Skip to content

SeetaResearch / Dragon

Go to a project
Commit b7365da6 by Ting PAN
Options

Dragon 0.2.1 preview

1 parent 6eeac5fe
Expand all
Inline Side-by-side
Showing with 2676 additions and 394 deletions
Dragon/CMakeLists.txt
# ---------------- Welcom To Use Dragon ----------------
PROJECT(dragon)
CMAKE_MINIMUM_REQUIRED(VERSION 2.8.0)
CMAKE_MINIMUM_REQUIRED(VERSION 3.0.0)
# ---------------- Welcom To Use Dragon ----------------
......@@ -12,7 +12,7 @@ option(WITH_PYTHON3 "Set ON to use PYTHON3 otherwise PYTHON2" OF
option(WITH_CUDA "Set ON to use CUDA" ON)
option(WITH_CUDNN "Set ON to use CUDNN" OFF)
option(WITH_BLAS "Set ON to use BLAS" OFF)
option(WITH_OMP "Set ON to use OpenMP" OFF)
option(WITH_OMP "Set ON to use OpenMP" ON)
option(WITH_SSE "Set ON to use SSE 4.1" ON)
option(WITH_MPI "Set ON to use MPI" OFF)
option(WITH_MPI_CUDA "Set ON to use MPI-CUDA" OFF)
......@@ -23,14 +23,13 @@ option(WITH_CUDA_FP16 "Set ON to use FP16" ON)
set(3RDPARTY_DIR ${PROJECT_SOURCE_DIR}/../3rdparty)
# set your python environment
set(PYTHON_DIR /usr/include/python2.7) # preferred
#set(PYTHON_DIR /usr/include/python3.x) # optional, set specific version
#set(ANACONDA_DIR /xxx/anaconda) # optional, root folder of anaconda, preset for 2.7, 3.5, and 3.6
set(NUMPY_DIR /xxx/numpy) # required, root folder of numpy package
set(PYTHON_INCLUDE_DIR /usr/include/python2.7) # preferred
#set(PYTHON_INCLUDE_DIR /usr/include/python3.x) # optional, set specific version
#set(ANACONDA_ROOT_DIR /xxx/anaconda) # optional, preset for 2.7, 3.5, and 3.6
set(NUMPY_ROOT_DIR /xxx/numpy) # required
# set CUDA compiling architecture
set(CUDA_ARCH -gencode arch=compute_20,code=sm_20
-gencode arch=compute_30,code=sm_30
set(CUDA_ARCH -gencode arch=compute_30,code=sm_30
-gencode arch=compute_35,code=sm_35
-gencode arch=compute_50,code=sm_50
-gencode arch=compute_60,code=sm_60)
......@@ -81,19 +80,20 @@ include_directories(${3RDPARTY_DIR}/include)
include_directories(${3RDPARTY_DIR}/include/mpi)
include_directories(${CUDA_INCLUDE_DIRS})
include_directories(${PROJECT_SOURCE_DIR}/src)
include_directories(${NUMPY_DIR}/core/include)
include_directories(${NUMPY_DIR})
include_directories(${NUMPY_DIR}/numpy)
include_directories(${ANACONDA_DIR}/include/python2.7)
include_directories(${ANACONDA_DIR}/include/python3.5)
include_directories(${ANACONDA_DIR}/include/python3.6)
include_directories(${PYTHON_DIR})
include_directories(${ANACONDA_DIR}/include)
include_directories(${PYTHON_INCLUDE_DIR})
include_directories(${ANACONDA_ROOT_DIR}/include)
include_directories(${ANACONDA_ROOT_DIR}/include/python2.7)
include_directories(${ANACONDA_ROOT_DIR}/include/python3.5)
include_directories(${ANACONDA_ROOT_DIR}/include/python3.6)
include_directories(${NUMPY_ROOT_DIR}/core/include)
include_directories(${NUMPY_ROOT_DIR}/include)
include_directories(${NUMPY_ROOT_DIR})
# ---[ libs
set(3RDPARTY_LIBS ${3RDPARTY_DIR}/lib)
set(UINX_CUDNN_LIBS /usr/local/cuda/lib64)
link_directories(${3RDPARTY_LIBS})
link_directories(/usr/local/cuda/lib64)
link_directories(${UINX_CUDNN_LIBS})
# ---[ Install
set(CMAKE_INSTALL_PREFIX ${PROJECT_SOURCE_DIR} CACHE STRING "set install prefix" FORCE)
......@@ -166,6 +166,12 @@ endif()
# ---[ Warnings
# ---[ execute
set (PROTOS_DIR ${PROJECT_SOURCE_DIR}/src/protos)
message(STATUS "Generate Protobuf Files")
execute_process(COMMAND protoc -I=${PROTOS_DIR} --cpp_out=${PROTOS_DIR} ${PROTOS_DIR}/caffemodel.proto)
execute_process(COMMAND protoc -I=${PROTOS_DIR} --cpp_out=${PROTOS_DIR} ${PROTOS_DIR}/dragon.proto)
# ---[ Subdirectories
add_subdirectory(modules/python)
......
Dragon/include/core/common.h
......@@ -45,7 +45,7 @@ using Set = std::unordered_set<Value> ;
#define CONCATENATE_IMPL(s1, s2) s1##s2
#define CONCATENATE(s1, s2) CONCATENATE_IMPL(s1,s2)
#define ANONYMOUS_VARIABLE(str) CONCATENATE(str, __LINE__)
#define NOT_IMPLEMENTED LOG(FATAL) << "this moudle is not implemented"
#define NOT_IMPLEMENTED LOG(FATAL) << "This module has not been implemented yet."
} // namespace dragon
......
Dragon/include/core/context.h
......@@ -42,7 +42,7 @@ class CPUContext{
#else
data = malloc(nbytes);
#endif
CHECK(data) << "malloc mem: " << nbytes << " bytes failed.";
CHECK(data) << "Malloc mem: " << nbytes << " bytes failed.";
return data;
}
......
Dragon/include/core/context_cuda.h
......@@ -91,13 +91,13 @@ class CUDAContext {
cudaStreamSynchronize(cudaStreamDefault);
cudaError_t error = cudaGetLastError();
CHECK_EQ(error, cudaSuccess)
<< "cuda error: " << cudaGetErrorString(error);
<< "CUDA Error: " << cudaGetErrorString(error);
}
inline static void* New(size_t nbytes) {
void* data;
cudaMalloc(&data, nbytes);
CHECK(data) << "malloc cuda mem: " << nbytes << " bytes failed.";
CHECK(data) << "Malloc cuda mem: " << nbytes << " bytes failed.";
return data;
}
......@@ -190,12 +190,12 @@ static inline cudnnHandle_t& cudnn_handle() {
#else // WITH_CUDA
class CUDAContext{
public:
CUDAContext(const DeviceOption& option) { LOG(FATAL) << "CUDA is not compilied."; }
CUDAContext(const int gpu_id = 0) { LOG(FATAL) << "CUDA is not compilied."; }
CUDAContext(const DeviceOption& option) { LOG(FATAL) << "CUDA was not compiled."; }
CUDAContext(const int gpu_id = 0) { LOG(FATAL) << "CUDA was not compiled."; }
template<class DstContext, class SrcContext>
static void Memcpy(size_t nbytes, void* dst, const void* src) {
LOG(FATAL) << "CUDA is not compilied.";
LOG(FATAL) << "CUDA was not compiled.";
}
};
#endif // WITH_CUDA
......
Dragon/include/core/operator.h
......@@ -61,12 +61,17 @@ class OperatorBase{
inline const Map<std::string, const Argument*>& args() { return args_; }
inline const Argument& arg(const string& name) { return *(args_[name]); }
typedef Map<string, vector<OperatorBase*> > RecomputeMap;
inline RecomputeMap& recompute_map() { return recompute_map_; }
void set_recompute_map(RecomputeMap recompute_map) { recompute_map_ = recompute_map; }
inline const OperatorDef& op_def() const { return op_def_; }
inline const string debug_string() const { return op_def_.DebugString(); }
protected:
string phase_;
Map<std::string, const Argument*> args_;
Map<string, vector<OperatorBase*> > recompute_map_;
vector<Tensor*> inputs_, outputs_;
OperatorDef op_def_;
Workspace* ws_;
......@@ -158,10 +163,10 @@ DECLARE_REGISTRY(CUDNNOperatorRegistry, OperatorBase, const OperatorDef&, Worksp
TIndex count = 1; \
for(int i = 0; i < shape.size(); i++) count *= shape[i]; \
CHECK_EQ(count, tensor.count()) \
<< "\nmodel request " << "Tensor(" << tensor.name() << ")'s " \
<< "size is " << count << "\n" \
<< "but now is " << tensor.count() << "\n" \
<< "may be feed the incorrect Tensor before ?"; \
<< "\nModel request " << "Tensor(" << tensor.name() << ")'s " \
<< "size is " << count << ", \n" \
<< "but now is " << tensor.count() << ", " \
<< "did you feed the incorrect Tensor before ?"; \
tensor.Reshape(shape); \
}
......
Dragon/include/core/operator_schema.h
......@@ -23,6 +23,8 @@ class OpSchema{
bool Verify(const OperatorDef& def) const;
inline OpSchema& IgnoreVerify() { ignore_verify_ = true; return *this; }
OpSchema& Inplace(set<pair<int, int> > inplace);
std::function<bool(int, int)> CheckInplace;
inline bool AllowInplace() const { return allow_inplace_; }
......@@ -38,14 +40,13 @@ class OpSchema{
min_input_ = min_output_= 0;
max_input_ = max_output_ = std::numeric_limits<int>::max();
CheckInplace = [](int, int) { return false; };
allow_inplace_ = false;
ignore_verify_ = allow_inplace_ = false;
}
string op_type_, file_;
int line_, min_input_, max_input_;
int min_output_, max_output_;
bool allow_inplace_;
bool allow_inplace_, ignore_verify_;
};
class OpSchemaRegistry {
......
Dragon/include/core/tensor.h
......@@ -30,10 +30,19 @@ class Tensor {
CHECK_GT(d, 0);
new_size *= d;
}
if (size_ != new_size && own_mem_ &&
capacity_ < TIndex(new_size * meta_.itemsize())) {
memory_.reset();
capacity_ = 0;
if (own_mem_) {
if (size_ != new_size &&
capacity_ < TIndex(new_size * meta_.itemsize())) {
memory_.reset();
capacity_ = 0;
}
} else {
if (ex_memory_ && TIndex(ex_memory_->nbytes()) <
TIndex(new_size * meta_.itemsize())) {
delete ex_memory_;
ex_memory_ = nullptr;
capacity_ = 0;
}
}
size_ = new_size;
}
......@@ -96,7 +105,7 @@ class Tensor {
MixedMemory* memory() const { return own_mem_ ? memory_.get() : ex_memory_; }
MixedMemory::State memory_state() const {
MixedMemory* mem = memory();
CHECK(mem) << "memory access before allowcating.";
CHECK(mem) << "Memory access before allowcating.";
return memory()->state();
}
......@@ -120,7 +129,7 @@ class Tensor {
} else if (TypeMeta::Id<Context>() == TypeMeta::Id<CUDAContext>()) {
*data_ptr = mem->mutable_cuda_data();
} else {
LOG(FATAL) << "unknown memory type access. only CPU or CUDA are supported.";
LOG(FATAL) << "Unknown memory type. Only CPU or CUDA is supported.";
}
}
}
......@@ -142,34 +151,16 @@ class Tensor {
template <class Context>
void* raw_mutable_data(const TypeMeta& meta) {
void* data_ptr;
if (own_mem_) {
mutable_data_ptr<Context>(&data_ptr);
if (meta_ == meta && data_ptr) {
return data_ptr;
} else {
meta_ = meta;
CHECK_GT(size_, 0);
memory_.reset(new MixedMemory(meta, size_* meta_.itemsize()));
mutable_data_ptr<Context>(&data_ptr); // malloc
if (meta.ctor()) meta_.ctor()(data_ptr, size_);
}
capacity_ = size_ * meta_.itemsize();
return data_ptr;
} else {
meta_ = meta;
CHECK_GT(size_, 0);
TIndex ex_capacity_ = ex_memory_->nbytes();
if (ex_capacity_ >= TIndex(size_ * meta.itemsize())) {
mutable_data_ptr<Context>(&data_ptr);
} else {
delete ex_memory_;
ex_memory_ = new MixedMemory(meta, size_* meta_.itemsize());
mutable_data_ptr<Context>(&data_ptr); // malloc
if (meta.ctor()) meta_.ctor()(data_ptr, size_);
capacity_ = size_ * meta.itemsize();
}
return data_ptr;
}
mutable_data_ptr<Context>(&data_ptr);
if (meta_ == meta && data_ptr) return data_ptr;
meta_ = meta;
CHECK_GT(size_, 0);
if (own_mem_) memory_.reset(new MixedMemory(meta, size_* meta_.itemsize()));
else ex_memory_ = new MixedMemory(meta, size_* meta_.itemsize());
mutable_data_ptr<Context>(&data_ptr); // malloc memory
if (meta.ctor()) meta_.ctor()(data_ptr, size_); // call the constructor
capacity_ = size_ * meta.itemsize();
return data_ptr;
}
template <class Context>
......@@ -181,7 +172,9 @@ class Tensor {
}
template <class Context>
const void* raw_data() const { return const_data_ptr<Context>(); }
const void* raw_data() const {
return const_data_ptr<Context>();
}
template <typename T, class Context>
T* mutable_data() {
......@@ -192,8 +185,8 @@ class Tensor {
}
template <typename T, class Context>
const T* data() const {
return static_cast<const T*>(raw_data<Context>());
const T* data() const {
return static_cast<const T*>(raw_data<Context>());
}
inline void Share(const Tensor& other) {
......
Dragon/include/core/workspace.h
......@@ -25,10 +25,8 @@ class Workspace{
typedef Map<string, unique_ptr<GraphBase> > GraphMap;
typedef Map<string, TensorFiller> FillerMap;
typedef Map<string, string> RenameMap;
typedef Map<string, vector<OperatorBase*> > RecomputeMap;
Workspace(): root_folder_(".") { init(); }
Workspace(string root_folder) : root_folder_(root_folder) { init(); }
Workspace() { init(); }
~Workspace();
void init() {
......@@ -122,7 +120,7 @@ class Workspace{
buffer_map_[category].pop();
return GetTensor(name);
}
LOG(FATAL) << "buffers of [" << category << "] "
LOG(FATAL) << "Buffers of [" << category << "] "
<< "are not enough, add more if necessary.";
return nullptr;
}
......@@ -161,38 +159,19 @@ class Workspace{
}
/******************** Utility ********************/
inline const string& GetRootFolder() const { return root_folder_; }
inline void CreateRename(const string& old_tensor,
const string& new_tensor) {
rename_map_[old_tensor] = new_tensor;
}
inline void AddRecompute(const string& tensor, OperatorBase* op) {
if (!recompute_map_.count(tensor)) {
recompute_map_[tensor] = vector<OperatorBase*>();
}
recompute_map_[tensor].push_back(op);
}
inline vector<OperatorBase*> GetRecompute(const string& tensor) {
if (recompute_map_.count(tensor)) {
return recompute_map_[tensor];
} else {
return vector<OperatorBase*>();
}
}
private:
TensorMap tensor_map_;
BufferMap buffer_map_;
LockMap lock_map_;
GraphMap graph_map_;
FillerMap filler_map_;
RenameMap rename_map_;
RecomputeMap recompute_map_;
string root_folder_;
RenameMap rename_map_;
};
} // namespace dragon
......
Dragon/include/operators/activation/dropout_op.h
......@@ -17,7 +17,7 @@ class DropoutOp final : public Operator<Context> {
public:
DropoutOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
prob(OperatorBase::GetSingleArg<float>("prob", 0)) {
prob(OperatorBase::GetSingleArg<float>("prob", 0.5)) {
bool use_scale = OperatorBase::GetSingleArg<bool>("scale", true);
threshold = static_cast<unsigned int>(UINT_MAX * prob);
if (use_scale) scale = 1.0 / (1.0 - prob);
......@@ -38,7 +38,7 @@ class DropoutGradientOp final : public Operator<Context> {
public:
DropoutGradientOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
prob(OperatorBase::GetSingleArg<float>("prob", 0)) {
prob(OperatorBase::GetSingleArg<float>("prob", 0.5)) {
bool use_scale = OperatorBase::GetSingleArg<bool>("scale", true);
threshold = static_cast<unsigned int>(UINT_MAX * prob);
if (use_scale) scale = 1.0 / (1.0 - prob);
......
