furiosa.quantizer.frontend.onnx package

Subpackages

• furiosa.quantizer.frontend.onnx.quantizer package
  • Submodules
  • furiosa.quantizer.frontend.onnx.quantizer.eliminate_clipper module
  • furiosa.quantizer.frontend.onnx.quantizer.quantizer module
  • furiosa.quantizer.frontend.onnx.quantizer.quantizer_mode module
  • furiosa.quantizer.frontend.onnx.quantizer.utils module
  • Module contents
• furiosa.quantizer.frontend.onnx.transformer package
  • Subpackages
    • furiosa.quantizer.frontend.onnx.transformer.experimental package
      • Submodules
      • furiosa.quantizer.frontend.onnx.transformer.experimental.eliminate_detection_postprocess module
      • furiosa.quantizer.frontend.onnx.transformer.experimental.fuse_div_for_bert module
      • furiosa.quantizer.frontend.onnx.transformer.experimental.reify_conv_for_bert module
      • Module contents
  • Submodules
  • furiosa.quantizer.frontend.onnx.transformer.convert_2d_sum_to_add module
  • furiosa.quantizer.frontend.onnx.transformer.convert_conv1d_to_conv2d module
  • furiosa.quantizer.frontend.onnx.transformer.eliminate_redundant_shape_pattern module
  • furiosa.quantizer.frontend.onnx.transformer.fuse_batchnorm module
  • furiosa.quantizer.frontend.onnx.transformer.fuse_conv module
  • furiosa.quantizer.frontend.onnx.transformer.fuse_depth_to_space module
  • furiosa.quantizer.frontend.onnx.transformer.fuse_gather_matmul module
  • furiosa.quantizer.frontend.onnx.transformer.fuse_gelu module
  • furiosa.quantizer.frontend.onnx.transformer.fuse_layer_normalization module
  • furiosa.quantizer.frontend.onnx.transformer.fuse_lp_normalization module
  • furiosa.quantizer.frontend.onnx.transformer.fuse_pad module
  • furiosa.quantizer.frontend.onnx.transformer.fuse_redundant_reshape_pattern module
  • furiosa.quantizer.frontend.onnx.transformer.infer_squeeze_axes module
  • furiosa.quantizer.frontend.onnx.transformer.polish_model module
  • furiosa.quantizer.frontend.onnx.transformer.utils module
  • Module contents
• furiosa.quantizer.frontend.onnx.utils package
  • Submodules
  • furiosa.quantizer.frontend.onnx.utils.check_model module
  • furiosa.quantizer.frontend.onnx.utils.inference_shape module
  • furiosa.quantizer.frontend.onnx.utils.version_checker module
  • Module contents

Submodules

furiosa.quantizer.frontend.onnx.calibrate module

class furiosa.quantizer.frontend.onnx.calibrate.CalibrationDataReaderForIterator(iterator: Iterator[Dict[str, numpy.ndarray]])

Bases: onnxruntime.quantization.calibrate.CalibrationDataReader

A CalibrationDataReader that wraps dicts mapping input tensor names to their values.

get_next()

generate the input data dict for ONNXinferenceSession run

exception furiosa.quantizer.frontend.onnx.calibrate.CalibrationError

Bases: Exception

The base class for all exceptions that are related to calibration.

furiosa.quantizer.frontend.onnx.calibrate.calibrate(model: onnx.onnx_ml_pb2.ModelProto, dataset: Iterable[Dict[str, numpy.ndarray]], augmented_model_path: Optional[str] = None) → Dict[str, Tuple[float, float]]

Estimates the range of tensors in a model, based on a dataset.

Parameters

• model – An ONNX model to calibrate.

• dataset – An Iterable that returns dicts mapping input tensor names to their values.

• augmented_model_path – A path to save an augmented model to.

Returns

A dict mapping tensors in the model to their minimum and maximum values.

furiosa.quantizer.frontend.onnx.calibrate.calibrate_with_random_data(model: onnx.onnx_ml_pb2.ModelProto, dataset_size: int = 8, augmented_model_path: Optional[str] = None) → Dict[str, Tuple[float, float]]

Estimates the range of tensors in a model, based on a random dataset.

Parameters

• model – An ONNX model to calibrate.

• dataset_size – the size of a random dataset to use.

• augmented_model_path – A path to save an augmented model to.

Returns

A dict mapping tensors in the model to their minimum and maximum values.

Module contents

furiosa.quantizer.frontend.onnx.optimize_model(model: onnx.onnx_ml_pb2.ModelProto, input_shapes: Optional[Dict[str, List[int]]] = None) → onnx.onnx_ml_pb2.ModelProto

furiosa.quantizer.frontend.onnx.parse_onnx_graph(model: onnx.onnx_ml_pb2.ModelProto) → Tuple[Dict[str, onnx.onnx_ml_pb2.ValueInfoProto], Dict[str, onnx.onnx_ml_pb2.NodeProto], Dict[str, List[onnx.onnx_ml_pb2.NodeProto]]]

furiosa.quantizer.frontend.onnx.post_training_quantization_with_random_calibration(model: onnx.onnx_ml_pb2.ModelProto, per_channel: bool, static: bool, mode: furiosa.quantizer.frontend.onnx.quantizer.utils.QuantizationMode, num_data: int = 8, check_idempotency: bool = False) → onnx.onnx_ml_pb2.ModelProto

furiosa.quantizer.frontend.onnx.post_training_quantize(model: onnx.onnx_ml_pb2.ModelProto, dataset: List[Dict[str, numpy.ndarray]], per_channel: bool = True, check_idempotency: bool = False) → onnx.onnx_ml_pb2.ModelProto

Post-training-quantizes an ONNX model with a calibration dataset.

Parameters

• model – An ONNX model to quantize.

• dataset – A calibration dataset.

• per_channel – If per_channel is True, Conv’s filters are per-channel quantized. Otherwise, they are per-tensor quantized.

Returns

An ONNX model post-training-quantized with the calibration dataset.

furiosa.quantizer.frontend.onnx.quantize(model: onnx.onnx_ml_pb2.ModelProto, per_channel: bool, static: bool, mode: furiosa.quantizer.frontend.onnx.quantizer.utils.QuantizationMode, dynamic_ranges: Dict[str, Tuple[float, float]]) → onnx.onnx_ml_pb2.ModelProto