def emit_BatchNorm()

in mmdnn/conversion/onnx/onnx_emitter.py [0:0]
85 lines of code
3 McCabe index (conditional complexity)

    def emit_BatchNorm(self, IR_node):
        epsilon = IR_node.get_attr('epsilon')
        if IR_node.get_attr('scale'):
            self.add_body(1, "{:15} = _weights_dict['{}']['scale'].squeeze()".format(
                IR_node.variable_name + '_scale_array',
                IR_node.name))
        else:
            self.add_body(1, "{:15} = np.ndarray(_weights_dict['{}']['bias'].shape, dtype=_weights_dict['{}']['bias'].dtype).squeeze()".format(
                              IR_node.variable_name + '_scale_array',
                              IR_node.name,
                              IR_node.name))
            self.add_body(1, "{:15}.fill(1)".format(IR_node.variable_name + '_scale_array'))

        self.add_body(1, "{:15} = helper.make_tensor_value_info('{}', onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[{}.dtype], list({}.shape))".format(
            IR_node.variable_name + '_scale',
            IR_node.variable_name + '_scale',
            IR_node.variable_name + '_scale_array',
            IR_node.variable_name + '_scale_array'))
    
        self.add_body(1, "{:15} = helper.make_tensor(name='{}', data_type=onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[{}.dtype], dims={}.shape, vals={})".format(
                          IR_node.variable_name + '_scale_init',
                          IR_node.variable_name + '_scale',
                          IR_node.variable_name + '_scale_array',
                          IR_node.variable_name + '_scale_array',
                          IR_node.variable_name + '_scale_array'))
        self.add_body(1, "{:15} = _weights_dict['{}']['bias'].squeeze()".format(
            IR_node.variable_name + '_bias_array',
            IR_node.name))
        self.add_body(1, "{:15} = helper.make_tensor_value_info('{}', onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[{}.dtype], list({}.shape))".format(
                        IR_node.variable_name + '_bias',
                        IR_node.variable_name + '_bias',
                        IR_node.variable_name + '_bias_array',
                        IR_node.variable_name + '_bias_array'))

        self.add_body(1, "{:15} = helper.make_tensor(name='{}', data_type=onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[{}.dtype], dims={}.shape, vals={}.flatten().astype(float))".format(
                            IR_node.variable_name + '_bias_init',
                            IR_node.variable_name + '_bias',
                            IR_node.variable_name + '_bias_array',
                            IR_node.variable_name + '_bias_array',
                            IR_node.variable_name + '_bias_array'))

        self.add_body(1, "{:15} = _weights_dict['{}']['mean'].squeeze()".format(
            IR_node.variable_name + '_mean_array',
            IR_node.name))
        self.add_body(1, "{:15} = helper.make_tensor_value_info('{}', onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[{}.dtype], list({}.shape))".format(
                        IR_node.variable_name + '_mean',
                        IR_node.variable_name + '_mean',
                        IR_node.variable_name + '_mean_array',
                        IR_node.variable_name + '_mean_array'))

        self.add_body(1, "{:15} = helper.make_tensor(name='{}', data_type=onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[{}.dtype], dims={}.shape, vals={}.flatten().astype(float))".format(
                            IR_node.variable_name + '_mean_init',
                            IR_node.variable_name + '_mean',
                            IR_node.variable_name + '_mean_array',
                            IR_node.variable_name + '_mean_array',
                            IR_node.variable_name + '_mean_array'))

        self.add_body(1, "{:15} = _weights_dict['{}']['var'].squeeze()".format(
                          IR_node.variable_name + '_var_array',
                          IR_node.name))
        self.add_body(1, "{:15} = helper.make_tensor_value_info('{}', onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[{}.dtype], list({}.shape))".format(
                            IR_node.variable_name + '_var',
                            IR_node.variable_name + '_var',
                            IR_node.variable_name + '_var_array',
                            IR_node.variable_name + '_var_array'))
    
        self.add_body(1, "{:15} = helper.make_tensor(name='{}', data_type=onnx.mapping.NP_TYPE_TO_TENSOR_TYPE[{}.dtype], dims={}.shape, vals={}.flatten().astype(float))".format(
                            IR_node.variable_name + '_var_init',
                            IR_node.variable_name + '_var',
                            IR_node.variable_name + '_var_array',
                            IR_node.variable_name + '_var_array',
                            IR_node.variable_name + '_var_array'))                         

        self.add_body(1, "{:15} = helper.make_node('BatchNormalization', inputs=['{}', '{}', '{}', '{}', '{}'],outputs=['{}'], epsilon={}, is_test={}, name='{}')".format(
                          IR_node.variable_name,
                          self.parent_variable_name(IR_node),
                          IR_node.variable_name + '_scale',
                          IR_node.variable_name + '_bias',
                          IR_node.variable_name + '_mean',
                          IR_node.variable_name + '_var',
                          IR_node.variable_name,
                          epsilon,
                          0 if self.phase == 'train' else 1,
                          IR_node.variable_name))
        self.initializer.append(IR_node.variable_name + '_scale_init')
        self.initializer.append(IR_node.variable_name + '_bias_init')
        self.initializer.append(IR_node.variable_name + '_mean_init')
        self.initializer.append(IR_node.variable_name + '_var_init')
        self.inputs.append(IR_node.variable_name + '_scale')
        self.inputs.append(IR_node.variable_name + '_bias')
        self.inputs.append(IR_node.variable_name + '_mean')
        self.inputs.append(IR_node.variable_name + '_var')
        self.nodes.append(IR_node.variable_name)