quark.onnx.finetuning.create_torch.create_model_ops
#
Module Contents#
Functions#
- quark.onnx.finetuning.create_torch.create_model_ops.param_is_symmetric(params: List[Any]) bool #
Check if parameters are symmetric, all values [2,2,2,2]. Then we can use only [2,2].
- quark.onnx.finetuning.create_torch.create_model_ops.extract_padding_params(params: List[Any]) Any #
Extract padding parameters for Pad layers.
- quark.onnx.finetuning.create_torch.create_model_ops.extract_padding_params_for_conv(params: List[Any]) Any #
Padding params in onnx are different than in pytorch. That is why we need to check if they are symmetric and cut half or return a padding layer.
- quark.onnx.finetuning.create_torch.create_model_ops.extract_weight_and_bias(params: List[Any]) Tuple[numpy.typing.NDArray[Any], numpy.typing.NDArray[Any] | None] #
Extract weights and biases.
- quark.onnx.finetuning.create_torch.create_model_ops.load_weight_and_bias(layer: torch.nn.Module, weight: numpy.typing.NDArray[Any], bias: numpy.typing.NDArray[Any] | None) None #
Load weight and bias to a given layer from onnx format.
- quark.onnx.finetuning.create_torch.create_model_ops.convert_conv(node: onnx.NodeProto, layer_params: List[Any], layer_qinfos: List[Any]) Tuple[quark.onnx.finetuning.create_torch.quant_base_ops.QuantizeWrapper, quark.onnx.finetuning.create_torch.quant_base_ops.QuantizeWrapper | None] #
- Use to convert Conv ONNX node to Torch module (or called layer).
This function supports onnx’s Conv and ConvTranspose from 1 to 11.
:param node : ONNX node. :param layer_params : Layer weight and bias parameters. :param layer_qinfos : Layer quantization information. :return: Converted conv layer, perhaps it has a pad layer.
- quark.onnx.finetuning.create_torch.create_model_ops.convert_matmul(node: onnx.NodeProto, layer_params: List[Any], layer_qinfos: List[Any]) Tuple[quark.onnx.finetuning.create_torch.quant_matmul_ops.QMatMul, None] #
Use to convert MatMul ONNX node to Torch module. This function supports onnx’s Matmul from 6.
:param node : ONNX node. :param layer_params : Layer weight parameters. :param layer_qinfos : Layer quantization informations. :return: Converted MatMul layer.
- quark.onnx.finetuning.create_torch.create_model_ops.convert_gemm(node: onnx.NodeProto, layer_params: List[Any], layer_qinfos: List[Any]) Tuple[quark.onnx.finetuning.create_torch.quant_gemm_ops.QGemm, None] #
- Use to convert Gemm ONNX node to Torch module.
This function supports onnx’s Instance Norm from 6.
:param node : ONNX node. :param layer_params : Layer weight and bias parameters. :param layer_qinfos : Layer quantization information. :return: Converted Gemm layer.
- quark.onnx.finetuning.create_torch.create_model_ops.convert_norm(node: onnx.NodeProto, layer_params: List[Any], layer_qinfos: List[Any]) Tuple[quark.onnx.finetuning.create_torch.quant_norm_ops.QInstanceNorm2d, None] #
- Use to convert norm (Instance Norm) ONNX node to Torch module.
This function supports onnx’s Instance Norm from 6.
:param node : ONNX node. :param layer_params : Layer weight and bias parameters. :param layer_qinfos : Layer quantization information. :return: Converted norm (Instance Norm) layer.
- quark.onnx.finetuning.create_torch.create_model_ops.convert_act(node: onnx.NodeProto) torch.nn.Module | None #
Use to convert Activation ONNX node to Torch module (or called layer).
:param node : ONNX node. :return: Converted act layer.
- quark.onnx.finetuning.create_torch.create_model_ops.convert_ops_to_modules(onnx_model: onnx.ModelProto) Tuple[torch.nn.Module | None, torch.nn.Module | None, torch.nn.Module | None, quark.onnx.finetuning.create_torch.quant_base_ops.QuantizationModule | None] #
Convert ONNX operations to Torch modules.
- quark.onnx.finetuning.create_torch.create_model_ops.set_modules_original_weight(module: torch.nn.Module, weight: numpy.typing.NDArray[Any]) None #
For setting original float weight
- quark.onnx.finetuning.create_torch.create_model_ops.get_modules_optimized_weight(module: torch.nn.Module) Any #
For getting optimized quantized weight
- quark.onnx.finetuning.create_torch.create_model_ops.set_modules_original_bias(module: torch.nn.Module, bias: numpy.typing.NDArray[Any]) None #
For setting original float bias
- quark.onnx.finetuning.create_torch.create_model_ops.get_modules_optimized_bias(module: torch.nn.Module) Any #
For getting optimized quantized bias