Dragon/include/operators/activation/elu_op.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_ACTIVATION_ELU_OP_H_
#define DRAGON_OPERATORS_ACTIVATION_ELU_OP_H_
#include "core/operator.h"
namespace dragon {
template <class Context>
class EluOp : public Operator<Context> {
public:
EluOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
alpha(OperatorBase::GetSingleArg<float>("alpha", 1.0)) {}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
float alpha;
};
template <class Context>
class EluGradientOp : public Operator<Context> {
public:
EluGradientOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
alpha(OperatorBase::GetSingleArg<float>("alpha", 1.0)) {
DISABLE_SHARE_GRADIENT;
}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
float alpha;
};
#ifdef WITH_CUDNN
#if CUDNN_VERSION_MIN(6, 0, 0)
template <class Context>
class CuDNNEluOp final : public EluOp<Context> {
public:
CuDNNEluOp(const OperatorDef& op_def, Workspace* ws)
: EluOp<Context>(op_def, ws) {
CUDNN_CHECK(cudnnCreateTensorDescriptor(&input_desc));
CUDNN_CHECK(cudnnCreateTensorDescriptor(&output_desc));
CUDNN_CHECK(cudnnCreateActivationDescriptor(&act_desc));
CUDNN_CHECK(cudnnSetActivationDescriptor(act_desc,
CUDNN_ACTIVATION_ELU, CUDNN_PROPAGATE_NAN, this->alpha));
}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
cudnnTensorDescriptor_t input_desc, output_desc;
cudnnActivationDescriptor_t act_desc;
};
template <class Context>
class CuDNNEluGradientOp final : public EluGradientOp<Context> {
public:
CuDNNEluGradientOp(const OperatorDef& op_def, Workspace* ws)
: EluGradientOp<Context>(op_def, ws) {
CUDNN_CHECK(cudnnCreateTensorDescriptor(&input_desc));
CUDNN_CHECK(cudnnCreateTensorDescriptor(&output_desc));
CUDNN_CHECK(cudnnCreateActivationDescriptor(&act_desc));
CUDNN_CHECK(cudnnSetActivationDescriptor(act_desc,
CUDNN_ACTIVATION_ELU, CUDNN_PROPAGATE_NAN, this->alpha));
}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
cudnnTensorDescriptor_t input_desc, output_desc;
cudnnActivationDescriptor_t act_desc;
};
#endif
#endif // WITH_CUDNN
} // namespace dragon
#endif // DRAGON_OPERATORS_ACTIVATION_ELU_OP_H_
\ No newline at end of file
Dragon/include/operators/activation/sigmoid_op.h
......@@ -12,7 +12,7 @@
namespace dragon {
template <class Context>
class SigmoidOp final : public Operator<Context> {
class SigmoidOp : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(SigmoidOp);
......@@ -21,7 +21,7 @@ class SigmoidOp final : public Operator<Context> {
};
template <class Context>
class SigmoidGradientOp final : public Operator<Context> {
class SigmoidGradientOp : public Operator<Context> {
public:
SigmoidGradientOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws) {
......@@ -32,6 +32,48 @@ class SigmoidGradientOp final : public Operator<Context> {
template <typename T> void RunWithType();
};
#ifdef WITH_CUDNN
template <class Context>
class CuDNNSigmoidOp final : public SigmoidOp<Context> {
public:
CuDNNSigmoidOp(const OperatorDef& op_def, Workspace* ws)
: SigmoidOp<Context>(op_def, ws) {
CUDNN_CHECK(cudnnCreateTensorDescriptor(&input_desc));
CUDNN_CHECK(cudnnCreateTensorDescriptor(&output_desc));
CUDNN_CHECK(cudnnCreateActivationDescriptor(&act_desc));
CUDNN_CHECK(cudnnSetActivationDescriptor(act_desc,
CUDNN_ACTIVATION_SIGMOID, CUDNN_PROPAGATE_NAN, 0));
}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
cudnnTensorDescriptor_t input_desc, output_desc;
cudnnActivationDescriptor_t act_desc;
};
template <class Context>
class CuDNNSigmoidGradientOp final : public SigmoidGradientOp<Context> {
public:
CuDNNSigmoidGradientOp(const OperatorDef& op_def, Workspace* ws)
: SigmoidGradientOp<Context>(op_def, ws) {
CUDNN_CHECK(cudnnCreateTensorDescriptor(&input_desc));
CUDNN_CHECK(cudnnCreateTensorDescriptor(&output_desc));
CUDNN_CHECK(cudnnCreateActivationDescriptor(&act_desc));
CUDNN_CHECK(cudnnSetActivationDescriptor(act_desc,
CUDNN_ACTIVATION_SIGMOID, CUDNN_PROPAGATE_NAN, 0));
}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
cudnnTensorDescriptor_t input_desc, output_desc;
cudnnActivationDescriptor_t act_desc;
};
#endif // WITH_CUDNN
} // namespace dragon
#endif // DRAGON_OPERATORS_ACTIVATION_SIGMOID_OP_HPP
\ No newline at end of file
Dragon/include/operators/activation/tanh_op.h
......@@ -12,7 +12,7 @@
namespace dragon {
template <class Context>
class TanhOp final : public Operator<Context> {
class TanhOp : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(TanhOp);
......@@ -21,7 +21,7 @@ class TanhOp final : public Operator<Context> {
};
template <class Context>
class TanhGradientOp final : public Operator<Context> {
class TanhGradientOp : public Operator<Context> {
public:
TanhGradientOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws) {
......@@ -32,6 +32,48 @@ class TanhGradientOp final : public Operator<Context> {
template <typename T> void RunWithType();
};
#ifdef WITH_CUDNN
template <class Context>
class CuDNNTanhOp final : public TanhOp<Context> {
public:
CuDNNTanhOp(const OperatorDef& op_def, Workspace* ws)
: TanhOp<Context>(op_def, ws) {
CUDNN_CHECK(cudnnCreateTensorDescriptor(&input_desc));
CUDNN_CHECK(cudnnCreateTensorDescriptor(&output_desc));
CUDNN_CHECK(cudnnCreateActivationDescriptor(&act_desc));
CUDNN_CHECK(cudnnSetActivationDescriptor(act_desc,
CUDNN_ACTIVATION_TANH, CUDNN_PROPAGATE_NAN, 0));
}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
cudnnTensorDescriptor_t input_desc, output_desc;
cudnnActivationDescriptor_t act_desc;
};
template <class Context>
class CuDNNTanhGradientOp final : public TanhGradientOp<Context> {
public:
CuDNNTanhGradientOp(const OperatorDef& op_def, Workspace* ws)
: TanhGradientOp<Context>(op_def, ws) {
CUDNN_CHECK(cudnnCreateTensorDescriptor(&input_desc));
CUDNN_CHECK(cudnnCreateTensorDescriptor(&output_desc));
CUDNN_CHECK(cudnnCreateActivationDescriptor(&act_desc));
CUDNN_CHECK(cudnnSetActivationDescriptor(act_desc,
CUDNN_ACTIVATION_TANH, CUDNN_PROPAGATE_NAN, 0));
}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
cudnnTensorDescriptor_t input_desc, output_desc;
cudnnActivationDescriptor_t act_desc;
};
#endif // WITH_CUDNN
} // namespace dragon
#endif // DRAGON_OPERATORS_ACTIVATION_TANH_OP_H_
\ No newline at end of file
Dragon/include/operators/arithmetic/add_op.h
......@@ -38,6 +38,33 @@ class AddGradientOp final : public Operator<Context> {
Tensor* bcast_multiplier;
};
template <class Context>
class RAddOp final : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(RAddOp);
void RunOnDevice() override;
template <typename T> void EltwiseRunWithType();
template <typename T> void BroadcastRunWithType(int type);
protected:
Tensor* bcast_multiplier;
};
template <class Context>
class RAddGradientOp final : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(RAddGradientOp);
void ShareGradient() override;
void RunOnDevice() override;
template <typename T> void EltwiseRunWithType();
template <typename T> void BroadcastRunWithType(int type);
protected:
Tensor* bcast_multiplier;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_ARITHMETIC_ADD_OP_H_
\ No newline at end of file
Dragon/include/operators/arithmetic/div_op.h
......@@ -38,6 +38,33 @@ class DivGradientOp final : public Operator<Context> {
Tensor* bcast_multiplier;
};
template <class Context>
class RDivOp final : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(RDivOp);
void RunOnDevice() override;
template <typename T> void EltwiseRunWithType();
template <typename T> void BroadcastRunWithType(int type);
protected:
Tensor* bcast_multiplier;
};
template <class Context>
class RDivGradientOp final : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(RDivGradientOp);
void ShareGradient() override;
void RunOnDevice() override;
template <typename T> void EltwiseRunWithType();
template <typename T> void BroadcastRunWithType(int type);
protected:
Tensor* bcast_multiplier;
};
} // namepsace dragon
#endif // DRAGON_OPERATORS_ARITHMETIC_DIV_OP_H_
\ No newline at end of file
Dragon/include/operators/arithmetic/eltwise_op.h
......@@ -20,7 +20,7 @@ class EltwiseOp final : public Operator<Context> {
coeffs(OperatorBase::GetRepeatedArg<float>("coeffs")) {
if (coeffs.size() > 0) {
CHECK_EQ(coeffs.size(), InputSize())
<< "\nop has " << InputSize() << " inputs, "
<< "\nOp has " << InputSize() << " inputs, "
<< "but provided " << coeffs.size() << " coeffs.";
} else coeffs.resize(InputSize(), float(1));
}
......
Dragon/include/operators/arithmetic/mul_op.h
......@@ -38,6 +38,33 @@ class MulGradientOp final : public Operator<Context> {
Tensor* bcast_multiplier;
};
template <class Context>
class RMulOp final : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(RMulOp);
void RunOnDevice() override;
template <typename T> void EltwiseRunWithType();
template <typename T> void BroadcastRunWithType(int type);
protected:
Tensor* bcast_multiplier;
};
template <class Context>
class RMulGradientOp final : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(RMulGradientOp);
void ShareGradient() override;
void RunOnDevice() override;
template <typename T> void EltwiseRunWithType();
template <typename T> void BroadcastRunWithType(int type);
protected:
Tensor* bcast_multiplier;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_ARITHMETIC_MUL_OP_H_
\ No newline at end of file
Dragon/include/operators/arithmetic/sub_op.h
......@@ -38,6 +38,33 @@ class SubGradientOp final : public Operator<Context> {
Tensor* bcast_multiplier;
};
template <class Context>
class RSubOp final : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(RSubOp);
void RunOnDevice() override;
template <typename T> void EltwiseRunWithType();
template <typename T> void BroadcastRunWithType(int type);
protected:
Tensor* bcast_multiplier;
};
template <class Context>
class RSubGradientOp final : public Operator<Context> {
public:
USE_SIMPLE_CTOR_DTOR(RSubGradientOp);
void ShareGradient() override;
void RunOnDevice() override;
template <typename T> void EltwiseRunWithType();
template <typename T> void BroadcastRunWithType(int type);
protected:
Tensor* bcast_multiplier;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_ARITHMETIC_SUB_OP_H_
\ No newline at end of file
Dragon/include/operators/utils/cast_op.h → Dragon/include/operators/cast/float2half_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_CAST_OP_H_
#define DRAGON_OPERATORS_UTILS_CAST_OP_H_
#ifndef DRAGON_OPERATORS_CAST_FLOAT2HALF_OP_H_
#define DRAGON_OPERATORS_CAST_FLOAT2HALF_OP_H_
#include "core/operator.h"
......@@ -20,4 +20,4 @@ class FloatToHalfOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_UTILS_CAST_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_CAST_FLOAT2HALF_OP_H_
\ No newline at end of file
Dragon/include/operators/utils/compare_op.h → Dragon/include/operators/control_flow/compare_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_COMPARE_OP_H_
#define DRAGON_OPERATORS_UTILS_COMPARE_OP_H_
#ifndef DRAGON_OPERATORS_CONTROL_FLOW_COMPARE_OP_H_
#define DRAGON_OPERATORS_CONTROL_FLOW_COMPARE_OP_H_
#include "core/operator.h"
......@@ -27,4 +27,4 @@ class CompareOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_UTILS_COMPARE_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_CONTROL_FLOW_COMPARE_OP_H_
\ No newline at end of file
Dragon/include/operators/utils/copy_op.h → Dragon/include/operators/control_flow/copy_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_COPY_OP_H_
#define DRAGON_OPERATORS_UTILS_COPY_OP_H_
#ifndef DRAGON_OPERATORS_CONTROL_FLOW_COPY_OP_H_
#define DRAGON_OPERATORS_CONTROL_FLOW_COPY_OP_H_
#include "core/operator.h"
......@@ -21,4 +21,4 @@ class CopyOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_UTILS_COPY_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_CONTROL_FLOW_COPY_OP_H_
\ No newline at end of file
Dragon/include/operators/common/scan_op.h → Dragon/include/operators/control_flow/scan_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_SCAN_OP_H_
#define DRAGON_OPERATORS_COMMON_SCAN_OP_H_
#ifndef DRAGON_OPERATORS_CONTROL_FLOW_SCAN_OP_H_
#define DRAGON_OPERATORS_CONTROL_FLOW_SCAN_OP_H_
#include "core/operator.h"
......@@ -80,4 +80,4 @@ class ScanGradientOp final: public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_SCAN_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_CONTROL_FLOW_SCAN_OP_H_
\ No newline at end of file
Dragon/include/operators/utils/accuracy_op.h → Dragon/include/operators/misc/accuracy_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_ACCURACY_OP_H_
#define DRAGON_OPERATORS_UTILS_ACCURACY_OP_H_
#ifndef DRAGON_OPERATORS_MISC_ACCURACY_OP_H_
#define DRAGON_OPERATORS_MISC_ACCURACY_OP_H_
#include "core/operator.h"
......@@ -16,7 +16,8 @@ class AccuracyOp final: public Operator<Context> {
public:
AccuracyOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
top_k(OperatorBase::GetSingleArg<int>("top_k", 1)) {
top_k(OperatorBase::GetSingleArg<int>("top_k", 1)),
axis(OperatorBase::GetSingleArg<int>("axis", 1)) {
vector<int> args = OperatorBase::GetRepeatedArg<int>("ignore_labels");
if (args.size()) {
ignore_labels.Reshape(vector<TIndex>(1, args.size()));
......@@ -29,10 +30,10 @@ class AccuracyOp final: public Operator<Context> {
template <typename T> void RunWithType();
protected:
TIndex top_k, outer_num, inner_num, classes;
TIndex top_k, axis, outer_dim, inner_dim, num_classes;
Tensor ignore_labels;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_UTILS_ACCURACY_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_MISC_ACCURACY_OP_H_
\ No newline at end of file
Dragon/include/operators/utils/gradient_op.h → Dragon/include/operators/misc/gradient_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_GRADIENT_GENERATE_OP_H_
#define DRAGON_OPERATORS_UTILS_GRADIENT_GENERATE_OP_H_
#ifndef DRAGON_OPERATORS_MISC_GRADIENT_GENERATE_OP_H_
#define DRAGON_OPERATORS_MISC_GRADIENT_GENERATE_OP_H_
#include "core/operator.h"
......@@ -59,4 +59,4 @@ class StopGradientOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_UTILS_GRADIENT_GENERATE_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_MISC_GRADIENT_GENERATE_OP_H_
\ No newline at end of file
Dragon/include/operators/utils/initialize_op.h → Dragon/include/operators/misc/initialize_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_INITIALIZE_OP_H_
#define DRAGON_OPERATORS_UTILS_INITIALIZE_OP_H_
#ifndef DRAGON_OPERATORS_MISC_INITIALIZE_OP_H_
#define DRAGON_OPERATORS_MISC_INITIALIZE_OP_H_
#include "core/operator.h"
#include "utils/filler.h"
......@@ -67,10 +67,12 @@ public:
TruncatedNormalOp(const OperatorDef& op_def, Workspace* ws)
: InitializeOp<Context>(op_def, ws) {
this->filler.set_type("truncated_normal");
this->filler.set_mean(OperatorBase::GetSingleArg<float>("mean", 0.0));
this->filler.set_std(OperatorBase::GetSingleArg<float>("std", 1.0));
this->filler.set_low(OperatorBase::GetSingleArg<float>("low", -2.0));
this->filler.set_high(OperatorBase::GetSingleArg<float>("high", 2.0));
float mu = OperatorBase::GetSingleArg<float>("mean", 0.0);
float sigma = OperatorBase::GetSingleArg<float>("std", 1.0);
this->filler.set_mean(mu);
this->filler.set_std(sigma);
this->filler.set_low(mu - 2 * sigma);
this->filler.set_high(mu + 2 * sigma);
}
};
......@@ -116,4 +118,4 @@ public:
} // namespace
#endif // DRAGON_OPERATORS_UTILS_INITIALIZE_OP_H_
#endif // DRAGON_OPERATORS_MISC_INITIALIZE_OP_H_
Dragon/include/operators/utils/memory_data_op.h → Dragon/include/operators/misc/memory_data_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_MEMORY_DATA_OP_H_
#define DRAGON_OPERATORS_UTILS_MEMORY_DATA_OP_H_
#ifndef DRAGON_OPERATORS_MISC_MEMORY_DATA_OP_H_
#define DRAGON_OPERATORS_MISC_MEMORY_DATA_OP_H_
#include "core/operator.h"
......@@ -30,4 +30,4 @@ class MemoryDataOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_UTILS_MEMORY_DATA_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_MISC_MEMORY_DATA_OP_H_
\ No newline at end of file
Dragon/include/operators/utils/proposal_op.h → Dragon/include/operators/misc/proposal_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_PROPOSAL_OP_H_
#define DRAGON_OPERATORS_UTILS_PROPOSAL_OP_H_
#ifndef DRAGON_OPERATORS_MISC_PROPOSAL_OP_H_
#define DRAGON_OPERATORS_MISC_PROPOSAL_OP_H_
#include "core/operator.h"
......@@ -37,4 +37,4 @@ class ProposalOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_UTILS_COMPARE_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_MISC_COMPARE_OP_H_
\ No newline at end of file
Dragon/include/operators/common/python_op.h → Dragon/include/operators/misc/python_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_PYTHON_OP_H_
#define DRAGON_OPERATORS_COMMON_PYTHON_OP_H_
#ifndef DRAGON_OPERATORS_MISC_PYTHON_OP_H_
#define DRAGON_OPERATORS_MISC_PYTHON_OP_H_
#include <Python.h>
......@@ -52,4 +52,4 @@ public:
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_PYTHON_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_MISC_PYTHON_OP_H_
\ No newline at end of file
Dragon/include/operators/mpi/base_mpi_op.h
......@@ -33,7 +33,7 @@ class ModelMPIBase : public Operator<Context> {
int world_root = OperatorBase::GetSingleArg<int>("root", 0);
MPI_Group_translate_ranks(world_group, 1, &world_root, group, &comm_root);
CHECK(comm_root != MPI_UNDEFINED) << "mpi root is not included in layer group.";
CHECK(comm_root != MPI_UNDEFINED) << "MPI root is not included in layer group.";
}
protected:
......
