|
DLPrimitives
|
Major object used for inference. More...
#include <include/dlprim/net.hpp>
Public Member Functions | |
| Net (Context &ctx) | |
| Create an empty network object for a context. | |
| Net const & | operator= (Net const &)=delete |
| Net (Net const &)=delete | |
| Net & | operator= (Net &&)=default |
| Net (Net &&)=default | |
| void | load_model (ModelBase &model) |
| Load from external model, shortcut to load_from_json(mode.network()) + setup() + load_parameters(model) | |
| void | load_from_json (json::value const &v) |
| Configure network graph from json. | |
| void | load_from_json_file (std::string const &name) |
| Configure network graph from json. | |
| void | add_input_tensor (std::string const &name, TensorSpecs const &ts, bool requires_gradient=false) |
| Define network input. | |
| void | mark_output_tensor (std::string const &name) |
| Make a tensor as an output tensor (will be preserverd) by name. | |
| void | set_loss_weight (std::string const &name, float loss_weight=1.0) |
| Make a tensor as an output loss tensor, regardless of its name (by default a name starting with "loss") considered loss. | |
| void | load_parameters (std::istream &s, bool allow_missing=false) |
| Load parameters from binary stream DLP format (file) s must be seekable. | |
| void | load_parameters (std::string const &name, bool allow_missing=false) |
| Load parameters from file name can be either DLP or HDF5 format. | |
| void | load_parameters (ModelBase &model, bool allow_missing=false) |
| Load parameters from external model. | |
| void | save_parameters (std::string const &fname) |
| Save network parameters to DLP format. | |
| void | load_parameters_from_hdf5 (std::string const &fname, bool allow_missing=false) |
| Copy all parameters for hdf5 file format. allow_missing - ignore if parameter is not defined. | |
| void | save_parameters_to_hdf5 (std::string const &fname) |
| Save all network parameters to a file. | |
| bool | keep_intermediate_tensors () const |
| True if all intermediate results are kept. | |
| void | keep_intermediate_tensors (bool keep) |
| Set if to keed intermediate results for debugging. Default is false - optimise memory use and reuse intermediate memory chunks. | |
| void | add_operator (std::unique_ptr< Operator > op, std::string const &name, std::vector< std::string > const &inputs, std::vector< std::string > const &outputs, std::vector< std::string > const ¶meters=std::vector< std::string >(), bool frozen=false) |
| Add an operator op to the network. name should be unique. More... | |
| void | mode (CalculationsMode mode) |
| Set the calculations mode. More... | |
| CalculationsMode | mode () const |
| get currenct calcuations mode | |
| std::shared_ptr< SharedResource > | shared_resource () |
| Get SharedResource object that can be shared between all operators in same network allowing operator access to network level features. | |
| void | setup () |
| Allocate all tensors and prepare network data. | |
| void | reshape () |
| reshape all tensors in network after reshaping the input data | |
| void | initialize_parameters (ExecutionContext const &e) |
| Initialize all parameters for training. | |
| void | copy_parameters_to_device () |
| void | copy_parameters_to_host () |
| void | clear_memory () |
| std::map< std::string, Tensor > & | tensors () |
| Get all intermediate tensors (operators inputs/outputs). Note if keep_intermediate_tensors() == false that same memory may be shared between different tensors. | |
| std::map< std::string, Tensor > & | tensor_diffs () |
| Get all intermediate tensor (operators inputs/outputs) gradient. Note if keep_intermediate_tensors() == false that same memory may be shared between different tensors. | |
| std::map< std::string, Tensor > & | params () |
| All operator parameters trainable and not trainable. | |
| std::map< std::string, Tensor > & | param_diffs () |
| All operator parameters gradients trainable and not trainable. | |
| Tensor & | tensor (std::string const &name) |
| Get tensor by name, throws ValidationError if does not exist. | |
| Tensor & | tensor_diff (std::string const &name) |
| Get tensor gradient by name, throws ValidationError if does not exist. | |
| Tensor & | param (std::string const &name) |
| Get parameter by name, throws ValidationError if does not exist. | |
| Tensor & | param_diff (std::string const &name) |
| Get parameter gradient by name, throws ValidationError if does not exist. | |
| void | forward (ExecutionContext const &ectx, bool sync=false) |
| void | backward (ExecutionContext const &ectx, bool sync=false) |
| std::vector< std::string > const & | input_names () |
| Get of all network inputs. | |
| std::vector< std::string > const & | output_names () |
| Get of all network outputs. | |
| Tensor & | input (unsigned id) |
| Get input tensors by their numeric id - according to the order in input_names() | |
| Tensor & | output (unsigned id) |
| Get output tensors by their numeric id - according to the order in output_names() | |
Major object used for inference.
| void dlprim::Net::add_operator | ( | std::unique_ptr< Operator > | op, |
| std::string const & | name, | ||
| std::vector< std::string > const & | inputs, | ||
| std::vector< std::string > const & | outputs, | ||
| std::vector< std::string > const & | parameters = std::vector< std::string >(), |
||
| bool | frozen = false |
||
| ) |
Add an operator op to the network. name should be unique.
inputs - tensor input names - must be defined (as input or output of another operatos) outputs - tensor output names parameters - give cutom names to parameters
Referenced by keep_intermediate_tensors().
| void dlprim::Net::mode | ( | CalculationsMode | mode | ) |
Set the calculations mode.
default is testing. If setup was called when the network in testing mode it can't be switched to training if you plan to use the network for training call it before setup()
1.8.11