDArray from,
const NDArray to,
int  priority = 0 
)

issue an copy operation from one NDArray to another the two ndarray can sit on different devices this operation will be scheduled by the engine

Parameters
fromthe ndarray we want to copy data from
tothe target ndarray
priorityPriority of the action.
Note
The function name explicitly marks the order of from and to due to different possible convention carried by copy function.
void mxnet::ElementwiseSum ( const std::vector< NDArray > &  source,
NDArray out,
int  priority = 0 
)

Perform elementwise sum over each data from source, store result into out.

Parameters
sourcethe ndarray we want to sum
outthe target ndarray
priorityPriority of the action.
size_t mxnet::num_aux_data ( NDArrayStorageType  stype)
Returns
the number of aux data used for given storage type
NDArray mxnet::operator* ( const NDArray lhs,
const NDArray rhs 
)

elementwise multiplication

Parameters
lhsleft operand
rhsright operand
Returns
a new result ndarray
NDArray mxnet::operator* ( const NDArray lhs,
const real_t rhs 
)

elementwise multiplication

Parameters
lhsleft operand
rhsright operand
Returns
a new result ndarray
NDArray mxnet::operator+ ( const NDArray lhs,
const NDArray rhs 
)

elementwise add

Parameters
lhsleft operand
rhsright operand
Returns
a new result ndarray
NDArray mxnet::operator+ ( const NDArray lhs,
const real_t rhs 
)

elementwise add

Parameters
lhsleft operand
rhsright operand
Returns
a new result ndarray
NDArray mxnet::operator- ( const NDArray lhs,
const NDArray rhs 
)

elementwise subtraction

Parameters
lhsleft operand
rhsright operand
Returns
a new result ndarray
NDArray mxnet::operator- ( const NDArray lhs,
const real_t rhs 
)

elementwise subtraction

Parameters
lhsleft operand
rhsright operand
Returns
a new result ndarray
NDArray mxnet::operator/ ( const NDArray lhs,
const NDArray rhs 
)

elementwise division

Parameters
lhsleft operand
rhsright operand
Returns
a new result ndarray
NDArray mxnet::operator/ ( const NDArray lhs,
const real_t rhs 
)

elementwise division

Parameters
lhsleft operand
rhsright operand
Returns
a new result ndarray
void mxnet::RandomSeed ( uint32_t  seed)

Seed the random number generator.

Parameters
seedthe seed to set to global random number generators.
void mxnet::SampleExponential ( real_t  lambda,
NDArray out 
)

Sample exponential distribution for each elements of out.

Parameters
lambdaparameter (rate) of the exponential distribution
outoutput NDArray.
void mxnet::SampleGamma ( real_t  alpha,
real_t  beta,
NDArray out 
)

Sample gamma distribution for each elements of out.

Parameters
alphaparameter (shape) of the gamma distribution
betaparameter (scale) of the gamma distribution
outoutput NDArray.
void mxnet::SampleGaussian ( real_t  mu,
real_t  sigma,
NDArray out 
)

Sample gaussian distribution for each elements of out.

Parameters
mumean of gaussian distribution.
sigmastandard deviation of gaussian distribution.
outoutput NDArray.
void mxnet::SampleGenNegBinomial ( real_t  mu,
real_t  alpha,
NDArray out 
)

Sample generalized negative binomial distribution for each elements of out.

Parameters
muparameter (mean) of the distribution
alphaparameter (over dispersion) of the distribution
outoutput NDArray.
void mxnet::SampleNegBinomial ( int32_t  k,
real_t  p,
NDArray out 
)

Sample negative binomial distribution for each elements of out.

Parameters
kfailure limit
psuccess probability
outoutput NDArray.
void mxnet::SamplePoisson ( real_t  lambda,
NDArray out 
)

Sample Poisson distribution for each elements of out.

Parameters
lambdaparameter (rate) of the Poisson distribution
outoutput NDArray.
void mxnet::SampleUniform ( real_t  begin,
real_t  end,
NDArray out 
)

Sample uniform distribution for each elements of out.

Parameters
beginlower bound of distribution.
endupper bound of distribution.
outoutput NDArray.