Dragon/include/operators/ndarray/arange_op.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_NDARRAY_ARGMAX_OP_H_
#define DRAGON_OPERATORS_NDARRAY_ARGMAX_OP_H_
#include "core/operator.h"
namespace dragon {
template <class Context>
class ArangeOp final : public Operator<Context> {
public:
ArangeOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
start(OperatorBase::GetSingleArg<int>("static_start", 0)),
stop(OperatorBase::GetSingleArg<int>("static_stop", -1)),
step(OperatorBase::GetSingleArg<int>("static_step", 1)),
dtype(OperatorBase::GetSingleArg<string>("dtype", "FLOAT32")) {
dynamic_start_ = OperatorBase::GetSingleArg<string>("dynamic_start", "");
dynamic_stop_ = OperatorBase::GetSingleArg<string>("dynamic_stop", "");
dynamic_step_ = OperatorBase::GetSingleArg<string>("dynamic_step", "");
}
void RunOnDevice() override;
void Reshape();
template <typename T> void RunWithType();
protected:
TIndex start, stop, step, count;
Tensor* dynamic_start, *dynamic_stop, *dynamic_step;
string dynamic_start_, dynamic_stop_, dynamic_step_;
string dtype;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_NDARRAY_ARANGE_OP_H_
\ No newline at end of file
Dragon/include/operators/common/argmax_op.h → Dragon/include/operators/ndarray/argmax_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_ARGMAX_OP_H_
#define DRAGON_OPERATORS_COMMON_ARGMAX_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_ARGMAX_OP_H_
#define DRAGON_OPERATORS_NDARRAY_ARGMAX_OP_H_
#include "core/operator.h"
......@@ -16,16 +16,18 @@ class ArgmaxOp final : public Operator<Context> {
public:
ArgmaxOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
axis(OperatorBase::GetSingleArg<int>("axis", 0)),
axis(OperatorBase::GetSingleArg<int>("axis", -1)),
keep_dims(OperatorBase::GetSingleArg<bool>("keep_dims", false)),
top_k(OperatorBase::GetSingleArg<int>("top_k", 1)) {}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
TIndex axis, top_k, count, inner_dim;
TIndex axis, axis_dim, top_k, count, inner_dim;
bool keep_dims;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_ARGMAX_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_ARGMAX_OP_H_
\ No newline at end of file
Dragon/include/operators/ndarray/argmin_op.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_NDARRAY_ARGMIN_OP_H_
#define DRAGON_OPERATORS_NDARRAY_ARGMIN_OP_H_
#include "core/operator.h"
namespace dragon {
template <class Context>
class ArgminOp final : public Operator<Context> {
public:
ArgminOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
axis(OperatorBase::GetSingleArg<int>("axis", -1)),
keep_dims(OperatorBase::GetSingleArg<bool>("keep_dims", false)),
top_k(OperatorBase::GetSingleArg<int>("top_k", 1)) {}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
TIndex axis, axis_dim, top_k, count, inner_dim;
bool keep_dims;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_NDARRAY_ARGMIN_OP_H_
\ No newline at end of file
Dragon/include/operators/common/at_op.h → Dragon/include/operators/ndarray/at_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_AT_OP_H_
#define DRAGON_OPERATORS_COMMON_AT_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_AT_OP_H_
#define DRAGON_OPERATORS_NDARRAY_AT_OP_H_
#include "core/operator.h"
......@@ -44,4 +44,4 @@ class AtGradientOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_AT_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_AT_OP_H_
\ No newline at end of file
Dragon/include/operators/common/concat_op.h → Dragon/include/operators/ndarray/concat_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_CONCAT_OP_H_
#define DRAGON_OPERATORS_COMMON_CONCAT_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_CONCAT_OP_H_
#define DRAGON_OPERATORS_NDARRAY_CONCAT_OP_H_
#include "core/operator.h"
......@@ -48,4 +48,4 @@ class ConcatGradientOp : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_CONCAT_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_CONCAT_OP_H_
\ No newline at end of file
Dragon/include/operators/common/crop_op.h → Dragon/include/operators/ndarray/crop_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_CROP_OP_H_
#define DRAGON_OPERATORS_COMMON_CROP_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_CROP_OP_H_
#define DRAGON_OPERATORS_NDARRAY_CROP_OP_H_
#include "core/operator.h"
......@@ -21,9 +21,9 @@ class CropOp: public Operator<Context> {
shape(OperatorBase::GetRepeatedArg<int>("shape")),
shape_like(OperatorBase::GetSingleArg<string>("shape_like", "")) {
CHECK(shape.size() * shape_like.size() == 0)
<< "\ncan not set shape and shape_like both.";
<< "\nCan not set shape and shape_like both.";
CHECK(shape.size() + shape_like.size() != 0)
<< "\nmust set shape and shape_like either.";
<< "\nMust set shape and shape_like either.";
}
void ComputeOutputShape();
......@@ -73,4 +73,4 @@ class CropGradientOp final : public Operator<Context > {
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_CROP_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_CROP_OP_H_
\ No newline at end of file
Dragon/include/operators/common/expand_dims_op.h → Dragon/include/operators/ndarray/expand_dims_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_EXPAND_DIMS_OP_H_
#define DRAGON_OPERATORS_COMMON_EXPAND_DIMS_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_EXPAND_DIMS_OP_H_
#define DRAGON_OPERATORS_NDARRAY_EXPAND_DIMS_OP_H_
#include "core/operator.h"
......@@ -36,4 +36,4 @@ class ExpandDimsGradientOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_EXPAND_DIMS_OP_H_
#endif // DRAGON_OPERATORS_NDARRAY_EXPAND_DIMS_OP_H_
Dragon/include/operators/common/flatten_op.h → Dragon/include/operators/ndarray/flatten_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_FLATTEN_OP_H_
#define DRAGON_OPERATORS_COMMON_FLATTEN_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_FLATTEN_OP_H_
#define DRAGON_OPERATORS_NDARRAY_FLATTEN_OP_H_
#include "core/operator.h"
......@@ -17,12 +17,15 @@ class FlattenOp final : public Operator<Context> {
FlattenOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
axis(OperatorBase::GetSingleArg<int>("axis", 0)),
num_axes(OperatorBase::GetSingleArg<int>("num_axes", -1)) {}
num_axes(OperatorBase::GetSingleArg<int>("num_axes", -1)),
keep_axes(OperatorBase::GetSingleArg<int>("keep_axes", INT_MAX)) {}
void RunOnDevice() override;
void SqueezeRun();
void KeepRun();
protected:
TIndex axis, num_axes;
TIndex axis, num_axes, keep_axes;
};
template <class Context>
......@@ -37,4 +40,4 @@ class FlattenGradientOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_FLATTEN_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_FLATTEN_OP_H_
\ No newline at end of file
Dragon/include/operators/utils/one_hot_op.h → Dragon/include/operators/ndarray/one_hot_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_ONE_HOT_OP_H_
#define DRAGON_OPERATORS_UTILS_ONE_HOT_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_ONE_HOT_OP_H_
#define DRAGON_OPERATORS_NDARRAY_ONE_HOT_OP_H_
#include "core/operator.h"
......@@ -29,4 +29,4 @@ class OneHotOp final : public Operator < Context > {
} // namespace dragon
#endif // DRAGON_OPERATORS_UTILS_ONE_HOT_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_ONE_HOT_OP_H_
\ No newline at end of file
Dragon/include/operators/ndarray/random_pick_op.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_NDARRAY_RANDOM_PICK_OP_H_
#define DRAGON_OPERATORS_NDARRAY_RANDOM_PICK_OP_H_
#include "core/operator.h"
namespace dragon {
template <class Context>
class RandomPickOp : public Operator<Context> {
public:
RandomPickOp(const OperatorDef& op_def, Workspace* ws) :
Operator<Context>(op_def, ws),
axis(OperatorBase::GetSingleArg<int>("axis", 0)),
max_samples(OperatorBase::GetSingleArg<int>("max_samples", 1)) {}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
TIndex axis, max_samples;
TIndex outer_dim, inner_dim, x_slice_dim, y_slice_dim;
vector<TIndex> output_dims;
Tensor* pick_indices;
};
template <class Context>
class RandomPickGradientOp final : public Operator<Context> {
public:
RandomPickGradientOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
axis(OperatorBase::GetSingleArg<int>("axis", 0)) {}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
TIndex axis;
TIndex outer_dim, inner_dim, x_slice_dim, y_slice_dim;
Tensor* pick_indices;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_NDARRAY_RANDOM_PICK_OP_H_
\ No newline at end of file
Dragon/include/operators/common/reduce_op.h → Dragon/include/operators/ndarray/reduce_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_REDUCE_OP_H_
#define DRAGON_OPERATORS_COMMON_REDUCE_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_REDUCE_OP_H_
#define DRAGON_OPERATORS_NDARRAY_REDUCE_OP_H_
#include "core/operator.h"
......@@ -50,4 +50,4 @@ class ReduceGradientOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_REDUCE_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_REDUCE_OP_H_
\ No newline at end of file
Dragon/include/operators/ndarray/repeat_op.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_NDARRAY_REPEAT_OP_H_
#define DRAGON_OPERATORS_NDARRAY_REPEAT_OP_H_
#include "core/operator.h"
namespace dragon {
template <class Context>
class RepeatOp : public Operator<Context> {
public:
RepeatOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
axis(OperatorBase::GetSingleArg<int>("axis", -1)),
repeats(OperatorBase::GetSingleArg<int>("repeats", 1)) {}
void RunOnDevice() override;
template<typename T> void RunWithType();
protected:
TIndex axis, repeats, outer_dim, dim, inner_dim;
};
template <class Context>
class RepeatGradientOp : public Operator<Context> {
public:
RepeatGradientOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
axis(OperatorBase::GetSingleArg<int>("axis", -1)),
repeats(OperatorBase::GetSingleArg<int>("repeats", 1)) {}
void RunOnDevice() override;
template<typename T> void RunWithType();
protected:
TIndex axis, repeats, outer_dim, dim, inner_dim;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_NDARRAY_REPEAT_OP_H_
\ No newline at end of file
Dragon/include/operators/common/reshape_op.h → Dragon/include/operators/ndarray/reshape_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_RESHAPE_OP_H_
#define DRAGON_OPERATORS_COMMON_RESHAPE_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_RESHAPE_OP_H_
#define DRAGON_OPERATORS_NDARRAY_RESHAPE_OP_H_
#include "core/operator.h"
......@@ -39,4 +39,4 @@ class ReshapeGradientOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_RESHAPE_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_RESHAPE_OP_H_
\ No newline at end of file
Dragon/include/operators/utils/shape_op.h → Dragon/include/operators/ndarray/shape_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UTILS_SHAPE_OP_H_
#define DRAGON_OPERATORS_UTILS_SHAPE_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_SHAPE_OP_H_
#define DRAGON_OPERATORS_NDARRAY_SHAPE_OP_H_
#include "core/operator.h"
......@@ -20,4 +20,4 @@ class ShapeOp final : public Operator<Context> {
} // namespace dragon
#endif //DRAGON_OPERATORS_UTILS_SHAPE_OP_H_
\ No newline at end of file
#endif //DRAGON_OPERATORS_NDARRAY_SHAPE_OP_H_
\ No newline at end of file
Dragon/include/operators/common/slice_op.h → Dragon/include/operators/ndarray/slice_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_SLICE_OP_H_
#define DRAGON_OPERATORS_COMMON_SLICE_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_SLICE_OP_H_
#define DRAGON_OPERATORS_NDARRAY_SLICE_OP_H_
#include "core/operator.h"
......@@ -51,4 +51,4 @@ class SliceGradientOp final : public Operator<Context> {
} // namespace dragon
#endif // #define DRAGON_OPERATORS_COMMON_SLICE_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_SLICE_OP_H_
\ No newline at end of file
Dragon/include/operators/ndarray/stack_op.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_NDARRAY_STACK_OP_H_
#define DRAGON_OPERATORS_NDARRAY_STACK_OP_H_
#include "core/operator.h"
namespace dragon {
template <class Context>
class StackOp : public Operator<Context> {
public:
StackOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
axis(OperatorBase::GetSingleArg<int>("axis", 0)),
nin(OperatorBase::GetSingleArg<int>("num_input", 1)) {}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
TIndex axis, nin, outer_dim, inner_dim, x_concat_dim, y_concat_dim;
TIndex x_offset, y_offset, concat_offset;
vector<TIndex> stack_dims, concat_dims;
};
template <class Context>
class StackGradientOp : public Operator<Context> {
public:
StackGradientOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
axis(OperatorBase::GetSingleArg<int>("axis", 0)),
nin(OperatorBase::GetSingleArg<int>("num_input", 1)) {}
void ShareGradient() override;
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
TIndex axis, nin, outer_dim, inner_dim, x_concat_dim, y_concat_dim;
TIndex x_offset, y_offset, concat_offset;
vector<TIndex> concat_dims;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_NDARRAY_STACK_OP_H_
\ No newline at end of file
Dragon/include/operators/common/tile_op.h → Dragon/include/operators/ndarray/tile_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_TILE_OP_H_
#define DRAGON_OPERATORS_COMMON_TILE_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_TILE_OP_H_
#define DRAGON_OPERATORS_NDARRAY_TILE_OP_H_
#include "core/operator.h"
......@@ -28,7 +28,7 @@ class TileOp : public Operator<Context> {
protected:
vector<int> multiples;
vector< pair<int, int> > process_axes;
TIndex axis, multiple, outer_dim, dim, inner_dim;
TIndex axis, multiple, outer_dim, ex_inner_dim;
Tensor* dest, *source;
};
......@@ -49,10 +49,10 @@ class TileGradientOp : public Operator<Context> {
protected:
vector<int> multiples;
vector< pair<int, int> > process_axes;
TIndex axis, multiple, outer_dim, dim, inner_dim;
TIndex axis, multiple, outer_dim, ex_inner_dim;
Tensor* dest, *source;
};
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_TILE_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_TILE_OP_H_
\ No newline at end of file
Dragon/include/operators/common/transpose_op.h → Dragon/include/operators/ndarray/transpose_op.h
......@@ -4,8 +4,8 @@
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_COMMON_TRANSPOSE_OP_H_
#define DRAGON_OPERATORS_COMMON_TRANSPOSE_OP_H_
#ifndef DRAGON_OPERATORS_NDARRAY_TRANSPOSE_OP_H_
#define DRAGON_OPERATORS_NDARRAY_TRANSPOSE_OP_H_
#include "core/operator.h"
......@@ -16,13 +16,17 @@ class TransposeOp final: public Operator<Context> {
public:
TransposeOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
perm(OperatorBase::GetRepeatedArg<int>("perm")) {}
perms(OperatorBase::GetRepeatedArg<int>("perms")) {
if (perms.size() > 0) reverse_dims = false;
else reverse_dims = true;
}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
vector<int> perm;
vector<int> perms;
bool reverse_dims;
Tensor* order, *old_steps, *new_steps;
};
......@@ -42,4 +46,4 @@ class TransposeGradientOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_COMMON_TRANSPOSE_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_NDARRAY_TRANSPOSE_OP_H_
\ No newline at end of file
Dragon/include/operators/update/async_update_op.h deleted 100644 → 0
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UPDATE_ASYNC_UPDATE_OP_H_
#define DRAGON_OPERATORS_UPDATE_ASYNC_UPDATE_OP_H_
#ifdef WITH_MPI
#include "operators/update/update_op_base.h"
#include "utils/thread.h"
namespace dragon {
template <class Context>
class AsyncUpdateOp final: public UpdateOpBase<Context> {
public:
AsyncUpdateOp(const OperatorDef& op_def, Workspace* ws);
int GetDelay(int tag);
void UpdateTimestamp(int tag);
void RunOnDevice() override;
void ComputeRunWithFloat() override { /* do nothing */ }
template <typename T> void RootRunWithType();
template <typename T> void ThreadRunWithType();
protected:
string mode;
unique_ptr<Tensor> recv_buffer;
Tensor** acc_buffers;
string* tags;
TIndex update_count;
int node_id, nsync, max_recv;
Map<int, int> local_timestamp;
std::unique_ptr<std::thread> thread;
#ifdef WITH_MPI_CUDA
cudaStream_t stream;
cublasHandle_t handle;
#endif
};
} // namespace dragon
#endif // WITH_MPI
#endif // DRAGON_OPERATORS_UPDATE_ASYNC_UPDATE_OP_H_
\ No newline at end of file
Dragon/include/operators/update/collective_update_op.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_UPDATE_COLLECTIVE_UPDATE_OP_H_
#define DRAGON_OPERATORS_UPDATE_COLLECTIVE_UPDATE_OP_H_
#include "core/operator.h"
namespace dragon {
#ifdef WITH_MPI
template <class Context>
class CollectiveUpdateOp : public Operator<Context> {
public:
CollectiveUpdateOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
mode(OperatorBase::GetSingleArg<string>("mode", "UNKNOWN")) {
InitMPI();
if (mode.find("NCCL") != string::npos) InitNCCL();
}
void InitMPI();
void InitNCCL();
void RunOnDevice() override;
void MPIAllReduceWithFloat();
void NCCLAllReduceWithFloat();
void MPIBcastWithFloat();
void NCCLBcastWithFloat();
protected:
int comm_size, comm_rank, comm_root;
int world_size, world_rank;
Tensor* buffer;
string mode;
MPI_Comm comm;
MPI_Group group;
#ifdef WITH_MPI_NCCL
ncclComm_t nccl_comm;
cudaStream_t stream;
#endif
};
#endif // WITH_MPI
} // namespace dragon
#endif // DRAGON_OPERATORS_UPDATE_COLLECTIVE_UPDATE_OP_H_
\ No newline at end of file
Dragon/include/operators/update/moving_average_op.h
......@@ -26,8 +26,6 @@ class MovingAverageOp final : public Operator<Context> {
};
} // namespace dragon
......
