package torch

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: ModuleDef

• required string type = 1

  The mypy type of this attribute

• required string name = 2

• required int64 id = 3

  Offset into attribute table

Represents all non-module code that the model depends on. Right now it's just a straight list of classes, defined in dependency order (i.e. dependencies appear before their dependers)

• optional RecordRef torchscript_arena = 1

• optional int64 proto_version = 1

  for the proto version, to keep both backward and forward compatibility, please bump the proto_version when we add any change in the proto. runtime decides whether accept the model based on the ir_version.

• optional ModuleDef main_module = 2

  main module of the model

• optional string producer_name = 3

  to distinguish whether exported from c2 or torch

• optional string producer_version = 4

  put build version here

• repeated TensorDef tensors = 5

  the table contains all the tensor information the tensor id is defined as TensorProto.name

Used in: ModelDef

• repeated ModuleDef submodules = 1

• optional RecordRef torchscript_arena = 2

• repeated caffe2.NetDef caffe2_nets = 3

• optional RecordRef pickle_arena = 4

  because the old pickle modules may not be supported by torch_script, have to stored as pickle_arena at this moment.

• optional RecordRef cpp_arena = 5

  should be exposed by the Class Archive, so user can save module specific data which cannot be store in the graph or torch_script

• repeated ParameterDef parameters = 6

  the parameters of this module

• optional string name = 7

• optional bool optimize = 8

  whether apply the optimizations to this module, only applicable to script modules

• repeated AttributeDef attributes = 9

• optional int64 get_state_attribute_id = 10

  Used for retrieving module state from the pickled IValues table

• optional RecordRef torchscript_debug_arena = 11

Used in: ModuleDef

• optional bool is_buffer = 1

  whether this parameter is registered as buffer or not

• optional int64 tensor_id = 2

  the offset into the tensor table where this parameter is stored

• optional string name = 3

• PROTO_VERSION_NEWEST = 6

Used in: LibDef, ModuleDef, TensorDef

• optional string key = 1

Used in: ModelDef

• repeated int64 dims = 1

• optional int64 offset = 2

• repeated int64 strides = 3

• optional bool requires_grad = 4

  whether we compute the gradient for the parameter

• optional caffe2.TensorProto.DataType data_type = 5

• optional RecordRef data = 6

• optional string device = 7

  device field stores the canonical device string, and it follows the format below: `(cpu|cuda)[:<device-index>]`, e.g., 'cuda:0'

• optional bool is_quantized = 8

• optional double scale = 9

• optional int64 zero_point = 10