package tensorflow.tf2xla

Get desktop application:
View/edit binary Protocol Buffers messages

Config represents configuration information for tf2xla conversion.

• repeated Feed feed = 1

  Each feed is a positional input argument for the generated computation. The order of each entry matches the order of each input argument.

• repeated Fetch fetch = 2

  Each fetch is a positional output argument for the generated computation. The order of each entry matches the order of each output argument.

• repeated Variable variable = 3

  Each variable is a named input and output of the generated computation.

• optional ConversionOptions conversion_options = 4

  Optional conversion options.

Options used during the conversion and compilation process.

Used in: Config

• bool custom_fake_quant_op_calls = 1

  When true tf.fake_quant_* ops will be emitted as custom calls to a 'fake_quant_with_min_max_vars' function accepting the input, min, max, num_bits, and narrow_range values as runtime arguments.

Feed represents a single feed tensor in the graph, which corresponds to an input argument for the generated computation.

Used in: Config

• optional TensorId id = 1

• optional TensorShapeProto shape = 2

• string name = 3

  Optional name for generated code.

• DataType type = 4

  Optional data type. This is not normally required, as the graph itself contains this information. However, if the node being fed is an op that is not linked into the binary, then the type cannot be inferred from the node; in this case, the type should be set here.

Fetch represents a single fetch tensor in the graph, which corresponds to an output argument for the generated computation.

Used in: Config

• optional TensorId id = 1

• string name = 2

  Optional name for generated code.

• optional TensorShapeProto shape = 3

  Optional shape and data type. If specified, may be used for validation.

• DataType type = 4

HostComputeMetadata describes all the sends and recvs from all host compute transfer ops in a computation.

Used in: xrt.XLAComputationConfig

• repeated HostTransferMetadata device_to_host = 1

  Metadata about each device_to_host transfer

• repeated HostTransferMetadata host_to_device = 2

  Metadata about each host_to_device transfer

HostTransferMetadata describes a transfer either from host to device or device to host. It has a key that is unique to the computation, and metadata about the list of tensors being transferred.

Used in: HostComputeMetadata

• string key = 1

  The key used to identify this transfer.

• repeated TensorMetadata metadata = 2

  For each Tensor being transferred, its type and shape.

TensorId identifies a tensor in a TensorFlow graph, by specifying the output index of a particular node in the graph. If the output of the named node feeds into other node(s), this corresponds to one or more edges. Otherwise it doesn't correspond to any existing edges at all, e.g. for output nodes.

Used in: Feed, Fetch

• string node_name = 1

• int64 output_index = 2

TensorMetadata indicates the type and shape of a Tensor that is part of a host compute transfer.

Used in: HostTransferMetadata

• DataType type = 1

• optional TensorShapeProto shape = 2

Variable represents a resource variable with the given name, shape and type.

Used in: Config

• string node_name = 1

• string name = 2

  Optional name for generated code. If empty, node_name will be used.

• optional TensorShapeProto shape = 3

• DataType type = 4

• bool readonly = 5

  Flag for variables that are never assigned. Assigments to a read-only variable or unassigned variables that are not read-only are invalid.