Method export_quantization_config

ppq/parser/caffe_exporter.py:180–218

(self, config_path: str, graph: BaseGraph)

class SNPECaffeExporter(CaffeExporter):
    def export_quantization_config(self, config_path: str, graph: BaseGraph):
        var_quant_info_recorder = {}
        for operation in graph.operations.values():
            if not isinstance(operation, QuantableOperation): continue
            for config, var in operation.config_with_variable:
                if not QuantizationStates.can_export(config.state):
                    raise PermissionError(
                        'Can not export quant config cause not all quantization configurations '
                        'have been correctly initialized(or some of them has been deactivated). '
                        f'Operation {operation.name} has an invalid quantization config({config.state}) '
                        f'at variable {var.name}.')

                # PATCH 2021.11.25
                # REMOVE BIAS FROM CONFIGURATION
                if config.num_of_bits > 8:
                        continue

                if config.state in {
                    QuantizationStates.FP32,
                    QuantizationStates.SOI
                }: continue
                # Simply override recorder is acceptable here,
                # we do not support mix precision quantization for CUDA backend now.
                # All configurations for this variable should keep identical towards each other.
                if config.state == QuantizationStates.PASSIVE and var.name in var_quant_info_recorder: continue
                var_quant_info_recorder[var.name] = config

        # ready to render config to json.
        for var in var_quant_info_recorder:
            config = var_quant_info_recorder[var]
            assert isinstance(config, TensorQuantizationConfig)
            tensorwise = config.policy.has_property(QuantizationProperty.PER_TENSOR)
            var_quant_info_recorder[var] = {
                'tensor_min' : convert_value(config.scale * (config.quant_min - config.offset), tensorwise, DataType.FP32),
                'tensor_max' : convert_value(config.scale * (config.quant_max - config.offset), tensorwise, DataType.FP32)
            }

        with open(file=config_path, mode='w') as file:
            json.dump(var_quant_info_recorder, file, indent=4)

    def export(self, file_path: str, graph: BaseGraph, config_path: str = None, input_shapes: List[List[int]] = [[1, 3, 224, 224]]):
        # dump config