Dragon/include/operators/update/update_op_base.h
......@@ -16,43 +16,21 @@ class UpdateOpBase : public Operator<Context> {
public:
UpdateOpBase(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
allow_parallel(false),
async_tag(-1),
lr_mult(OperatorBase::GetSingleArg<float>("lr_mult", 1.0)),
decay_mult(OperatorBase::GetSingleArg<float>("decay_mult", 1.0)),
domain(OperatorBase::GetSingleArg<string>("domain", "_")),
mode(OperatorBase::GetSingleArg<string>("mode", "Sync")) { InitMPI(); }
domain(OperatorBase::GetSingleArg<string>("domain", "_")) {}
float param(const string& name) const;
void InitMPI();
void RunOnDevice() override;
template <typename T> void ReduceRunWithType();
template <typename T> void PreprocessRunWithType();
virtual void ComputeRunWithFloat() = 0;
template <typename T> void UpdateRunWithType();
template <typename T> void RecvRunWithType();
protected:
float lr_mult, decay_mult;
float l2_decay, clip_thresh, scale_factor;
int comm_size, comm_rank, comm_root;
int world_size, world_rank;
bool allow_parallel;
int async_tag;
Tensor* buffer;
string domain, mode;
#ifdef WITH_MPI
MPI_Comm comm;
MPI_Group group;
#endif // WITH_MPI
#ifdef WITH_MPI_NCCL
ncclComm_t nccl_comm;
cudaStream_t stream;
#endif // WITH_MPI_NCCL
string domain;
};
} // namespace dragon
......
Dragon/include/operators/vision/bilinear_resize_op.h 0 → 100644
// --------------------------------------------------------
// Dragon
// Copyright(c) 2017 SeetaTech
// Written by Ting Pan
// --------------------------------------------------------
#ifndef DRAGON_OPERATORS_VISION_BILINEAR_RESIZE_OP_H_
#define DRAGON_OPERATORS_VISION_BILINEAR_RESIZE_OP_H_
#include "core/operator.h"
namespace dragon {
template <class Context>
class BilinearResizeOp : public Operator<Context> {
public:
BilinearResizeOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
static_dsize(OperatorBase::GetRepeatedArg<int>("static_dsize")),
dynamic_dsize(OperatorBase::GetRepeatedArg<string>("dynamic_dsize")),
fy(OperatorBase::GetSingleArg<float>("fy", -1.0)),
fx(OperatorBase::GetSingleArg<float>("fx", -1.0)) {}
void RunOnDevice() override;
template <typename T> void RunWithType();
protected:
vector<int> static_dsize;
vector<string> dynamic_dsize;
vector<TIndex> dims;
float h_scale, w_scale, fy, fx;
};
template <class Context>
class BilinearResizeGradientOp : public Operator<Context> {
public:
BilinearResizeGradientOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws) {}
void RunOnDevice() override;
template <typename T> void RunWithType();
};
} // namespace dragon
#endif // DRAGON_OPERATORS_VISION_BILINEAR_RESIZE_OP_H_
\ No newline at end of file
Dragon/include/operators/vision/dense_concat_op.h
......@@ -7,7 +7,7 @@
#ifndef DRAGON_OPERATORS_VISION_DENSE_CONCAT_OP_H_
#define DRAGON_OPERATORS_VISION_DENSE_CONCAT_OP_H_
#include "operators/common/concat_op.h"
#include "operators/ndarray/concat_op.h"
namespace dragon {
......
Dragon/include/operators/vision/nn_resize_op.h
......@@ -16,7 +16,8 @@ class NNResizeOp : public Operator<Context> {
public:
NNResizeOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
dsize(OperatorBase::GetRepeatedArg<int>("dsize")),
static_dsize(OperatorBase::GetRepeatedArg<int>("static_dsize")),
dynamic_dsize(OperatorBase::GetRepeatedArg<string>("dynamic_dsize")),
fy(OperatorBase::GetSingleArg<float>("fy", -1.0)),
fx(OperatorBase::GetSingleArg<float>("fx", -1.0)) {}
......@@ -24,7 +25,8 @@ class NNResizeOp : public Operator<Context> {
template <typename T> void RunWithType();
protected:
vector<int> dsize;
vector<int> static_dsize;
vector<string> dynamic_dsize;
vector<TIndex> dims;
float h_scale, w_scale, fy, fx;
};
......
Dragon/include/operators/vision/pooling_op.h
......@@ -16,22 +16,24 @@ enum PoolingMode { MAX_POOLING, AVG_POOLING };
template <class Context>
class PoolingOp: public Operator <Context> {
public:
PoolingOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws) {
vector<int> ks = OperatorBase::GetRepeatedArg<int>("kernel_size");
for (int i = 0; i < 2; i++)
kernel_size.push_back(i < ks.size() ? ks[i] : ks[0]);
vector<int> s = OperatorBase::GetRepeatedArg<int>("stride");
for (int i = 0; i < 2; i++)
stride.push_back(i < s.size() ? s[i] : s[0]);
vector<int> p = OperatorBase::GetRepeatedArg<int>("pad");
for (int i = 0; i < 2; i++)
pad.push_back(i < p.size() ? p[i] : p[0]);
mode = PoolingMode(OperatorBase::GetSingleArg<int>("mode", MAX_POOLING));
PoolingOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws),
mode(PoolingMode(OperatorBase::GetSingleArg<int>("mode", MAX_POOLING))),
global_pooling(OperatorBase::GetSingleArg<bool>("global_pooling", false)) {
vector<int> ks = OperatorBase::GetRepeatedArg<int>("kernel_size");
vector<int> s = OperatorBase::GetRepeatedArg<int>("stride");
vector<int> p = OperatorBase::GetRepeatedArg<int>("pad");
for (int i = 0; i < 2; i++) {
if (global_pooling) {
kernel_size.push_back(-1);
stride.push_back(1);
pad.push_back(0);
} else {
kernel_size.push_back(i < ks.size() ? ks[i] : ks[0]);
stride.push_back(i < s.size() ? s[i] : s[0]);
pad.push_back(i < p.size() ? p[i] : p[0]);
}
}
}
void Reshape();
......@@ -45,27 +47,30 @@ class PoolingOp: public Operator <Context> {
PoolingMode mode;
TIndex num, channels, height, width;
TIndex pool_height, pool_width;
bool global_pooling;
};
template <class Context>
class PoolingGradientOp: public Operator<Context> {
public:
PoolingGradientOp(const OperatorDef& op_def, Workspace* ws)
: Operator<Context>(op_def, ws) {
vector<int> ks = OperatorBase::GetRepeatedArg<int>("kernel_size");
for (int i = 0; i < 2; i++)
kernel_size.push_back(i < ks.size() ? ks[i] : ks[0]);
vector<int> s = OperatorBase::GetRepeatedArg<int>("stride");
for (int i = 0; i < 2; i++)
stride.push_back(i < s.size() ? s[i] : s[0]);
vector<int> p = OperatorBase::GetRepeatedArg<int>("pad");
for (int i = 0; i < 2; i++)
pad.push_back(i < p.size() ? p[i] : p[0]);
mode = PoolingMode(OperatorBase::GetSingleArg<int>("mode", MAX_POOLING));
: Operator<Context>(op_def, ws),
mode(PoolingMode(OperatorBase::GetSingleArg<int>("mode", MAX_POOLING))),
global_pooling(OperatorBase::GetSingleArg<bool>("global_pooling", false)) {
vector<int> ks = OperatorBase::GetRepeatedArg<int>("kernel_size");
vector<int> s = OperatorBase::GetRepeatedArg<int>("stride");
vector<int> p = OperatorBase::GetRepeatedArg<int>("pad");
for (int i = 0; i < 2; i++) {
if (global_pooling) {
kernel_size.push_back(-1);
stride.push_back(1);
pad.push_back(0);
} else {
kernel_size.push_back(i < ks.size() ? ks[i] : ks[0]);
stride.push_back(i < s.size() ? s[i] : s[0]);
pad.push_back(i < p.size() ? p[i] : p[0]);
}
}
}
void Reshape();
......@@ -79,6 +84,7 @@ class PoolingGradientOp: public Operator<Context> {
PoolingMode mode;
TIndex num, channels, height, width;
TIndex pool_height, pool_width;
bool global_pooling;
};
#ifdef WITH_CUDNN
......
Dragon/include/operators/vision/roi_pooling_op.h
......@@ -53,4 +53,4 @@ class ROIPoolingGradientOp final : public Operator<Context> {
} // namespace dragon
#endif // DRAGON_OPERATORS_VISION_ROI_POOLING_OP_H_
\ No newline at end of file
#endif // DRAGON_OPERATORS_VISION_ROI_POOLING_OP_H_
\ No newline at end of file
Dragon/include/utils/caffemodel.h
......@@ -49,7 +49,7 @@ inline bool ReadProtoFromBinaryFile(const char* filename, Message* proto) {
return success;
}
inline void LoadCaffeModel(string file, string scope, Workspace* ws) {
inline void LoadCaffeModel(string file, Workspace* ws) {
NetParameter net_param;
ReadProtoFromBinaryFile(file.c_str(), &net_param);
LOG(INFO) << "Restore From Model @: " << file << "......";
......@@ -57,7 +57,7 @@ inline void LoadCaffeModel(string file, string scope, Workspace* ws) {
for (int i = 0; i < net_param.layer_size(); i++) {
const LayerParameter& layer = net_param.layer(i);
const string& layer_name = layer.name();
string prefix = scope + layer_name + "@param";
string prefix = layer_name + "@param";
for (int j = 0; j < layer.blobs_size(); j++) {
string tensor_name = prefix + dragon_cast<string, int>(j);
if (!ws->HasTensor(tensor_name))
......@@ -111,8 +111,8 @@ inline void SavaCaffeModel(string file, const vector<Tensor*>& tensors) {
}
std::fstream output(file, std::ios::out | std::ios::trunc | std::ios::binary);
CHECK(net_param.SerializeToOstream(&output));
LOG(INFO) << "save the model @: " << file << "......";
LOG(INFO) << "model format: caffemodel";
LOG(INFO) << "Save the model @: " << file << "......";
LOG(INFO) << "Model format: caffemodel";
}
} // namespace dragon
......
Dragon/include/utils/cast.h
......@@ -8,7 +8,9 @@
#define DRAGON_UTILS_CAST_H_
#include <cstring>
#include "core/types.h"
#include "utils/cuda_device.h"
namespace dragon {
......@@ -113,6 +115,45 @@ template<> inline float32 dragon_cast<float32, float>(float val) {
return dragon_cast<float32, float16>(t);
}
#ifdef WITH_CUDA_FP16
template<> inline half dragon_cast<half, float>(float val) {
#if CUDA_VERSION_MIN(9, 0, 0)
__half_raw fp16_raw;
fp16_raw.x = dragon_cast<float16, float>(val).x;
return half(fp16_raw);
#else
half fp16;
fp16.x = dragon_cast<float16, float>(val).x;
return fp16;
#endif
}
template<> inline half2 dragon_cast<half2, float>(float val) {
#if CUDA_VERSION_MIN(9, 0, 0)
half fp16 = dragon_cast<half, float>(val);
return half2(fp16, fp16);
#else
half2 fp32;
fp32.x = dragon_cast<float32, float>(val).x;
return fp32;
#endif
}
template<> inline half2 dragon_cast<half2, float16>(float16 val) {
#if CUDA_VERSION_MIN(9, 0, 0)
__half_raw fp16_raw;
fp16_raw.x = val.x;
return half2(half(fp16_raw), half(fp16_raw));
#else
half2 fp32;
fp32.x = dragon_cast<float32, float16>(val).x;
return fp32;
#endif
}
#endif // WITH_CUDA_FP16
} // namespace dragon
......
Dragon/include/utils/cuda_device.h
......@@ -15,7 +15,7 @@
#include <cuda.h>
#ifdef WITH_MPI_NCCL
#include <nccl/nccl.h>
#include <nccl.h>
#endif // WITH_MPI_NCCL
#include "core/common.h"
......@@ -25,6 +25,12 @@ namespace dragon {
static const int CUDA_NUM_THREADS = 1024;
#define MAX_GPUS 8
#define CUDA_VERSION_MIN(major, minor, patch) \
(CUDA_VERSION >= (major * 1000 + minor * 100 + patch))
#define CUDA_VERSION_MAX(major, minor, patch) \
(CUDA_VERSION < (major * 1000 + minor * 100 + patch))
#define CUDA_CHECK(condition) \
do { \
cudaError_t error = condition; \
......@@ -61,6 +67,10 @@ inline int GET_BLOCKS(const int N) {
#define CUDA_POST_KERNEL_CHECK CUDA_CHECK(cudaPeekAtLastError())
#if CUDA_VERSION_MAX(9, 0, 0)
#define __hdiv hdiv
#endif
inline int NUM_DEVICES() {
static int count = -1;
if (count < 0) {
......
