#ifndef __MANAGER_H__
#define __MANAGER_H__
+#include "tensor.h"
+#include "tensor_wrap_specs.h"
#ifdef __cplusplus
#include <functional>
#include <memory>
#include <unordered_map>
+#include <utility>
#include <vector>
#include <basic_planner.h>
public:
/**
+ * @brief Tensor Group Type
+ * @note this is not mutually exclusive list, a tensor might be identified as
+ * input as well as output
+ *
+ */
+ enum TensorGroupType {
+ INPUT = 0, /**< Input of an operation */
+ OUTPUT = 1, /**< Output of an operation */
+ WEIGHT = 2, /**< Weight of an operation */
+ TENSORS = 3, /**< Extra states of an operation */
+ };
+
+ constexpr inline static unsigned NUM_TENSOR_GROUP_TYPE =
+ 4; /**< number of tensor group type */
+
+ /**
* @brief Constructor of Manager
*/
Manager() : enable_optimizations(true) {}
}
}
+ /**
+ * @brief request Tensor with weight specification
+ *
+ * @param spec specification
+ * @param identify_as identify as tensor as a group
+ * @return Tensor* tensor
+ */
+ Tensor *requestTensor(const WeightSpecV2 &spec, TensorGroupType identify_as);
+
+ /**
+ * @brief request Tensor with variable + gradient specification
+ *
+ * @param spec specification
+ * @param identify_as identify as tensor as a group
+ * @param exec_order execution order to refer to
+ * @param scope common scope to attach in front of current specification name
+ * @return Tensor* tensor
+ */
+ Tensor *requestTensor(const VarGradSpecV2 &spec, TensorGroupType identify_as,
+ const GraphNode::ExecutionOrder &exec_order,
+ const std::string &scope = "");
+
private:
/** @todo: merge this list to one */
std::vector<std::unique_ptr<Weight>>
std::vector<std::unique_ptr<Var_Grad>>
tensors_v2; /**< extra tensors required by the layers */
+ std::array<std::vector<std::unique_ptr<Var_Grad>>, NUM_TENSOR_GROUP_TYPE>
+ tensor_book; /**< reference to tensor book kept */
+
TensorPool weight_pool; /**< tensor pool to request tensors */
TensorPool tensor_pool; /**< tensor pool to request tensors */
#ifndef __TENSOR_WRAP_SPECS_H__
#define __TENSOR_WRAP_SPECS_H__
+#include <memory>
#include <tuple>
#include <tensor.h>
*
*/
enum class TensorLifespan {
- UNMANAGED = 0b0, /**< tensor with no lifespan, will not be allocated */
- FORWARD_FUNC_LIFESPAN = 0b01, /**< tensor must not be reset before during the
- forward function call, eg. temporary tensors
- needed during forward operations */
- BACKWARD_FUNC_LIFESPAN = 0b10, /**< tensor must not be reset before during the
- backward function call, eg. temporary tensors
- needed during backward operations */
- ITERATION_LIFESPAN = 0b11, /**< tensor must not be reset until the owning
- layer finishes its execution in the current
- iteration, eg. hidden memory/cells of RNN */
- EPOCH_LIFESPAN = 0b111, /**< tensor must not be reset before the epoch ends */
- MAX_LIFESPAN = 0b1111, /**< tensor must not be reset until the end of the
- model execution, eg. layer weights */
+ UNMANAGED = 0b000, /**< tensor with no lifespan, will not be allocated */
+ FORWARD_FUNC_LIFESPAN = 0b001, /**< tensor must not be reset before during the
+ forward function call, eg. temporary tensors
+ needed during forward operations */
+ CALC_DERIV_LIFESPAN = 0b010, /**< must be valid during calcDerivative() */
+ CALC_GRAD_LIFESPAN = 0b100, /**< tensor must be valid during calcGradient() */
+ CALC_GRAD_DERIV_LIFESPAN = 0b110, /**< tensor must not be reset before during
+ the calc_grad and clac_deriv call, eg. temporary
+ tensors needed during backward operations */
+ BACKWARD_FUNC_LIFESPAN =
+ CALC_GRAD_DERIV_LIFESPAN, /**< Alias of CALC_GRAD_DERIV_LIFESPAN */
+ ITERATION_LIFESPAN = 0b111, /**< tensor must not be reset until the owning
+ layer finishes its execution in the current
+ iteration, eg. hidden memory/cells of RNN */
+ EPOCH_LIFESPAN = 0b1111, /**< tensor must be valid before the epoch ends */
+ MAX_LIFESPAN = 0b11111, /**< tensor must not be reset until the end of the
+ model execution, eg. layer weights */
};
/**
TensorLifespan>
VarGradSpec;
+/**
+ * @brief Tensor Specification which describes how this tensor should be
+ * allocated and managed
+ *
+ */
+struct TensorSpecV2 {
+
+ /**
+ * @brief Tensor is being managed by nntrainer, this enum defines how the
+ * value should be recognized inside nntrainer tensor managing scheme.
+ *
+ */
+ enum class RequestType {
+ PLACEHOLDER, /**< Placeholder defines that nntrainer should never care about
+ the memory inside the particualar tensor */
+ UNIQUE, /**< Unique means a simple tensor that will be owned explicitly the
+ current request */
+ READ_ONLY_VIEW, /**< Readonly view defines a view of which ownership of @a
+ underlying memory is at another tensor, also hinting
+ nntrainer that the operation upon this particular tensor
+ will never change value of the underlying memory */
+ MAYBE_MODIFYING_VIEW, /**< Maybe modifying view defines a (possible) view of
+ which ownership of @a underlying memory is at another
+ tensor, while hinting the nntrainer this tensor will do
+ some modification of the underlying memory. nntrainer
+ will try to make this particular tensor a view of the
+ stated reference. If making a view of reference is likely
+ to break the data integrity, nntrainer will request an
+ independent memory slot, in this case, it is user's
+ responsibility to copy the data. */
+ SHARED, /**< Shared defines a shared tensor ownership for the given
+ identifier, it is user's responsibility to guarantee that
+ dimension and initializer of shared tensor if exactly same as the
+ user will be agnostic about when and who will actually request
+ the certain tensor. */
+ };
+
+ RequestType request_type = RequestType::UNIQUE; /**< Type of request */
+ std::string name; /**< Identifier */
+ TensorDim dim; /**< dimension */
+ TensorLifespan ls; /**< lifespan */
+ Tensor::Initializer initializer =
+ Tensor::Initializer::NONE; /**< initializer */
+
+ /** ONLY USED FOR READ_ONLY_VIEW, MAYBE_MODIFYING_VIEW */
+ unsigned int offset; /**< tensor offset */
+ std::string reference_name; /**< reference name */
+};
+
+/**
+ * @brief variable + gradient specification
+ *
+ */
+struct VarGradSpecV2 {
+ TensorSpecV2 variable_spec; /**< variable spec */
+ std::unique_ptr<TensorSpecV2> gradient_spec =
+ nullptr; /**< gradient spec, if null it cannot be trained*/
+};
+
+/**
+ * @brief weight specification
+ *
+ */
+struct WeightSpecV2 {
+ VarGradSpecV2 vg_spec; /**< variable + graident specification */
+ WeightRegularizer regularizer = WeightRegularizer::NONE; /**< regularizer */
+ float regularizer_constant = 0.0f; /**< regularizer constant */
+};
+
} // namespace nntrainer
#endif /** __TENSOR_WRAP_SPECS_H__ */
projects / platform / core / ml / nntrainer.git / commitdiff