本文共 2793 字，大约阅读时间需要 9 分钟。

code：https://github.com/OAID/Tengine

version: 88b4b7a2

图片，代码都来自以上项目。

1. 简介

…

2. 正题

1. 创建一个空的graph，ir_graph_t* ir_graph = create_ir_graph((struct context*)context);
2. 找到模型序列化器struct serializer* loader = find_serializer_via_name(model_format);，然后调用load_model方法
3. load_graph，会初始化graph的tensor_list,node_list,输入输出的node，sub_graphes

if (load_graph_tensors(tm2_s, graph, priv) < 0)         // 载入tensor_list    goto error;if (load_graph_nodes(tm2_s, graph, priv) < 0)           // 初始化node_list    goto error;if (set_graph_io_nodes(tm2_s, graph, priv) < 0)    goto error;if (load_graph_sub_info(tm2_s, graph, priv) < 0)    goto error;

4. load_graph_tensors中会遍历预训练的模型，创建tensor，包括type，data，shape等属性，并添加到graph.tensor_list中。
5. load_graph_nodes中会遍历预训练模型中的nodes，创建node，设置node的index，input_num, output_num, input_tensors, output_tensors, op等属性，并添加到graph.node_list
6. source/operator/prototype目录下所有op文件是用于node的op初始化，op->param_mem参数空间申请。

#include "graph/tensor.h"#include "graph/node.h"#include "graph/graph.h"#include "module/module.h"#include "utility/sys_port.h"static int infer_shape(struct node* node)       // static 函数指针的位置不变{
       struct graph* ir_graph = node->graph;    struct tensor* input = get_ir_graph_tensor(ir_graph, node->input_tensors[0]);    struct tensor* output = get_ir_graph_tensor(ir_graph, node->output_tensors[0]);    set_ir_tensor_shape(output, input->dims, input->dim_num);           // 设置 output tensor 的尺寸    return 0;}static int init_op(struct op* op){
       op->param_mem = NULL;    op->param_size = 0;    op->same_shape = 0;    op->infer_shape = infer_shape;    return 0;}static void release_op(struct op* op){
       sys_free(op->param_mem);}int register_absval_op(){
       struct method m;    m.version = 1;    m.init = init_op;               // 函数指针    m.release = release_op;         // 函数指针    return register_op(OP_ABSVAL, OP_ABSVAL_NAME, &m);          // OP_ABSVAL：enum， OP_ABSVAL_NAME：字符串}int unregister_absval_op(){
       return unregister_op(OP_ABSVAL, 1);}

7. source/serializer/tmfile/op文件夹下的所有op文件，是对node->op->param_mem进行赋值，上一步是申请内存初始化。如果用c++的话，可以把这两步的操作放到一个类中，框架更加清晰

static int batchnorm_op_map(int op){
       return OP_BATCHNORM;}static int tm2_load_batchnorm(struct graph* ir_graph, struct node* ir_node, const TM2_Node* tm_node, const TM2_Operator* tm_op){
       struct batchnorm_param* batchnorm_param = ( struct batchnorm_param* )ir_node->op.param_mem;    const struct tm2_priv* tm2_priv = (struct tm2_priv*)ir_graph->serializer_privacy;    const char* mem_base = tm2_priv->base;    const TM2_BatchNormParam* tm_param = ( TM2_BatchNormParam* )(mem_base + tm_op->offset_t_param);    batchnorm_param->rescale_factor = tm_param->rescale_factor;         // op 的参数    batchnorm_param->eps = tm_param->eps;    batchnorm_param->caffe_flavor = tm_param->caffe_flavor;    return 0;}

转载地址：http://cuhws.baihongyu.com/