Dragon/include/utils/math_functions.h
......@@ -107,7 +107,7 @@ template<typename T, class Context>
void AddScalar(const int n, const float alpha, T* y);
template<typename T, class Context>
void MulScalar(const int n, const T alpha, T* y);
void MulScalar(const int n, const float alpha, T* y);
template<typename T, class Context>
void Axpy(const int n, float alpha, const T* x, T *y);
......
Dragon/include/utils/omp_alternative.h
......@@ -14,7 +14,7 @@
namespace dragon {
#define OMP_MIN_ITERATORS_PER_CORE 256
#define OMP_MIN_ITERATORS_PER_CORE 200000
inline int GET_OMP_THREADS(const int N) {
int threads = std::max(N / OMP_MIN_ITERATORS_PER_CORE, 1);
......
Dragon/modules/python/CMakeLists.txt
......@@ -23,6 +23,9 @@ endif()
if (UNIX AND WITH_BLAS)
TARGET_LINK_LIBRARIES(${PROJECT_NAME}_python openblas)
endif()
if (UNIX AND WITH_MPI_NCCL)
TARGET_LINK_LIBRARIES(${PROJECT_NAME}_python nccl)
endif()
# ---[ link platforms
if(UNIX)
......
Dragon/modules/python/dragon.h
......@@ -96,9 +96,9 @@ class NumpyFetcher : public TensorFetcherBase {
CHECK_GT(tensor.count(), 0);
vector<npy_intp> npy_dims;
for (const auto dim : tensor.dims()) npy_dims.push_back(dim);
int numpy_type = DragonToNumpyType(tensor.meta()); // translate a Meta to a int
int numpy_type = DragonToNumpyType(tensor.meta());
if (numpy_type == -1) {
string s = "Tensor(" + tensor.name() + "): unknown type yet, really run the net?";
string s = "The data type of Tensor(" + tensor.name() + ") is unknown. Have you solved it ?";
PyErr_SetString(PyExc_RuntimeError, s.c_str());
return nullptr;
}
......@@ -134,12 +134,12 @@ class NumpyFeeder : public TensorFeederBase {
PyArrayObject* array = PyArray_GETCONTIGUOUS(original_array);
const TypeMeta& meta = NumpyTypeToDragon(PyArray_TYPE(array));
if (meta.id() == 0) {
PyErr_SetString(PyExc_TypeError, "numpy data type is not supported.");
PyErr_SetString(PyExc_TypeError, "Unsupported data type.");
return nullptr;
}
if (meta.id() != tensor->meta().id() && tensor->meta().id() != 0)
LOG(WARNING) << "feed Tensor(" << tensor->name() << ")"
<< " with different dtype from original's.";
LOG(WARNING) << "Feed Tensor(" << tensor->name() << ")"
<< " with different data type from original one.";
tensor->SetMeta(meta);
int ndim = PyArray_NDIM(array);
npy_intp* npy_dims = PyArray_DIMS(array);
......@@ -154,7 +154,7 @@ class NumpyFeeder : public TensorFeederBase {
tensor->raw_mutable_data<CUDAContext>(),
static_cast<void*>(PyArray_DATA(array)));
#else
LOG(FATAL) << "CUDA is not compilied.";
LOG(FATAL) << "CUDA was not compiled.";
#endif
} else{
CPUContext::Memcpy<CPUContext, CPUContext>(tensor->nbytes(),
......
Dragon/modules/python/py_mpi.h
......@@ -24,7 +24,7 @@ inline PyObject* MPIInitCC(PyObject* self, PyObject* args) {
int thread_type;
MPI_Init_thread(NULL, NULL, MPI_THREAD_MULTIPLE, &thread_type);
CHECK_EQ(thread_type, MPI_THREAD_MULTIPLE)
<< "require enable <MPI_THREAD_MULTIPLE> support.";
<< "\nRequire to enable <MPI_THREAD_MULTIPLE> support.";
Py_RETURN_TRUE;
}
......@@ -72,7 +72,7 @@ inline PyObject* MPICreateGroupCC(PyObject* self, PyObject* args) {
all_ranks.insert(ranks[i]);
}
err_code = MPI_Group_incl(world_group, size, ranks, &local_group);
CHECK(err_code == MPI_SUCCESS) << "failed to create mpi group.";
CHECK(err_code == MPI_SUCCESS) << "\nFail to create mpi group.";
}
// check exclude ranks
......@@ -88,11 +88,11 @@ inline PyObject* MPICreateGroupCC(PyObject* self, PyObject* args) {
for (int i = 0; i < world_size; i++)
if (!tmp.count(i)) all_ranks.insert(i);
err_code = MPI_Group_excl(world_group, size, ranks, &local_group);
CHECK(err_code == MPI_SUCCESS) << "failed to create mpi group.";
CHECK(err_code == MPI_SUCCESS) << "Fail to create mpi group.";
}
err_code = MPI_Comm_create(MPI_COMM_WORLD, local_group, &local_comm);
CHECK(err_code == MPI_SUCCESS) << "failed to create mpi group.";
CHECK(err_code == MPI_SUCCESS) << "Fail to create mpi group.";
if (local_comm != MPI_COMM_NULL) {
int world_rank, local_size;
......@@ -120,7 +120,7 @@ inline PyObject* MPICreateGroupCC(PyObject* self, PyObject* args) {
#else // WITH_MPI
#define MPI_NOT_IMPLEMENTED \
LOG(FATAL) << "MPI is not compilied."; \
LOG(FATAL) << "MPI was not compiled."; \
Py_RETURN_TRUE
inline PyObject* MPIInitCC(PyObject* self, PyObject* args) { MPI_NOT_IMPLEMENTED; }
......
Dragon/python/dragon/config.py
......@@ -16,21 +16,52 @@ option = {}
REGISTERED_OPERATORS = set(s for s in RegisteredOperatorsCC())
NO_GRADIENT_OPERATORS = set(s for s in NoGradientOperatorsCC())
# The current device, 'CPU' or 'CUDA'
option['device'] = 'CPU'
# The device id
option['gpu_id'] = 0
# Whether to use cuDNN if possible
option['use_cudnn'] = False
# The global random seed
option['random_seed'] = 3
# if True, disable Dragon-Memonger
# Disable the memonger if true
option['debug_mode'] = False
option['share_grads'] = False # set it by Dragon-Memonger
option['allow_mirrow_stage'] = True # default
# Set it by the memonger
option['share_grads'] = False
def EnableCPU():
"""Enable CPU mode globally.
Returns
-------
None
"""
global option
option['device'] = 'CPU'
def EnableCUDA(gpu_id=0, use_cudnn=True):
"""Enable CUDA mode globally.
Parameters
----------
gpu_id : int
The id of GPU to use.
use_cudnn : boolean
Whether to use cuDNN if available.
Returns
-------
None
"""
global option
option['device'] = 'CUDA'
option['gpu_id'] = gpu_id
......@@ -39,32 +70,99 @@ def EnableCUDA(gpu_id=0, use_cudnn=True):
# TODO(PhyscalX): please not use @setter
# TODO(PhyscalX): seems that it can't change the global value
def SetRandomSeed(seed):
"""Set the global random seed.
Parameters
----------
seed : int
The seed to use.
Returns
-------
None
"""
global option
option['random_seed'] = seed
def GetRandomSeed():
"""Get the global random seed.
Returns
-------
int
The global random seed.
"""
global option
return option['random_seed']
def SetGPU(id):
"""Set the global id GPU.
Parameters
----------
id : int
The id of GPU to use.
Returns
-------
None
"""
global option
option['gpu_id'] = id
def GetGPU(id):
"""Get the global id of GPU.
Returns
-------
int
The global id of GPU.
"""
global option
return option['gpu_id']
def SetDebugMode(mode):
def SetDebugMode(enabled=True):
"""Enable Debug mode globally.
It will disable all memory sharing optimizations.
Parameters
----------
enabled : boolean
Whether to enable debug mode.
Returns
-------
None
"""
global option
option['debug_mode'] = mode
option['debug_mode'] = enabled
def SetLoggingLevel(level):
"""
set the minimum level of logging
:param level: a str of DEBUG, INFO(default), WARNING, ERROR, FATAL
"""
"""Set the minimum level of Logging.
Parameters
----------
level : str
The level, ``DEBUG``, ``INFO``, ``WARNING``, ``ERROR`` or ``FATAL``.
Notes
-----
The default level is ``INFO``.
"""
SetLogLevelCC(level)
global logger
logger.setLevel({
......
Dragon/python/dragon/core/gradient_maker.py
......@@ -13,10 +13,34 @@ from dragon.__init__ import *
from .scope import GetOperatorName
class GraphGradientMaker(object):
"""
GraphGradientMaker is deigned to generate gradient operators automatically.
It relies on the generating rules defined in the C++ backend.
"""
@classmethod
def CreateGradientForOp(cls, op_def, g_output):
""" parse ops from string """
g_ops, g_inputs, defaults = CreateGradientDefsCC(op_def.SerializeToString(), g_output)
def CreateGradientForOp(cls, forward_op, g_output):
"""Generate the OperatorDef for ``BackwardOp`` by ``ForwardOp``.
Parameters
----------
forward_op : dragon_pb2.OperatorDef
The OperatorDef of ``ForwardOp``.
g_output : list of str
The inputs of ``BackwardOp`` (Precomputed Grads).
Returns
-------
tuple
The OpDef, outputs and defaults of ``BackwardOp``.
References
----------
The wrapper of ``CreateGradientDefsCC``.
"""
g_ops, g_inputs, defaults = \
CreateGradientDefsCC(forward_op.SerializeToString(), g_output)
for idx, g_op in enumerate(g_ops):
new_def = pb.OperatorDef()
new_def.ParseFromString(g_op)
......@@ -24,9 +48,28 @@ class GraphGradientMaker(object):
g_ops[idx] = new_def
return g_ops, g_inputs, defaults
@classmethod
def CheckMissingGrad(cls, forward_op, inputs_to_grads, blacklist, targets):
""" check the missing grads, if True, skip this op """
"""Check if missing Grads. If True, skip this Op.
Parameters
----------
forward_op : dragon_pb2.OperatorDef
The OperatorDef of ``ForwardOp``.
inputs_to_grads : dict
The dict of <input, g_input>.
blacklist : set of str
The set of ``NoGradient`` tensors.
targets : list of str
The solving targets.
Returns
-------
tuple
The result of checking and generated filling grads.
"""
if forward_op.type in config.NO_GRADIENT_OPERATORS:
for input in forward_op.input: blacklist.add(input)
return (True, None)
......@@ -49,22 +92,44 @@ class GraphGradientMaker(object):
# check pass, even if missing some grads
return (False, gen_grads)
@classmethod
def Make(cls, ops, targets):
def Make(cls, forward_ops, targets):
"""Make ``BackwardOps`` based on ``ForwardOps``.
Parameters
----------
forward_ops : list of dragon_pb2.OperatorDef
The operators of ``ForwardOp``.
targets : list of str
The solving targets.
Returns
-------
tuple
The ``ForwardOps`` and ``BackwardOps``.
See Also
--------
`theano.function(*args, **kwargs)`_ - How to make a graph. [**Theano Style**]
"""
inputs_to_grads = {}
inputs_count = defaultdict(int)
grads_count = defaultdict(int)
all_g_ops = []
all_split_grads = set()
blacklist = set()
backward_ops = []
# PLAY for the forward
for op in ops:
if op.type in config.NO_GRADIENT_OPERATORS: continue
for input in op.input: inputs_count[input] += 1
for forward_op in forward_ops:
if forward_op.type in config.NO_GRADIENT_OPERATORS: continue
for input in forward_op.input: inputs_count[input] += 1
# PLAY for the backward
for forward_op in ops[::-1]:
for forward_op in forward_ops[::-1]:
is_skip, gen_grads = cls.CheckMissingGrad(forward_op, inputs_to_grads, blacklist, targets)
g_outputs = list(inputs_to_grads.get(name, None) for name in forward_op.output)
g_ops, g_inputs, defaults = cls.CreateGradientForOp(forward_op, g_outputs)
......@@ -81,8 +146,8 @@ class GraphGradientMaker(object):
GetOperatorName()[1], defaults=values)
if forward_op.HasField('device_option'):
gen_op.device_option.CopyFrom(forward_op.device_option)
all_g_ops.append(gen_op)
for g_op in g_ops: all_g_ops.append(g_op)
backward_ops.append(gen_op)
for g_op in g_ops: backward_ops.append(g_op)
# split & gather grads for multi-used input
for g_op in g_ops:
......@@ -107,11 +172,11 @@ class GraphGradientMaker(object):
if g_op.HasField('device_option'):
gather_op.device_option.CopyFrom(g_op.device_option)
_, gather_op.name = GetOperatorName()
all_g_ops.append(gather_op)
backward_ops.append(gather_op)
g_op.output[g_output_idx] = split_name
# done
if not is_skip:
for name, grad in zip(forward_op.input, g_inputs):
if grad != '': inputs_to_grads[name] = grad
return ops, all_g_ops
\ No newline at end of file
return forward_ops, backward_ops
\ No newline at end of file
Dragon/python/dragon/core/mpi.py
......@@ -13,43 +13,162 @@ from dragon import MPIInitCC, MPIRankCC, MPISizeCC, \
_is_init = False
_snapshot_ranks = []
_parallel_groups = []
_parallel_mode = 'Sync'
_parallel_mode = 'MPI'
def init():
__all__ = [
'Init',
'Is_Init',
'Rank',
'Size',
'CreateGroup',
'Snapshot',
'AllowSnapshot',
'Parallel',
'AllowParallel',
'SetParallelMode',
'GetParallelMode',
'Finalize'
]
def _check_init():
global _is_init
if _is_init is False: Init()
def Init():
"""Init the MPI env.
Returns
-------
None
Notes
-----
This function can only be called once.
References
----------
The wrapper of ``MPIInitCC``
"""
MPIInitCC()
global _is_init
global _snapshot_ranks
_is_init = True
_snapshot_ranks = [i for i in xrange(size())]
_snapshot_ranks = [i for i in xrange(Size())]
def check_init():
global _is_init
if _is_init is False: init()
def is_init():
def Is_Init():
"""Whether the MPI env has initialized.
Returns
-------
boolean
"""
return _is_init
def rank():
check_init()
def Rank():
"""The world rank of current MPI node.
Returns
-------
int
The world rank.
References
----------
The wrapper of ``MPIRankCC``.
"""
_check_init()
return MPIRankCC()
def size():
check_init()
def Size():
"""The world size of current MPI env.
Returns
-------
int
The world size.
References
----------
The wrapper of ``MPISizeCC``.
"""
_check_init()
return MPISizeCC()
def group(root=0, incl=[], excl=[]):
check_init()
def CreateGroup(root=0, incl=[], excl=[]):
"""Construct a ``MPIGroup`` with specific members.
Parameters
----------
root : int
The root of this group.
incl : list
The include nodes.
excl: list
The exclude nodes.
Returns
-------
tuple
The local common and group id.
References
----------
The wrapper of ``MPICreateGroupCC``.
"""
_check_init()
comm, group = MPICreateGroupCC(root, incl, excl)
return np.int64(comm), np.int64(group)
def snapshot(incl):
check_init()
def Snapshot(incl):
"""Set the specific MPI nodes to snapshot.
The exclude nodes will not snapshot through `workspace.Snapshot(*args, **kwargs)`_.
Parameters
----------
incl : int or list
Returns
-------
None
"""
_check_init()
if not isinstance(incl, list): incl = [incl]
global _snapshot_ranks
_snapshot_ranks = incl
def parallel(conf):
check_init()
def Parallel(conf):
"""Set the specific MPI nodes for data parallelism.
Parameters
----------
conf : list
The list of configures. Each configure should be a list also.
Returns
-------
None
Examples
--------
>>> mpi.parallel([0, 1]) # rank(0, 1) will be into a parallel group.
>>> mpi.parallel([0, 1], [2, 3]) # rank(0, 1), rank(2, 3) will be into two parallel groups.
"""
_check_init()
if not isinstance(conf[0], list): conf = [conf]
for ele in conf:
if not isinstance(ele, list):
......@@ -57,28 +176,81 @@ def parallel(conf):
global _parallel_groups
_parallel_groups = conf
def allow_snapshot():
def AllowSnapshot():
"""Whether this node can snapshot.
Returns
-------
boolean
"""
global _snapshot_ranks
return rank() in _snapshot_ranks
return Rank() in _snapshot_ranks
def AllowParallel():
"""Whether this node was set for data parallelism.
