Android 工程示例
Android工程示例用于展示 Tengine 基于 Android 系统的各种 CPU 架构的硬件后端运行网络模型推理。
编译
参考 源码编译（Android） 章节生成部署所需要的以下库文件：
build-android/install/lib/
└── libtengine-lite.so
运行
模型格式
CPU 后端支持加载 Float32/Float16/Uint8/Int8 tmfile，其中 Float16/Uint8/Int8 需要通过相应的模型量化工具获取。
推理精度设置
CPU 支持 Float32/Float16/Uint8/Int8 四种精度模型进行网络模型推理，需要在执行 prerun_graph_multithread(graph_t graph, struct options opt) 之前通过 struct options opt 显式设置推理精度。
Enable CPU FP32 mode
/* set runtime options */
struct options opt;
opt.num_thread = num_thread;
opt.cluster = TENGINE_CLUSTER_ALL;
opt.precision = TENGINE_MODE_FP32;
opt.affinity = 0;
Enable CPU FP16 mode
/* set runtime options */
struct options opt;
opt.num_thread = num_thread;
opt.cluster = TENGINE_CLUSTER_ALL;
opt.precision = TENGINE_MODE_FP16;
opt.affinity = 0;
Enable CPU Uint8 mode
/* set runtime options */
struct options opt;
opt.num_thread = num_thread;
opt.cluster = TENGINE_CLUSTER_ALL;
opt.precision = TENGINE_MODE_UINT8;
opt.affinity = 0;
Enable CPU Int8 mode
/* set runtime options */
struct options opt;
opt.num_thread = num_thread;
opt.cluster = TENGINE_CLUSTER_ALL;
opt.precision = TENGINE_MODE_INT8;
opt.affinity = 0;
参考 Demo
源码请参考 tm_classification.c
源码请参考 tm_classification_fp16.c
源码请参考 tm_classification_int8.c
使用 C API 预测
Android demo 大多数基于 C API 开发，调用 C API 大致分为以下几个步骤。更详细的 API 描述请参考：Tengine C API。
/* set runtime options */
struct options opt;
opt.num_thread = num_thread;
opt.cluster = TENGINE_CLUSTER_ALL;
opt.precision = TENGINE_MODE_FP32;
opt.affinity = affinity;
/* inital tengine */
init_tengine();
/* create graph, load tengine model xxx.tmfile */
graph_t graph = create_graph(NULL, "tengine", model_file);
/* set the shape, data buffer of input_tensor of the graph */
int img_size = img_h * img_w * 3;
int dims[] = {1, 3, img_h, img_w}; // nchw
float* input_data = ( float* )malloc(img_size * sizeof(float));
tensor_t input_tensor = get_graph_input_tensor(graph, 0, 0);
set_tensor_shape(input_tensor, dims, 4);
set_tensor_buffer(input_tensor, input_data, img_size * 4);
/* prerun graph, set work options(num_thread, cluster, precision) */
prerun_graph_multithread(graph, opt);
/* prepare process input data, set the data mem to input tensor */
get_input_data(image_file, input_data, img_h, img_w, mean, scale);
/* run graph */
run_graph(graph, 1);
/* get the result of classification */
tensor_t output_tensor = get_graph_output_tensor(graph, 0, 0);
float* output_data = ( float* )get_tensor_buffer(output_tensor);
/* release tengine */
free(input_data);
postrun_graph(graph);
destroy_graph(graph);
release_tengine();
使用 C++ API 预测
Android demo 同时提供 C++ API 简化开发流程，调用 C++ API 大致分为以下几个步骤。更详细的 API 描述请参考：Tengine C++ API。
/* inital tengine */
init_tengine();
tengine::Net somenet;
tengine::Tensor input_tensor;
tengine::Tensor output_tensor;
/* set runtime options of Net */
somenet.opt.num_thread = num_thread;
somenet.opt.cluster = TENGINE_CLUSTER_ALL;
somenet.opt.precision = TENGINE_MODE_FP32;
somenet.opt.affinity = affinity;
/* load model */
somenet.load_model(nullptr, "tengine", model_file.c_str());
/* prepare input data */
input_tensor.create(1, 3, img_h, img_w);
get_input_data(image_file.c_str(), ( float* )input_tensor.data, img_h, img_w, mean, scale);
/* set input data */
somenet.input_tensor("data", input_tensor);
/* forward */
somenet.run();
/* get result */
somenet.extract_tensor("prob", output_tensor);
/* release tengine */
release_tengine();
执行结果
使用adb 连接上Android 设备，以ubuntu环境为例，命令如下：
sudo apt install adb #安装adb，使电脑可以与Android设备通信。并查看Android设备的ip。
adb connect [安卓设备ip]
adb devices #确保可以看到设备
adb push tm_classification /data/local/tmp/
adb push cat.jpg /data/local/tmp/
adb push mobilenet.tmfile /data/local/tmp/
adb push libtengine-lite.so /data/local/tmp/
adb shell
#此时进入了Android设备的终端
cd /data/local/tmp
./tm_classification -m mobilenet.tmfile -i cat.jpg -g 224,224 -s 0.017,0.017,0.017 -w 104.007,116.669,122.679
start to run register cpu allocator
tengine-lite library version: 1.0-dev
model file : ./temp/models/mobilenet.tmfile
image file : ./temp/images/cat.jpg
img_h, img_w, scale[3], mean[3] : 224 224 , 0.017 0.017 0.017, 104.0 116.7 122.7
Repeat 1 times, thread 1, avg time 656.76 ms, max_time 656.76 ms, min_time 656.76 ms
--------------------------------------
8.574148, 282
7.880116, 277
7.812579, 278
7.286453, 263
6.357488, 281
--------------------------------------