Data models

The protobuf and on-disk formats this repo defines. Most are under tensorflow/core/framework/*.proto, tensorflow/core/protobuf/*.proto, and tensorflow/lite/schema/*.fbs.

Core types

Tensor

A tensor has a DataType, a TensorShape, and a TensorBuffer of bytes. Definition: tensorflow/core/framework/tensor.h. Wire format: tensorflow/core/framework/tensor.proto (TensorProto).

DataType enum: tensorflow/core/framework/types.proto. Notable values: DT_FLOAT, DT_DOUBLE, DT_INT8/16/32/64, DT_UINT8/16/32/64, DT_BOOL, DT_STRING, DT_BFLOAT16, DT_HALF, DT_VARIANT, DT_RESOURCE, plus the reference variants (DT_FLOAT_REF, …).

TensorShape

tensorflow/core/framework/tensor_shape.proto (TensorShapeProto). Optional unknown_rank, list of dim with optional name and size (where -1 denotes an unknown dimension).

OpDef

The schema of an op. tensorflow/core/framework/op_def.proto:

message OpDef {
  string name = 1;
  repeated ArgDef input_arg = 2;
  repeated ArgDef output_arg = 3;
  repeated AttrDef attr = 4;
  ...
}

NodeDef, GraphDef

A NodeDef (tensorflow/core/framework/node_def.proto) is one node in a graph: name, op, inputs, attrs, device. A GraphDef is a list of NodeDefs plus versions (tensorflow/core/framework/graph.proto).

FunctionDef, FunctionDefLibrary

tensorflow/core/framework/function.proto. A function is a OpDef plus a list of node definitions and ret-name mappings. FunctionDefLibrary packages many functions.

MetaGraphDef, SavedModel

tensorflow/core/protobuf/meta_graph.proto and tensorflow/core/protobuf/saved_model.proto. The on-disk format for SavedModel; see features/saved-model for usage.

RunOptions, RunMetadata

tensorflow/core/protobuf/config.proto and tensorflow/core/protobuf/run_metadata.proto. Per-run configuration and tracing.

Configuration protos

Proto	File	Purpose
ConfigProto	tensorflow/core/protobuf/config.proto	Session-level configuration.
GPUOptions	tensorflow/core/protobuf/config.proto	GPU-specific knobs.
RewriterConfig	tensorflow/core/protobuf/rewriter_config.proto	Grappler optimizer toggles.
ClusterDef / JobDef / TaskDef	tensorflow/core/protobuf/cluster.proto	Distributed cluster description.
DebugOptions	tensorflow/core/protobuf/debug.proto	Tensor-level debugger options.

Service definitions

gRPC service contracts:

Proto	Service
tensorflow/core/protobuf/master_service.proto	Master
tensorflow/core/protobuf/worker_service.proto	Worker
tensorflow/core/protobuf/eager_service.proto	Remote eager
tensorflow/core/distributed_runtime/coordination/coordination_service.proto	Coordination service

Examples and records

tensorflow/core/example/ holds the canonical training-record formats:

• Example (example.proto) — a map of feature name → Feature (bytes_list/float_list/int64_list).
• SequenceExample — variable-length sequences of features.
• Feature, Features (feature.proto) — building blocks.

Disk format: each record is length-prefixed inside a TFRecord file (tensorflow/core/lib/io/record_*).

TFLite flatbuffer schema

tensorflow/lite/schema/schema.fbs — the entire .tflite file format. Key tables:

• Model — root: version, operator codes, subgraphs, buffers, metadata.
• SubGraph — tensors, op codes, inputs, outputs.
• Operator — refers to an operator code and the input/output tensor indices.
• Buffer — raw bytes for constants.
• Tensor — name, shape, dtype, quantization parameters, sparsity.

Schema versioning is in tensorflow/lite/version.h. See apps/tensorflow-lite.

Auxiliary protos

A handful of useful ones to know:

• Saver-related — tensorflow/core/protobuf/saver.proto, meta_graph.proto.
• Profiler — tensorflow/core/profiler/protobuf/xplane.proto (XSpace/XPlane/XStat).
• Trackable graph — tensorflow/core/protobuf/trackable_object_graph.proto.

Related

• features/saved-model — uses many of the above.
• systems/distributed-runtime — uses the gRPC service protos.
• apps/tensorflow-lite — uses schema.fbs.