Returns
-------
boolean
def allow_parallel():
"""
global _parallel_groups
world_rank = rank()
world_rank = Rank()
for idx, g in enumerate(_parallel_groups):
if world_rank in g: return idx, g
return -1, []
def set_parallel_mode(mode):
assert mode == 'Sync' or \
mode == 'Async' \
or mode == 'Async_No_Lock'
def SetParallelMode(mode):
"""Set the mode of data parallelism.
Parameters
----------
mode : str
The mode, ``MPI``, ``NCCL`` or ``MIXED``.
Returns
-------
None
Notes
-----
The default mode is ``MPI``.
"""
assert mode == 'MPI' or \
mode == 'NCCL' \
or mode == 'MIXED'
global _parallel_mode
_parallel_mode = mode
def get_parallel_mode():
def GetParallelMode():
"""Get the current mode of data parallelism.
Returns
-------
str
The mode, ``MPI``, ``NCCL`` or ``MIXED``.
"""
global _parallel_mode
return _parallel_mode
def finalize():
check_init()
def Finalize():
"""Finalize the MPI env.
Returns
-------
None
Notes
-----
This function should be called to close the initialized MPI env.
"""
_check_init()
MPIFinalizeCC()
\ No newline at end of file
Dragon/python/dragon/core/scope.py
......@@ -11,34 +11,100 @@ PHASE_SCOPE = ''
DEVICE_SCOPE = ''
ENGINE_SCOPE = ''
SEPARATOR = '/'
_CURRENT_OP_IDX = 0
_SCOPE_TENSOR_IDX = defaultdict(int)
__all__ = [
'GetTensorIdx',
'GetTensorName',
'GetOperatorIdx',
'GetOperatorName',
'TensorScope',
'PhaseScope',
'DeviceScope'
]
def GetOperatorIdx():
"""Get the available operator index.
Returns
-------
int
The operator index.
"""
global _CURRENT_OP_IDX
_CURRENT_OP_IDX = _CURRENT_OP_IDX + 1
return _CURRENT_OP_IDX - 1
def GetTensorIdx():
"""Get the available tensor index.
Returns
-------
int
The tensor index.
"""
global _SCOPE_TENSOR_IDX
_SCOPE_TENSOR_IDX[TENSOR_SCOPE] += 1
return _SCOPE_TENSOR_IDX[TENSOR_SCOPE] - 1
def GetOperatorName(name=None):
"""Get the available operator name.
Parameters
----------
name : str
The optional name to use.
Returns
-------
str
The operator name.
"""
op_idx = GetOperatorIdx()
if name is None:
return op_idx, 'Op_' + str(op_idx)
else: return op_idx, name
def GetTensorName():
"""Get the available tensor name.
Returns
-------
str
The operator name.
"""
return 'Tensor_' + str(GetTensorIdx())
class TensorScope(object):
SEPARATOR = '/'
"""TensorScope is the basic variable scope.
Examples
--------
>>> with TensorScope('conv1'): a = Tensor('weights')
>>> a.name
>>> conv1/weight
>>> import dragon
>>> with dragon.name_scope('conv1'): a = Tensor('weights')
>>> a.name
>>> conv1/weight
"""
def __init__(self, prefix):
assert isinstance(prefix, type('str')), \
"TensorScope takes in a string as its argument."
self.prefix = prefix + TensorScope.SEPARATOR
self.prefix = prefix + SEPARATOR
def __enter__(self):
global TENSOR_SCOPE
......@@ -49,7 +115,20 @@ class TensorScope(object):
assert TENSOR_SCOPE.endswith(self.prefix)
TENSOR_SCOPE = TENSOR_SCOPE[:-len(self.prefix)]
class PhaseScope(object):
"""PhaseScope is a auxiliary to assign the specific phase.
Examples
--------
>>> import dragon.vm.theano as theano
>>> a = ops.RandomUniform([2, 3])
>>> with PhaseScope(phase='train'): f = theano.function(outputs=a)
>>> import dragon
>>> with dragon.phase_scope(phase='test'): f = theano.function(outputs=a)
"""
def __init__(self, phase):
assert isinstance(phase, type('str')), \
"PhaseScope takes in a string as its argument."
......@@ -64,7 +143,18 @@ class PhaseScope(object):
assert PHASE_SCOPE == self.phase
PHASE_SCOPE = ''
class DeviceScope(object):
"""DeviceScope is a auxiliary to assign the specific device.
Examples
--------
>>> with DeviceScope(device='cpu'): a = ops.RandomUniform([2, 3])
>>> import dragon
>>> with dragon.device_scope(device='gpu', id=0, use_cudnn=True): a = ops.RandomUniform([2, 3])
"""
def __init__(self, device, id=0, use_cudnn=True):
self.device = device.lower()
self.engine = 'CUDNN' if use_cudnn else 'DRAGON'
......@@ -77,8 +167,7 @@ class DeviceScope(object):
DEVICE_SCOPE = '/' + self.device + ':' + str(self.id)
ENGINE_SCOPE = self.engine
def __exit__(self, type, value, traceback):
global DEVICE_SCOPE, ENGINE_SCOPE
DEVICE_SCOPE = ''
ENGINE_SCOPE = ''
ENGINE_SCOPE = ''
\ No newline at end of file
Dragon/python/dragon/docs/Makefile 0 → 100644
# Makefile for Sphinx documentation
#
# You can set these variables from the command line.
SPHINXOPTS =
SPHINXBUILD = sphinx-build
PAPER =
BUILDDIR = _build
# User-friendly check for sphinx-build
ifeq ($(shell which $(SPHINXBUILD) >/dev/null 2>&1; echo $$?), 1)
$(error The '$(SPHINXBUILD)' command was not found. Make sure you have Sphinx installed, then set the SPHINXBUILD environment variable to point to the full path of the '$(SPHINXBUILD)' executable. Alternatively you can add the directory with the executable to your PATH. If you don't have Sphinx installed, grab it from http://sphinx-doc.org/)
endif
# Internal variables.
PAPEROPT_a4 = -D latex_paper_size=a4
PAPEROPT_letter = -D latex_paper_size=letter
ALLSPHINXOPTS = -d $(BUILDDIR)/doctrees $(PAPEROPT_$(PAPER)) $(SPHINXOPTS) .
# the i18n builder cannot share the environment and doctrees with the others
I18NSPHINXOPTS = $(PAPEROPT_$(PAPER)) $(SPHINXOPTS) .
.PHONY: help clean html dirhtml singlehtml pickle json htmlhelp qthelp devhelp epub latex latexpdf text man changes linkcheck doctest coverage gettext
help:
@echo "Please use \`make <target>' where <target> is one of"
@echo " html to make standalone HTML files"
@echo " dirhtml to make HTML files named index.html in directories"
@echo " singlehtml to make a single large HTML file"
@echo " pickle to make pickle files"
@echo " json to make JSON files"
@echo " htmlhelp to make HTML files and a HTML help project"
@echo " qthelp to make HTML files and a qthelp project"
@echo " applehelp to make an Apple Help Book"
@echo " devhelp to make HTML files and a Devhelp project"
@echo " epub to make an epub"
@echo " latex to make LaTeX files, you can set PAPER=a4 or PAPER=letter"
@echo " latexpdf to make LaTeX files and run them through pdflatex"
@echo " latexpdfja to make LaTeX files and run them through platex/dvipdfmx"
@echo " text to make text files"
@echo " man to make manual pages"
@echo " texinfo to make Texinfo files"
@echo " info to make Texinfo files and run them through makeinfo"
@echo " gettext to make PO message catalogs"
@echo " changes to make an overview of all changed/added/deprecated items"
@echo " xml to make Docutils-native XML files"
@echo " pseudoxml to make pseudoxml-XML files for display purposes"
@echo " linkcheck to check all external links for integrity"
@echo " doctest to run all doctests embedded in the documentation (if enabled)"
@echo " coverage to run coverage check of the documentation (if enabled)"
clean:
rm -rf $(BUILDDIR)/*
html:
$(SPHINXBUILD) -b html $(ALLSPHINXOPTS) $(BUILDDIR)/html
@echo
@echo "Build finished. The HTML pages are in $(BUILDDIR)/html."
dirhtml:
$(SPHINXBUILD) -b dirhtml $(ALLSPHINXOPTS) $(BUILDDIR)/dirhtml
@echo
@echo "Build finished. The HTML pages are in $(BUILDDIR)/dirhtml."
singlehtml:
$(SPHINXBUILD) -b singlehtml $(ALLSPHINXOPTS) $(BUILDDIR)/singlehtml
@echo
@echo "Build finished. The HTML page is in $(BUILDDIR)/singlehtml."
pickle:
$(SPHINXBUILD) -b pickle $(ALLSPHINXOPTS) $(BUILDDIR)/pickle
@echo
@echo "Build finished; now you can process the pickle files."
json:
$(SPHINXBUILD) -b json $(ALLSPHINXOPTS) $(BUILDDIR)/json
@echo
@echo "Build finished; now you can process the JSON files."
htmlhelp:
$(SPHINXBUILD) -b htmlhelp $(ALLSPHINXOPTS) $(BUILDDIR)/htmlhelp
@echo
@echo "Build finished; now you can run HTML Help Workshop with the" \
".hhp project file in $(BUILDDIR)/htmlhelp."
qthelp:
$(SPHINXBUILD) -b qthelp $(ALLSPHINXOPTS) $(BUILDDIR)/qthelp
@echo
@echo "Build finished; now you can run "qcollectiongenerator" with the" \
".qhcp project file in $(BUILDDIR)/qthelp, like this:"
@echo "# qcollectiongenerator $(BUILDDIR)/qthelp/Dragon.qhcp"
@echo "To view the help file:"
@echo "# assistant -collectionFile $(BUILDDIR)/qthelp/Dragon.qhc"
applehelp:
$(SPHINXBUILD) -b applehelp $(ALLSPHINXOPTS) $(BUILDDIR)/applehelp
@echo
@echo "Build finished. The help book is in $(BUILDDIR)/applehelp."
@echo "N.B. You won't be able to view it unless you put it in" \
"~/Library/Documentation/Help or install it in your application" \
"bundle."
devhelp:
$(SPHINXBUILD) -b devhelp $(ALLSPHINXOPTS) $(BUILDDIR)/devhelp
@echo
@echo "Build finished."
@echo "To view the help file:"
@echo "# mkdir -p $$HOME/.local/share/devhelp/Dragon"
@echo "# ln -s $(BUILDDIR)/devhelp $$HOME/.local/share/devhelp/Dragon"
@echo "# devhelp"
epub:
$(SPHINXBUILD) -b epub $(ALLSPHINXOPTS) $(BUILDDIR)/epub
@echo
@echo "Build finished. The epub file is in $(BUILDDIR)/epub."
latex:
$(SPHINXBUILD) -b latex $(ALLSPHINXOPTS) $(BUILDDIR)/latex
@echo
@echo "Build finished; the LaTeX files are in $(BUILDDIR)/latex."
@echo "Run \`make' in that directory to run these through (pdf)latex" \
"(use \`make latexpdf' here to do that automatically)."
latexpdf:
$(SPHINXBUILD) -b latex $(ALLSPHINXOPTS) $(BUILDDIR)/latex
@echo "Running LaTeX files through pdflatex..."
$(MAKE) -C $(BUILDDIR)/latex all-pdf
@echo "pdflatex finished; the PDF files are in $(BUILDDIR)/latex."
latexpdfja:
$(SPHINXBUILD) -b latex $(ALLSPHINXOPTS) $(BUILDDIR)/latex
@echo "Running LaTeX files through platex and dvipdfmx..."
$(MAKE) -C $(BUILDDIR)/latex all-pdf-ja
@echo "pdflatex finished; the PDF files are in $(BUILDDIR)/latex."
text:
$(SPHINXBUILD) -b text $(ALLSPHINXOPTS) $(BUILDDIR)/text
@echo
@echo "Build finished. The text files are in $(BUILDDIR)/text."
man:
$(SPHINXBUILD) -b man $(ALLSPHINXOPTS) $(BUILDDIR)/man
@echo
@echo "Build finished. The manual pages are in $(BUILDDIR)/man."
texinfo:
$(SPHINXBUILD) -b texinfo $(ALLSPHINXOPTS) $(BUILDDIR)/texinfo
@echo
@echo "Build finished. The Texinfo files are in $(BUILDDIR)/texinfo."
@echo "Run \`make' in that directory to run these through makeinfo" \
"(use \`make info' here to do that automatically)."
info:
$(SPHINXBUILD) -b texinfo $(ALLSPHINXOPTS) $(BUILDDIR)/texinfo
@echo "Running Texinfo files through makeinfo..."
make -C $(BUILDDIR)/texinfo info
@echo "makeinfo finished; the Info files are in $(BUILDDIR)/texinfo."
gettext:
$(SPHINXBUILD) -b gettext $(I18NSPHINXOPTS) $(BUILDDIR)/locale
@echo
@echo "Build finished. The message catalogs are in $(BUILDDIR)/locale."
changes:
$(SPHINXBUILD) -b changes $(ALLSPHINXOPTS) $(BUILDDIR)/changes
@echo
@echo "The overview file is in $(BUILDDIR)/changes."
linkcheck:
$(SPHINXBUILD) -b linkcheck $(ALLSPHINXOPTS) $(BUILDDIR)/linkcheck
@echo
@echo "Link check complete; look for any errors in the above output " \
"or in $(BUILDDIR)/linkcheck/output.txt."
doctest:
$(SPHINXBUILD) -b doctest $(ALLSPHINXOPTS) $(BUILDDIR)/doctest
@echo "Testing of doctests in the sources finished, look at the " \
"results in $(BUILDDIR)/doctest/output.txt."
coverage:
$(SPHINXBUILD) -b coverage $(ALLSPHINXOPTS) $(BUILDDIR)/coverage
@echo "Testing of coverage in the sources finished, look at the " \
"results in $(BUILDDIR)/coverage/python.txt."
xml:
$(SPHINXBUILD) -b xml $(ALLSPHINXOPTS) $(BUILDDIR)/xml
@echo
@echo "Build finished. The XML files are in $(BUILDDIR)/xml."
pseudoxml:
$(SPHINXBUILD) -b pseudoxml $(ALLSPHINXOPTS) $(BUILDDIR)/pseudoxml
@echo
@echo "Build finished. The pseudo-XML files are in $(BUILDDIR)/pseudoxml."
Dragon/python/dragon/docs/_extensions/mathmacro.py 0 → 100644
"""Sphinx extension provide a new directive *mathmacro*.
This extension has to be added after the other math extension since it
redefined the math directive and the math role. For example, like this
(in the conf.py file)::
extensions = [
'sphinx.ext.autodoc', 'sphinx.ext.doctest',
'sphinx.ext.mathjax',
'sphinx.ext.viewcode', 'sphinx.ext.autosummary',
'numpydoc',
'mathmacro']
"""
from __future__ import print_function
import re
from docutils.parsers.rst.directives.misc import Replace
from sphinx.ext.mathbase import MathDirective
from sphinx.ext.mathbase import math_role
def multiple_replacer(replace_dict):
"""Return a function replacing doing multiple replacements.
The produced function replace `replace_dict.keys()` by
`replace_dict.values`, respectively.
