Crate caffe2_predictor

Modules

• emulator

Structs

• AllThreadLocalHelperVector
  | A thread safe vector of all ThreadLocalHelper, | this will be used to encapuslate the | locking in the APIs for the changes to | the global AllThreadLocalHelperVector instance. |
• BenchmarkParam
• BenchmarkRunner
  | benchmark runner takes an @emulator | to run nets. | | The runtime will be measured by @profiler. | | The output will be formatted by @formatter |
• DataNetFiller
  | @init_net: a reader net to generate | parameters | | @data_net: a reader net to generate | inputs |
• DataRandomFiller
  | @run_net: the predict net with parameter | and input names | | @input_dims: the input dimensions | of all operator inputs of run_net | | @input_types: the input types of all | operator inputs of run_net |
• Filler
• InferenceGraph
  | This struct stores information about | the inference graph which defines underlying | math of BlackBoxPredictor. Other parts of | it such as various threading optimizations | don’t belong here. |
• MutatingNetSupplier

  ---

• Predictor
• PredictorConfig
  | Stores parameters nessasary for creating | a PredictorInterface object. |
• PredictorTest
• RunnableNet
• SingleLoadedNetSupplier
  | A simple net supplier that always return | the same net and filler pair. | | The SingleLoadedNetSupplier contains | a shared ptr to a workspace with parameters | already loaded by net loader. |
• SingleNetSupplier
  | A simple net supplier that always return | the same net and filler pair. |
• StdOutputFormatter
  | Print the output of the emulator run | to stdout. |
• TestDataRandomFiller

  ---

• ThreadLocalHelper
  | ThreadLocalHelper is per thread |
• ThreadLocalPtr

  ---

• ThreadLocalPtrImpl
  | ThreadLocalPtrImpl is per object |

Constants

• MS_IN_SECOND
• initSpec
• metaSpec
• predictSpec

Traits

• Emulator
  | A net emulator. In short, it can run nets | with given @iterations. |
• FillerTrait
  | A filler to initialize the parameters | and inputs of a predictor |
• NetSupplier
  | An interface to supplier a pair of net | and its filler. | | The net should be able to run once the | filler fills the workspace. | | The supplier should take the ownership | of both net and filler. |
• OutputFormatter
  | An interface that formats the output | of the emulator runs. |
• Profiler
  | An interface to profile the metrics | of a @runnable. | | It should return execution walltime | in milliseconds. |

Functions

• enforce_is_tensor
• extract_meta_net_def
• fill_random_network_inputs
  Convenient helpers to fill data to workspace.
• fill_with_type
• found_op_candidate
• get_all_thread_local_helper_vector
  meyer’s singleton
• get_blob
• get_tensor
• get_thread_local_helper
• make_predictor_config
• make_predictor_config_with_init_net_and_run_net
• make_workspace
• parse_meta_net_def
• parse_net_def
• random_tensor
• remove_ops_by_type
  | Conceptually it’s a pretty easy process | and consists of 3 steps: | | -1) SSA rewrite; | | -2) propagate inputs forwards; | | -3) propagate outputs backwards and | then forwards again. | | However, because of model outputs which | can’t be overwritten during the SSA | process, and the fact that inputs could | be overwritten by ops and also appear | in outputs, it adds a lot of extra complexity | to handle these special cases. | | A lot of this extra complexity is handled | in | | FoundOpCandidate. |
• rename_inputs
• rename_inputs_in_children
• rename_output_in_parents
• rename_outputs
• rewrite_subnets_for_if_op
• run_global_initialization
  | Extract the MetaNetDef from db, and | run the global init net on the master | workspace. |

Type Definitions

• PredictorParameters
  | Parameters for a Predictor provided | by name. | | They are stored as shared_ptr to accommodate | parameter sharing |
• Predictor_TensorList
• Predictor_TensorMap
• UnsafeAllThreadLocalHelperVector
• UnsafeThreadLocalMap
  | Map of object pointer to instance in | each thread to achieve per thread(using | thread_local) per object(using the | map) thread local pointer |
• filler_type_pair_t

  ---