yobx.litert.litert_helper

Minimal pure-Python TFLite/LiteRT FlatBuffer parser and helper utilities for the LiteRT→ONNX converter.

The parser requires no external dependencies beyond the Python standard library. It reads the binary FlatBuffer format that every .tflite file uses and exposes the model graph as plain Python dataclasses so that the converter can inspect operator types, tensor shapes, weights, and graph connectivity.

TFLite dtype enum mapping

Value	TFLite TensorType
0	FLOAT32
1	FLOAT16
2	INT32
3	UINT8
4	INT64
5	STRING
6	BOOL
7	INT16
8	COMPLEX64
9	INT8
10	FLOAT64
11	COMPLEX128
12	UINT64
14	UINT32
15	UINT16
16	INT4

class yobx.litert.litert_helper.ActivationFunctionType

NONE = 0

RELU = 1

RELU6 = 3

RELU_N1_TO_1 = 2

SIGN_BIT = 5

TANH = 4

class yobx.litert.litert_helper.BuiltinOperator

Subset of the TFLite BuiltinOperator enum.

ABS = 101

ADD = 0

AVERAGE_POOL_2D = 1

BATCH_MATMUL = 126

CEIL = 104

CONCATENATION = 2

CONV_2D = 3

CUSTOM = 32

DEPTHWISE_CONV_2D = 4

DEQUANTIZE = 6

DIV = 42

ELU = 111

EXP = 47

EXPAND_DIMS = 70

FLOOR = 8

FLOOR_DIV = 90

FULLY_CONNECTED = 9

GELU = 150

HARD_SWISH = 117

LEAKY_RELU = 98

LOG = 73

LOGICAL_AND = 86

LOGICAL_NOT = 87

LOGICAL_OR = 84

LOG_SOFTMAX = 50

MAX_POOL_2D = 17

MEAN = 40

MUL = 18

NEG = 59

PAD = 34

POW = 78

REDUCE_MAX = 82

REDUCE_MIN = 89

RELU = 19

RELU_N1_TO_1 = 20

RESHAPE = 22

ROUND = 116

RSQRT = 76

SIN = 66

SOFTMAX = 25

SQRT = 75

SQUARED_DIFFERENCE = 99

SQUEEZE = 43

SUB = 41

SUM = 74

TANH = 28

TRANSPOSE = 39

TRANSPOSE_CONV = 67

class yobx.litert.litert_helper.Padding

SAME = 0

VALID = 1

class yobx.litert.litert_helper.TFLiteModel(version: int, subgraphs: List[TFLiteSubgraph] | None = None)

Top-level parsed TFLite model.

class yobx.litert.litert_helper.TFLiteOperator(opcode: int, custom_code: str, inputs: Tuple[int, ...], outputs: Tuple[int, ...], builtin_options: Dict)

Parsed representation of a TFLite Operator.

The builtin_options dict contains decoded attributes for the current opcode (e.g. {"fused_activation": 0} for FULLY_CONNECTED). Unrecognised op options are left unparsed (empty dict).

property name: str

Human-readable op name for diagnostics.

class yobx.litert.litert_helper.TFLiteSubgraph(name: str, tensors: List[TFLiteTensor], inputs: Tuple[int, ...], outputs: Tuple[int, ...], operators: List[TFLiteOperator])

Parsed representation of a TFLite SubGraph.

class yobx.litert.litert_helper.TFLiteTensor(index: int, name: str, dtype: int, shape: Tuple[int, ...], data: ndarray | None)

Parsed representation of a TFLite Tensor.

yobx.litert.litert_helper.builtin_op_name(code: int) → str

Return the name of a TFLite BuiltinOperator code, e.g. 'RELU'.

yobx.litert.litert_helper.litert_dtype_to_np_dtype(dtype_int: int) → dtype

Map a TFLite TensorType integer to a numpy.dtype.

Parameters:

dtype_int – TFLite dtype integer from the FlatBuffer

Returns:

corresponding numpy.dtype

Raises:

ValueError – if the dtype is not supported

yobx.litert.litert_helper.parse_tflite_model(model: str | PathLike | bytes) → TFLiteModel

Parse a .tflite file (or raw bytes) and return a TFLiteModel.

Parameters:

model – path to a .tflite file or its raw bytes

Returns:

parsed TFLiteModel