"""
def replacement_function(match):
s = match.group(0)
end = s[-1]
if re.match(r'[\W_]', end):
return replace_dict[s[:-1]]+end
else:
return replace_dict[s]
pattern = "|".join([re.escape(k)+r'[\W_]'
for k in replace_dict.keys()])
pattern = re.compile(pattern, re.M)
return lambda string: pattern.sub(replacement_function, string)
def multiple_replace(string, replace_dict):
mreplace = multiple_replacer(replace_dict)
return mreplace(string)
class MathMacro(Replace):
"""Directive defining a math macro."""
def run(self):
if not hasattr(self.state.document, 'math_macros'):
self.state.document.math_macros = {}
latex_key = '\\'+self.state.parent.rawsource.split('|')[1]
self.state.document.math_macros[latex_key] = ''.join(self.content)
self.content[0] = ':math:`'+self.content[0]
self.content[-1] = self.content[-1]+'`'
return super(MathMacro, self).run()
class NewMathDirective(MathDirective):
"""New math block directive parsing the latex code."""
def run(self):
try:
math_macros = self.state.document.math_macros
except AttributeError:
pass
else:
if math_macros:
multiple_replace = multiple_replacer(math_macros)
for i, c in enumerate(self.content):
self.content[i] = multiple_replace(c)
for i, a in enumerate(self.arguments):
self.arguments[i] = multiple_replace(a)
return super(NewMathDirective, self).run()
def new_math_role(role, rawtext, text, lineno, inliner,
options={}, content=[]):
"""New math role parsing the latex code."""
try:
math_macros = inliner.document.math_macros
except AttributeError:
pass
else:
if math_macros:
rawtext = multiple_replace(rawtext, math_macros)
text = rawtext.split('`')[1]
return math_role(role, rawtext, text, lineno, inliner,
options=options, content=content)
def setup(app):
app.add_role('math', new_math_role)
app.add_directive('math', NewMathDirective)
app.add_directive('mathmacro', MathMacro)
\ No newline at end of file
Dragon/python/dragon/docs/_templates/layout.html 0 → 100644
{% extends "basic/layout.html" %}
<!----------------------- Defines ----------------------->
{% if theme_bootstrap_version == "3" %}
{% set bootstrap_version, navbar_version = "3.3.7", "" %}
{% set bs_span_prefix = "col-md-" %}
{% else %}
{% set bootstrap_version, navbar_version = "2.3.2", "-2" %}
{% set bs_span_prefix = "span" %}
{% endif %}
{%- set render_sidebar = (not embedded) and (not theme_nosidebar|tobool) and sidebars %}
{%- set bs_content_width = render_sidebar and "9" or "12"%}
<!----------------------- CSS/JS ----------------------->
{% set css_files = css_files + [
'_static/css/dragon.css',
'_static/fonts/font-awesome/css/font-awesome.min.css',
'_static/fonts/lato/css/latofonts.css'
]
%}
{% set script_files = script_files + [
'_static/js/jquery-1.11.0.min.js',
'_static/js/jquery-fix.js',
'_static/bootstrap-' + bootstrap_version + '/js/bootstrap.min.js',
'_static/js/bootstrap-sphinx.js',
]
%}
{%- if render_sidebar %}
{% set script_files = script_files + ['_static/js/sidebar.js']%}
{%- endif %}
<!----------------------- Macros ----------------------->
{%- macro bsidebar() %}
{%- if render_sidebar %}
<div class="sphinxsidebar leftsidebar" role="navigation" aria-label="main navigation">
<div class="sphinxsidebarwrapper">
{%- block sidebartoc %}
{%- include "localtoc.html" %}
{%- endblock %}
</div>
</div>
{%- endif %}
{%- endmacro %}
<!----------------------- Blocks ----------------------->
{%- block doctype -%}
<!DOCTYPE html>
{%- endblock %}
{%- block extrahead %}
<meta charset='utf-8'>
<meta http-equiv='X-UA-Compatible' content='IE=edge,chrome=1'>
<meta name='viewport' content='width=device-width, initial-scale=1.0, maximum-scale=1'>
<meta name="apple-mobile-web-app-capable" content="yes">
{% endblock %}
{# Silence the sidebar's, relbar's #}
{% block header %}{% endblock %}
{% block relbar1 %}{% endblock %}
{% block relbar2 %}{% endblock %}
{% block sidebarsourcelink %}{% endblock %}
{% block content %}
{% include "navbar.html" %}
<div class="container doc-container">
<div class="row">
{% block sidebar1 %}{{ bsidebar() }}{% endblock %}
<div class="content">
{% block body %}{% endblock %}
</div>
</div>
</div>
{% endblock %}
Dragon/python/dragon/docs/_templates/localtoc.html 0 → 100644
{{ toctree(maxdepth=theme_globaltoc_depth|toint, collapse=True,includehidden=theme_globaltoc_includehidden|tobool) }}
\ No newline at end of file
Dragon/python/dragon/docs/_templates/navbar.html 0 → 100644
{# Import the theme's layout. #}
<div id="navbar" class="{{ theme_navbar_class }} navbar-default {% if theme_navbar_fixed_top|tobool -%} navbar-fixed-top{%- endif -%}">
<div class="container">
<div class="navbar-header">
<button type="button" class="navbar-toggle" data-toggle="collapse" data-target=".nav-collapse">
<span class="icon-bar"></span>
<span class="icon-bar"></span>
<span class="icon-bar"></span>
</button>
<a class="navbar-brand" href="{{ pathto("../../index") }}">
{%- block sidebarlogo %}
{%- if logo %}<span><img src="{{ pathto('_static/images/' + logo, 1) }}"></span>{%- endif %}
{%- endblock %}
{% if theme_navbar_title -%}{{ theme_navbar_title|e }}{%- else -%}{{ project|e }}{%- endif -%}
</a>
</div>
<div class="collapse navbar-collapse nav-collapse">
<ul class="nav navbar-nav">
<!-- Install -->
<li><a href="{{ pathto("../../helper/install") }}">Install</a></li>
<!-- Github -->
<li><a href="https://github.com/neopenx/Dragon">Github</a></li>
<!-- API -->
<li class="dropdown globaltoc-container">
<a role="button" id="dLabelGlobalToc" data-toggle="dropdown" data-target="#" href="#" aria-expanded="true">API<b class="caret"></b></a>
<ul class="dropdown-menu globaltoc" role="menu" aria-labelledby="dLabelGlobalToc">
<li class="toctree-l1 dropdown-submenu"><a class="reference internal" href="{{ pathto(master_doc) }}" tabindex="-1">Python</a><li>
</ul>
</li>
</ul>
{% block navbarsearch %}
{% include "navbarsearchbox.html" %}
{% endblock %}
</div>
</div>
</div>
Dragon/python/dragon/docs/_templates/search.html 0 → 100644
{%- extends "layout.html" %}
{% if theme_bootstrap_version == "3" %}
{% set bootstrap_version, navbar_version = "3.3.7", "" %}
{% set bs_span_prefix = "col-md-" %}
{% else %}
{% set bootstrap_version, navbar_version = "2.3.2", "-2" %}
{% set bs_span_prefix = "span" %}
{% endif %}
{%- set render_sidebar = False %}
{% set title = _('Search') %}
{% set script_files = script_files + ['_static/searchtools.js'] %}
{% block extrahead %}
<script type="text/javascript" xmlns="http://www.w3.org/1999/html">
jQuery(function () { Search.loadIndex("{{ pathto('searchindex.js', 1) }}"); });
</script>
{# this is used when loading the search index using $.ajax fails,
such as on Chrome for documents on localhost #}
<script type="text/javascript" id="searchindexloader"></script>
{{ super() }}
{% endblock %}
{% block content %}
{% include "navbar.html" %}
<div class="container doc-container">
<div class="col-lg-2"></div>
<div class="col-lg-8">
{% block sidebar1 %}{{ bsidebar() }}{% endblock %}
<div class="content" style="width: 100%">
<h1 id="search-documentation">{{ _('Search') }}</h1>
<div id="fallback" class="admonition warning">
<script type="text/javascript">$('#fallback').hide();</script>
<p>
{% trans %}Please activate JavaScript to enable the search
functionality.{% endtrans %}
</p>
</div>
<p>
From here you can search these documents. Enter your search words into the box below and click <strong>Search</strong>.
</p>
{% if theme_bootstrap_version == "3" %}
<form class="form-inline" action="" method="get">
<div class="form-group">
<input type="text" class="form-control" name="q" value="" />
</div>
<input type="submit" class="btn btn-default" value="Search" />
<span id="search-progress" style="padding-left: 10px"></span>
</form>
{% else %}
<form class="form-search">
<input type="text" class="input-medium search-query" name="q" value="" />
<input type="submit" class="btn btn-default" value="{{ _('search') }}" />
<span id="search-progress" style="padding-left: 10px"></span>
</form>
{% endif %}
{% if search_performed %}
<h2>{{ _('Search Results') }}</h2>
{% if not search_results %}
<p>{{ _('Your search did not match any documents. Please make sure that all words are spelled correctly and that you\'ve selected enough categories.') }}</p>
{% endif %}
{% endif %}
<div id="search-results">
{% if search_results %}
<ul>
{% for href, caption, context in search_results %}
<li>
<a href="{{ pathto(item.href) }}">{{ caption }}</a>
<div class="context">{{ context|e }}</div>
</li>
{% endfor %}
</ul>
{% endif %}
</div>
</div>
</div>
<div class="col-lg-2"></div>
</div>
{% endblock %}
Dragon/python/dragon/docs/conf.py 0 → 100644
# --------------------------------------------------------
# Dragon
# Copyright(c) 2017 SeetaTech
# Written by Ting Pan
# --------------------------------------------------------
import sys
import os
import sphinx_bootstrap_theme
# basic
html_static_path = ['_static']
templates_path = ['_templates']
exclude_patterns = ['_build']
source_suffix = '.rst'
master_doc = 'index'
pygments_style = 'sphinx'
todo_include_todos = True
# extensions
sys.path.insert(0, os.path.abspath('_extensions'))
extensions = [
'sphinx.ext.autodoc',
'sphinx.ext.mathjax',
'sphinx.ext.napoleon',
#'sphinx.ext.viewcode',
'mathmacro',
]
# project
project = ''
copyright = '2017, Ting Pan'
author = 'Ting Pan'
html_logo = "dragon.png"
html_title = ""
html_short_title = ""
html_favicon = 'images/favicon.png'
version = ''
release = ''
language = None
# theme
html_theme = 'bootstrap'
html_theme_path = sphinx_bootstrap_theme.get_html_theme_path()
html_show_sourcelink = False
html_show_sphinx = False
html_show_copyright = False
html_theme_options = {
'globaltoc_depth': -1,
'navbar_class': "navbar navbar-inverse",
'navbar_fixed_top': "true",
'bootswatch_theme': "yeti",
}
html_sidebars = {'index': ['localtoc.html'],
'install': ['localtoc.html'],
'contents/**': ['localtoc.html']}
# overloads
def setup(app):
app.config.values['autodoc_member_order'] = ('bysource', True)
\ No newline at end of file
Dragon/python/dragon/docs/contents/config.rst 0 → 100644
====================
:mod:`dragon.config`
====================
.. toctree::
:hidden:
Quick Shortcut
--------------
==================== =============================================================================
List Brief
==================== =============================================================================
`EnableCPU`_ Enable CPU mode globally.
`EnableCUDA`_ Enable CUDA mode globally.
`SetRandomSeed`_ Set the global random seed.
`GetRandomSeed`_ Get the global random seed.
`SetGPU`_ Set the global id GPU.
`GetGPU`_ Get the global id of GPU.
`SetDebugMode`_ Enable Debug mode globally.
`SetLoggingLevel`_ Set the minimum level of Logging.
==================== =============================================================================
API Reference
-------------
.. automodule:: dragon.config
:members:
.. _EnableCPU: #dragon.config.EnableCPU
.. _EnableCUDA: #dragon.config.EnableCUDA
.. _SetRandomSeed: #dragon.config.SetRandomSeed
.. _GetRandomSeed: #dragon.config.GetRandomSeed
.. _SetGPU: #dragon.config.SetGPU
.. _GetGPU: #dragon.config.GetGPU
.. _SetDebugMode: #dragon.config.SetDebugMode
.. _SetLoggingLevel: #dragon.config.SetLoggingLevel
\ No newline at end of file
Dragon/python/dragon/docs/contents/core.rst 0 → 100644
==================
:mod:`dragon.core`
==================
Data Structure
--------------
.. toctree::
:hidden:
core/tensor
core/scope
============================== =======================================================================
List Brief
============================== =======================================================================
`dragon.core.scope`_ The Scope and Namespace.
`dragon.core.tensor`_ The basic structure of VM.
============================== =======================================================================
C++ Binding Wrapper
-------------------
.. toctree::
:hidden:
core/workspace
core/mpi
core/gradient_maker
============================== =======================================================================
List Brief
============================== =======================================================================
`dragon.core.workspace`_ The interfaces of Workspace, mostly are the wrappers of C++.
`dragon.core.gradient_maker`_ The generator of GradientOps.
`dragon.core.mpi`_ The MPI utilities.
============================== =======================================================================
.. _dragon.core.mpi: core/mpi.html
.. _dragon.core.scope: core/scope.html
.. _dragon.core.tensor: core/tensor.html
.. _dragon.core.workspace: core/workspace.html
.. _dragon.core.gradient_maker: core/gradient_maker.html
\ No newline at end of file
Dragon/python/dragon/docs/contents/core/gradient_maker.rst 0 → 100644
====================
:mod:`GradientMaker`
====================
.. toctree::
:hidden:
.. currentmodule:: dragon.core.gradient_maker
.. autoclass:: GraphGradientMaker
:members:
.. _theano.function(*args, **kwargs): ../vm/theano/compile.html#dragon.vm.theano.compile.function.function
\ No newline at end of file
Dragon/python/dragon/docs/contents/core/mpi.rst 0 → 100644
==========
:mod:`MPI`
==========
.. toctree::
:hidden:
Basic
-----
============================== =============================================================================
List Brief
============================== =============================================================================
`Init`_ Init the MPI env.
`Is_Init`_ Whether the MPI env has initialized.
`Rank`_ The world rank of current MPI node.
`Size`_ The world size of current MPI env.
`CreateGroup`_ Construct a MPIGroup with specific members.
`Finalize`_ Finalize the MPI env.
============================== =============================================================================
Parallelism
-----------
============================== =============================================================================
List Brief
============================== =============================================================================
`Snapshot`_ Set the specific MPI nodes to snapshot.
`Parallel`_ Set the specific MPI nodes for data parallelism.
`AllowSnapshot`_ Whether this node can snapshot.
`AllowParallel`_ Whether this node was set for data parallelism.
`SetParallelMode`_ Set the mode of data parallelism.
`GetParallelMode`_ Get the current mode of data parallelism.
============================== =============================================================================
.. automodule:: dragon.core.mpi
:members:
.. _Init: #dragon.core.mpi.Init
.. _Is_Init: #dragon.core.mpi.Is_Init
.. _Rank: #dragon.core.mpi.Rank
.. _Size: #dragon.core.mpi.Size
.. _CreateGroup: #dragon.core.mpi.CreateGroup
.. _Finalize: #dragon.core.mpi.Finalize
.. _Snapshot: #dragon.core.mpi.Snapshot
.. _Parallel: #dragon.core.mpi.Parallel
.. _AllowSnapshot: #dragon.core.mpi.AllowSnapshot
.. _AllowParallel: #dragon.core.mpi.AllowParallel
.. _SetParallelMode: #dragon.core.mpi.SetParallelMode
.. _GetParallelMode: #dragon.core.mpi.GetParallelMode
.. _workspace.Snapshot(*args, **kwargs): workspace.html#dragon.core.workspace.Snapshot
\ No newline at end of file
Dragon/python/dragon/docs/contents/core/scope.rst 0 → 100644
============
:mod:`Scope`
============
.. toctree::
:hidden:
.. currentmodule:: dragon.core.scope
.. autoclass:: dragon.core.scope.TensorScope
:members:
.. autoclass:: dragon.core.scope.PhaseScope
:members:
.. autoclass:: dragon.core.scope.DeviceScope
:members:
Dragon/python/dragon/docs/contents/core/tensor.rst 0 → 100644
=============
:mod:`Tensor`
=============
.. toctree::
:hidden:
Quick Shortcut
--------------
============================== =============================================================================
List Brief
============================== =============================================================================
`Tensor.name`_ Return or Set the name.
`Tensor.shape`_ Return or Set the shape.
`Tensor.dtype`_ Return or Set the data type.
`Tensor.set_value`_ Feed the values to C++ backend.
`Tensor.get_value`_ Fetch the values from C++ backend.
`Tensor.copy`_ Return a Tensor with same content.
`Tensor.reshape`_ Reshape the dimensions of input.
`Tensor.dimshuffle`_ Shuffle the dimen`sions.
`Tensor.CreateOperator`_ Construct a new Tensor with specific operator descriptor.
`Tensor.Fill`_ Fill self with the specific type of filler.
`Tensor.PrintExpressions`_ Return the stringified internal expressions.
============================== =============================================================================
Register
--------
============================== =============================================================================
List Brief
============================== =============================================================================
`Tensor.Variable`_ Register as an empty variable.
`Tensor.Placeholder`_ Register as a placeholder.
`Tensor.Constant`_ Register as a variable with constant initializer.
`Tensor.Uniform`_ Register as a variable with uniform initializer.
`Tensor.Normal`_ Register as a variable with normal initializer.
`Tensor.TruncatedNormal`_ Register as a variable with truncated normal initializer.
`Tensor.Gaussian`_ Register as a variable with gaussian initializer.
`Tensor.Xavier`_ Register as a variable with xavier initializer.
`Tensor.MSRA`_ Register as a variable with msra initializer.
`Tensor.GlorotUniform`_ Register as a variable with glorot uniform initializer.
`Tensor.GlorotNormal`_ Register as a variable with glorot normal initializer.
