spec

Typed export specification; importable without Keras or a training backend.

VALID_IOconstant
CALIBRATEDconstant
PrecisionclassNumeric format of an exported graph.
IODTypeclassElement type of an exported model's inputs and outputs.
ConversionModeclassHow the Keras model is traced for conversion.
TensorRoleclassRole of a model input or output.
state_input_namefunctionName of the input of state pair k.
state_output_namefunctionName of the output of state pair k.
state_pairfunction("in", k) or ("out", k) for a state tensor name, otherwise None.
ExportSpecclassWhat to export.
check_dense_per_channelfunctionRefuse per-tensor dense weights for a precision without quantized weights.
BackendUnavailableclassThe active Keras backend cannot export: LiteRT export needs the TensorFlow backend.
check_resetsfunctionRefuse state resets that are not increasing steps from 1 to steps - 1 of a streaming model.

VALID_IO: dict[Precision, frozenset[IODType]] = {Precision.FP32: frozenset({IODType.FLOAT32}), Precision.FP16: frozenset({IODType.FLOAT16}), Precision.A8W8: frozenset({IODType.INT8, IODType.FLOAT32}), Precision.A16W8: frozenset({IODType.INT16, IODType.FLOAT32})}

CALIBRATED: frozenset[Precision] = frozenset({Precision.A8W8, Precision.A16W8})

Precision()

Numeric format of an exported graph.

FP32 = 'fp32'

float32 weights and compute.

FP16 = 'fp16'

native float16 graph; inputs, weights, activations and outputs are float16.

A8W8 = 'a8w8'

int8 activations and weights, calibrated.

A16W8 = 'a16w8'

int16 activations and int8 weights, calibrated.

IODType()

Element type of an exported model’s inputs and outputs.

FLOAT32 = 'float32'

FLOAT16 = 'float16'

INT8 = 'int8'

INT16 = 'int16'

ConversionMode()

How the Keras model is traced for conversion.

KERAS = 'keras'

SAVED_MODEL = 'saved_model'

CONCRETE = 'concrete'

TensorRole()

Role of a model input or output.

SIGNAL = 'signal'

STATE = 'state'

AUX = 'aux'

state_input_name(k: int) -> str

Name of the input of state pair k.

state_output_name(k: int) -> str

Name of the output of state pair k.

state_pair(name: str) -> tuple[str, int] | None

("in", k) or ("out", k) for a state tensor name, otherwise None.

ExportSpec()

What to export. precision, io_dtype and mode have no defaults.

model_config = ConfigDict(frozen=True, extra='forbid')

format: Literal['litert'] = 'litert'

Export format: litert, a .tflite flatbuffer (the only one).

precision: Precision

Numeric format of the graph.

io_dtype: IODType

Input and output element type; must be valid for precision (VALID_IO).

mode: ConversionMode

How the model is traced for conversion.

strict: StrictBool = True

For calibrated precisions, refuse operators without an integer kernel instead of falling back to float operators. Defaults to True; it does not affect float precisions.

state_tie_tolerance: StrictFloat = Field(default=0.01, ge=0.0, le=0.5)

For calibrated precisions with integer I/O, export gives each state pair (state_in_k, state_out_k) one scale and zero point covering both tensors’ ranges, and refuses when that scale differs from either original by more than this fraction of it (at most 0.5). Models without state pairs ignore it.

dense_per_channel: StrictBool = True

For calibrated precisions, quantize FULLY_CONNECTED weights per output channel (True, the converter’s default) or with one scale per tensor (False). Convolutions stay per channel either way. The CMSIS-NN int16 FULLY_CONNECTED kernel of LiteRT for Microcontrollers accepts per-tensor weights only, so a16w8 exports for it use False; the reference and heliaRT kernels run either. Float precisions refuse False.

check_dense_per_channel(precision: Precision, dense_per_channel: bool) -> None

Refuse per-tensor dense weights for a precision without quantized weights.

Raises

Errors raised by check_dense_per_channel
Type	Description
`ValueError`	If ``dense_per_channel`` is False and ``precision`` is not a8w8 or a16w8.

BackendUnavailable()

The active Keras backend cannot export: LiteRT export needs the TensorFlow backend.

check_resets(resets: Collection[int], steps: int, stateful: bool, what: str) -> None

Refuse state resets that are not increasing steps from 1 to steps - 1 of a streaming model.

Raises

Errors raised by check_resets
Type	Description
`ValueError`	If ``resets`` are given for a model without state, or are not increasing steps within the sequence (step 0 always starts from zero states).