============================== =============================================================================
Override
--------
============================== =============================================================================
List Brief
============================== =============================================================================
`Tensor.__add__`_ x.__add__(y) <=> x + y
`Tensor.__radd__`_ x.__radd__(y) <=> y + x
`Tensor.__sub__`_ x.__sub__(y) <=> x - y
`Tensor.__rsub__`_ x.__rsub__(y) <=> y - x
`Tensor.__mul__`_ x.__mul__(y) <=> x * y
`Tensor.__rmul__`_ x.__rmul__(y) <=> y * x
`Tensor.__div__`_ x.__div__(y) <=> x / y
`Tensor.__rdiv__`_ x.__rdiv__(y) <=> y / x
`Tensor.__neg__`_ x.__neg__() <=> -x
`Tensor.__str__`_ Return the information(name/shape).
`Tensor.__getitem__`_ Return a Tensor with specific indices.
`Tensor.__call__`_ Print the expressions.
============================== =============================================================================
API Reference
-------------
.. currentmodule:: dragon.core.tensor
.. autoclass:: Tensor
:members:
.. automethod:: __init__
.. _Tensor.Variable: #dragon.core.tensor.Tensor.Variable
.. _Tensor.Placeholder: #dragon.core.tensor.Tensor.Placeholder
.. _Tensor.Constant: #dragon.core.tensor.Tensor.Constant
.. _Tensor.Uniform: #dragon.core.tensor.Tensor.Uniform
.. _Tensor.Normal: #dragon.core.tensor.Tensor.Normal
.. _Tensor.TruncatedNormal: #dragon.core.tensor.Tensor.TruncatedNormal
.. _Tensor.Gaussian: #dragon.core.tensor.Tensor.Gaussian
.. _Tensor.Xavier: #dragon.core.tensor.Tensor.Xavier
.. _Tensor.MSRA: #dragon.core.tensor.Tensor.MSRA
.. _Tensor.GlorotUniform: #dragon.core.tensor.Tensor.GlorotUniform
.. _Tensor.GlorotNormal: #dragon.core.tensor.Tensor.GlorotNormal
.. _Tensor.__add__: #dragon.core.tensor.Tensor.__add__
.. _Tensor.__radd__: #dragon.core.tensor.Tensor.__radd__
.. _Tensor.__sub__: #dragon.core.tensor.Tensor.__sub__
.. _Tensor.__rsub__: #dragon.core.tensor.Tensor.__rsub__
.. _Tensor.__mul__: #dragon.core.tensor.Tensor.__mul__
.. _Tensor.__rmul__: #dragon.core.tensor.Tensor.__rmul__
.. _Tensor.__div__: #dragon.core.tensor.Tensor.__div__
.. _Tensor.__rdiv__: #dragon.core.tensor.Tensor.__rdiv__
.. _Tensor.__neg__: #dragon.core.tensor.Tensor.__neg__
.. _Tensor.__str__: #dragon.core.tensor.Tensor.__str__
.. _Tensor.__getattr__: #dragon.core.tensor.Tensor.__getattr__
.. _Tensor.__getitem__: #dragon.core.tensor.Tensor.__getitem__
.. _Tensor.__call__: #dragon.core.tensor.Tensor.__call__
.. _Tensor.name: #dragon.core.tensor.Tensor.name
.. _Tensor.shape: #dragon.core.tensor.Tensor.shape
.. _Tensor.dtype: #dragon.core.tensor.Tensor.dtype
.. _Tensor.set_value: #dragon.core.tensor.Tensor.set_value
.. _Tensor.get_value: #dragon.core.tensor.Tensor.get_value
.. _Tensor.copy: #dragon.core.tensor.Tensor.copy
.. _Tensor.reshape: #dragon.core.tensor.Tensor.reshape
.. _Tensor.dimshuffle: #dragon.core.tensor.Tensor.dimshuffle
.. _Tensor.CreateOperator: #dragon.core.tensor.Tensor.CreateOperator
.. _Tensor.Fill: #dragon.core.tensor.Tensor.Fill
.. _Tensor.PrintExpressions: #dragon.core.tensor.Tensor.PrintExpressions
.. _workspace.FeedTensor(*args, **kwargs): workspace.html#dragon.core.workspace.FeedTensor
.. _workspace.FetchTensor(*args, **kwargs): workspace.html#dragon.core.workspace.FetchTensor
.. _ops.Copy(*args, **kwargs): ../operators/control_flow.html#dragon.operators.control_flow.Copy
Dragon/python/dragon/docs/contents/core/workspace.rst 0 → 100644
================
:mod:`Workspace`
================
.. toctree::
:hidden:
Tensor
------
============================== =============================================================================
List Brief
============================== =============================================================================
`HasTensor`_ Query whether tensor has registered in current workspace.
`GetTensorName`_ Query the name represented in current workspace.
`CreateFiller`_ Create the filler in the backend.
`FetchTensor`_ Fetch the values of given tensor.
`FeedTensor`_ Feed the values to the given tensor.
============================== =============================================================================
Graph
-----
============================== =============================================================================
List Brief
============================== =============================================================================
`CreateGraph`_ Create the graph in the backend.
`RunGraph`_ Run the specific graph.
============================== =============================================================================
Misc
----
============================== =============================================================================
List Brief
============================== =============================================================================
`Snapshot`_ Snapshot tensors into a binary file.
`Restore`_ Restore tensors from a binary file.
`SwitchWorkspace`_ Switch to the specific Workspace.
`PrintRawGraphDef`_ Print the raw prototxt.
`PrintOptimizedGraph`_ Print the optimized prototxt.
`WriteOptimizedGraph`_ Generate the optimized prototxt into a file.
============================== =============================================================================
API Reference
-------------
.. automodule:: dragon.core.workspace
:members:
:undoc-members:
:show-inheritance:
.. _SwitchWorkspace: #dragon.core.workspace.SwitchWorkspace
.. _CreateGraph: #dragon.core.workspace.CreateGraph
.. _HasTensor: #dragon.core.workspace.HasTensor
.. _GetTensorName: #dragon.core.workspace.GetTensorName
.. _CreateFiller: #dragon.core.workspace.CreateFiller
.. _FetchTensor: #dragon.core.workspace.FetchTensor
.. _FeedTensor: #dragon.core.workspace.FeedTensor
.. _RunGraph: #dragon.core.workspace.RunGraph
.. _Snapshot: #dragon.core.workspace.Snapshot
.. _Restore: #dragon.core.workspace.Restore
.. _PrintRawGraphDef: #dragon.core.workspace.PrintRawGraphDef
.. _PrintOptimizedGraph: #dragon.core.workspace.PrintOptimizedGraph
.. _WriteOptimizedGraph: #dragon.core.workspace.WriteOptimizedGraph
.. _theano.function(*args, **kwargs): ../vm/theano/compile.html#dragon.vm.theano.compile.function.function
\ No newline at end of file
Dragon/python/dragon/docs/contents/io.rst 0 → 100644
================
:mod:`dragon.io`
================
Wrapper
-------
.. toctree::
:hidden:
io/data_batch
========================== =====================================================================
List Brief
========================== =====================================================================
`dragon.io.data_batch`_ Efficient I/O based on `LMDB`_.
========================== =====================================================================
Component
---------
.. toctree::
:hidden:
io/data_reader
io/data_transformer
io/blob_fetcher
============================== =====================================================================
List Brief
============================== =====================================================================
`dragon.io.data_reader`_ Queue encoded string from `LMDB`_.
`dragon.io.data_transformer`_ Queue transformed images from `DataReader`_.
`dragon.io.blob_fetcher`_ Queue blobs from `DataTransformer`_.
============================== =====================================================================
.. _LMDB: http://lmdb.readthedocs.io/en/release
.. _DataReader: io/data_reader.html#dragon.io.data_reader
.. _DataTransformer: io/data_transformer.html#dragon.io.data_transformer
.. _dragon.io.data_batch: io/data_batch.html
.. _dragon.io.data_reader: io/data_reader.html
.. _dragon.io.data_transformer: io/data_transformer.html
.. _dragon.io.blob_fetcher: io/blob_fetcher.html
\ No newline at end of file
Dragon/python/dragon/docs/contents/io/blob_fetcher.rst 0 → 100644
==================
:mod:`BlobFetcher`
==================
.. toctree::
:hidden:
.. currentmodule:: dragon.io.blob_fetcher
.. autoclass:: BlobFetcher
:members:
.. automethod:: __init__
.. _DataTransformer: data_transformer.html#dragon.io.data_transformer
\ No newline at end of file
Dragon/python/dragon/docs/contents/io/data_batch.rst 0 → 100644
================
:mod:`DataBatch`
================
.. toctree::
:hidden:
.. currentmodule:: dragon.io.data_batch
.. autoclass:: DataBatch
:members:
.. automethod:: __init__
\ No newline at end of file
Dragon/python/dragon/docs/contents/io/data_reader.rst 0 → 100644
=================
:mod:`DataReader`
=================
.. toctree::
:hidden:
.. currentmodule:: dragon.io.data_reader
.. autoclass:: DataReader
:members:
.. automethod:: __init__
.. _LMDB: http://lmdb.readthedocs.io/en/release
\ No newline at end of file
Dragon/python/dragon/docs/contents/io/data_transformer.rst 0 → 100644
======================
:mod:`DataTransformer`
======================
.. toctree::
:hidden:
.. currentmodule:: dragon.io.data_transformer
.. autoclass:: DataTransformer
:members:
.. automethod:: __init__
.. _DataReader: data_reader.html#dragon.io.data_reader
\ No newline at end of file
Dragon/python/dragon/docs/contents/memonger.rst 0 → 100644
======================
:mod:`dragon.memonger`
======================
.. toctree::
:hidden:
Quick Shortcut
--------------
==================== =============================================================================
List Brief
==================== =============================================================================
`ShareGrads`_ Enable gradients sharing globally.
`Drop`_ Drop(Share) the inputs for outputs.
==================== =============================================================================
API Reference
-------------
.. automodule:: dragon.memonger
:members:
.. _ShareGrads: #dragon.memonger.ShareGrads
.. _Drop: #dragon.memonger.Drop
\ No newline at end of file
Dragon/python/dragon/docs/contents/operators.rst 0 → 100644
=======================
:mod:`dragon.operators`
=======================
Common
------
.. toctree::
:hidden:
operators/data
operators/initializer
operators/arithmetic
operators/ndarray
operators/control_flow
operators/misc
operators/cast
operators/mpi
=================================== =====================================================================
List Brief
=================================== =====================================================================
`dragon.operators.data`_ The data preparing operators.
`dragon.operators.initializer`_ The initializing operators.
`dragon.operators.arithmetic`_ The arithmetic operators.
`dragon.operators.ndarray`_ The ndarray operators.
`dragon.operators.control_flow`_ The control flow operators.
`dragon.operators.misc`_ The misc operators.
`dragon.operators.cast`_ The cast operators.
`dragon.operators.mpi`_ The MPI operators.
=================================== =====================================================================
Neural Networks
---------------
.. toctree::
:hidden:
operators/norm
operators/activation
operators/vision
operators/recurrent
operators/loss
=================================== =====================================================================
List Brief
=================================== =====================================================================
`dragon.operators.norm`_ The normalization operators.
`dragon.operators.activation`_ The activation operators.
`dragon.operators.vision`_ The vision operators.
`dragon.operators.recurrent`_ The recurrent operators.
`dragon.operators.loss`_ The loss operators.
=================================== =====================================================================
Custom
------
.. toctree::
:hidden:
operators/custom/minibatch
operators/custom/data_process
operators/custom/vec_mult
========================================= =====================================================================
List Brief
========================================= =====================================================================
`dragon.operators.custom.minibatch`_ How to form a minibatch based on `dragon.io`_ package.
`dragon.operators.custom.data_process`_ How to custom a RunOp for data processing.
`dragon.operators.custom.vec_mult`_ How to custom a TemplateOp for Vector Multiplication.
========================================= =====================================================================
.. _dragon.operators.data: operators/data.html
.. _dragon.operators.initializer: operators/initializer.html
.. _dragon.operators.arithmetic: operators/arithmetic.html
.. _dragon.operators.ndarray: operators/ndarray.html
.. _dragon.operators.control_flow: operators/control_flow.html
.. _dragon.operators.misc: operators/misc.html
.. _dragon.operators.cast: operators/cast.html
.. _dragon.operators.mpi: operators/mpi.html
.. _dragon.operators.activation: operators/activation.html
.. _dragon.operators.vision: operators/vision.html
.. _dragon.operators.recurrent: operators/recurrent.html
.. _dragon.operators.loss: operators/loss.html
.. _dragon.operators.norm: operators/norm.html
.. _dragon.io: io.html
.. _dragon.operators.custom.minibatch: operators/custom/minibatch.html
.. _dragon.operators.custom.data_process: operators/custom/data_process.html
.. _dragon.operators.custom.vec_mult: operators/custom/vec_mult.html
\ No newline at end of file
Dragon/python/dragon/docs/contents/operators/activation.rst 0 → 100644
=================
:mod:`Activation`
=================
.. toctree::
:hidden:
.. automodule:: dragon.operators.activation
:members:
.. |sigmoid_function| mathmacro:: \, y = \frac{1}{1 + {e}^{-x}}
.. |tanh_function| mathmacro:: \, y = \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}
.. |relu_function| mathmacro:: \, y = \max(0, x)
.. |elu_function| mathmacro:: \, y = \left\{ \begin{array} \\ x & & (x > 0) \\ Alpha * (e^{x} - 1) & & (x <= 0) \\ \end{array} \right.
.. |leaky_relu_function| mathmacro:: \, y = \max(x, 0) + Slope * \min(x, 0)
.. |dropout_function| mathmacro:: \, y = x * Bernoulli(p=1 - prob)
.. |softmax_function| mathmacro:: \, y = \frac{e^{x_{i}}}{\sum e^{x_{j}}}
Dragon/python/dragon/docs/contents/operators/arithmetic.rst 0 → 100644
=================
:mod:`Arithmetic`
=================
.. toctree::
:hidden:
.. automodule:: dragon.operators.arithmetic
:members:
.. |power_function| mathmacro:: \\ \, y = [(Scale * x) + Shift]^{Power}
.. |scale_function| mathmacro:: \\ [Axis, Axis + NumAxes)
.. |gram_matrix_function| mathmacro:: \\ \, y = xx^{T}
\ No newline at end of file
Dragon/python/dragon/docs/contents/operators/cast.rst 0 → 100644
===========
:mod:`Cast`
===========
.. toctree::
:hidden:
.. automodule:: dragon.operators.cast
:members:
\ No newline at end of file
Dragon/python/dragon/docs/contents/operators/control_flow.rst 0 → 100644
===================
:mod:`Control Flow`
===================
.. toctree::
:hidden:
.. automodule:: dragon.operators.control_flow
:members:
\ No newline at end of file
Dragon/python/dragon/docs/contents/operators/custom/data_process.rst 0 → 100644
==================
:mod:`DataProcess`
==================
.. toctree::
:hidden:
.. currentmodule:: dragon.operators.custom.data_process
.. autoclass:: DataProcessOp
:members:
\ No newline at end of file
Dragon/python/dragon/docs/contents/operators/custom/minibatch.rst 0 → 100644
================
:mod:`MiniBatch`
================
.. toctree::
:hidden:
.. currentmodule:: dragon.operators.custom.minibatch
.. autoclass:: MiniBatchOp
:members:
.. _dragon.io: ../../io.html
\ No newline at end of file
Dragon/python/dragon/docs/contents/operators/custom/vec_mult.rst 0 → 100644
==============
:mod:`VecMult`
==============
.. toctree::
:hidden:
.. currentmodule:: dragon.operators.custom.vec_mult
.. autoclass:: VecMultOp
:members:
\ No newline at end of file
Dragon/python/dragon/docs/contents/operators/data.rst 0 → 100644
===========
:mod:`Data`
===========
.. toctree::
:hidden:
.. automodule:: dragon.operators.data
:members:
.. _LMDB: http://lmdb.readthedocs.io/en/release
.. _DataBatch: ../io/data_batch.html#dragon.io.data_batch
.. _DataReader: ../io/data_reader.html#dragon.io.data_reader
.. _DataTransformer: ../io/data_transformer.html#dragon.io.data_transformer
.. _BlobFetcher: ../io/blob_fetcher.html#dragon.io.blob_fetcher
\ No newline at end of file
Dragon/src/protos/dragon.pb.cc deleted 100644 → 0
This diff could not be displayed because it is too large.
Dragon/src/protos/dragon.pb.h deleted 100644 → 0
This diff could not be displayed because it is too large.
Dragon/src/utils/op_kernel.cu
This diff could not be displayed because it is too large.
Markdown is supported
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!