Add a inference demo

inference/C/README.md

@@ -0,0 +1,14 @@
+# Inference demo
+
+This is an inference demo program based on the Paddle C API. But this demo is based on the c++ code, so need to use g++ or clang++ to compile. 
+The demo can be run from the command line and used to test the inference performance of various models.
+
+## Android
+To compile and run this demo in the Android environment, follow these steps:
+
+1. Refer to [this document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/cross_compiling/cross_compiling_for_android_cn.md) to compile the paddle of android version.
+2. Compile this inference.cc to an executable program for the Android environment.
+3. Run the demo program by logging into the Android environment via adb and specifying the paddle model from the command line.
+```
+./inference --merged_model ./model/mobilenet.paddle --input_size 150528
+```

inference/C/inference.cc

@@ -0,0 +1,171 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <iostream>
+#include <time.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <paddle/capi.h>
+
+inline paddle_error& operator |=(paddle_error& a, paddle_error b) {
+ return a =
+ static_cast<paddle_error>(static_cast<int>(a) | static_cast<int>(b));
+}
+
+class Timer {
+public:
+ Timer(std::string name, int iter = 1) : name_(name), iter_(iter) {
+ clock_gettime(CLOCK_MONOTONIC, &tp_start);
+ }
+ ~Timer() {
+ struct timespec tp_end;
+ clock_gettime(CLOCK_MONOTONIC, &tp_end);
+ float time = ((tp_end.tv_nsec - tp_start.tv_nsec)/1000000.0f);
+ time += (tp_end.tv_sec - tp_start.tv_sec)*1000;
+ time /= iter_;
+ std::cout << "Time of " << name_ << " " << time << " ms." << std::endl;
+ }
+
+private:
+ std::string name_;
+ int iter_;
+ struct timespec tp_start;
+};
+
+void read_file(const char* file, void** buf, long* size) {
+ FILE* fp = fopen(file, "r");
+ if (fp) {
+ if (fseek(fp, 0L, SEEK_END) == 0) {
+ *size = ftell(fp);
+ fseek(fp, 0L, SEEK_SET);
+ *buf = malloc(*size);
+ fread(*buf, 1, *size, fp);
+ }
+ }
+ fclose(fp);
+}
+
+int main(int argc, char* argv[]) {
+ // parse command line arguments
+ std::string predict_config;
+ std::string predict_model;
+ std::string merged_model;
+ int input_size;
+ for (int i = 1; i < argc; ++i) {
+ if (std::string(argv[i]) == "--predict_config") {
+ predict_config = std::string(argv[++i]);
+ } else if (std::string(argv[i]) == "--predict_model") {
+ predict_model = std::string(argv[++i]);
+ } else if (std::string(argv[i]) == "--merged_model") {
+ merged_model = std::string(argv[++i]);
+ } else if (std::string(argv[i]) == "--input_size") {
+ input_size = atoi(argv[++i]);
+ }
+ }
+
+ {
+ Timer time("init paddle");
+ char* argv[] = {"--use_gpu=False"};
+ if (paddle_init(1, (char**)argv) != kPD_NO_ERROR) {
+ std::cout << "paddle init error!" << std::endl;
+ }
+ }
+
+ // Create a gradient machine for inference.
+ paddle_gradient_machine machine;
+ paddle_error error = kPD_NO_ERROR;
+ if (!merged_model.empty()) {
+ Timer time("create from merged model file");
+ long size = 0;
+ void* buf = NULL;
+ read_file(merged_model.c_str(), &buf, &size);
+ paddle_gradient_machine_create_for_inference_with_parameters(
+ &machine, buf, size);
+ free(buf);
+ } else {
+ // Reading config binary file. It is generated by `convert_protobin.sh`
+ if (predict_config.empty()) return -1;
+ long size = 0;
+ void* buf = NULL;
+ {
+ Timer time("read model config");
+ read_file(predict_config.c_str(), &buf, &size);
+ }
+
+ error |= 
+ paddle_gradient_machine_create_for_inference(&machine, buf, (int)size);
+
+ if (predict_model.empty()) {
+ error |= paddle_gradient_machine_randomize_param(machine);
+ } else {
+ Timer time("load model parameter");
+ error |= paddle_gradient_machine_load_parameter_from_disk(
+ machine, predict_model.c_str());
+ }
+ free(buf);
+ }
+
+ if (error != kPD_NO_ERROR) {
+ std::cout << "paddle create inference machine error!" << std::endl;
+ }
+
+ // Create input matrix.
+ paddle_arguments in_args = paddle_arguments_create_none();
+ error |= paddle_arguments_resize(in_args, 1);
+ paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1,
+ /* size */ input_size,
+ /* useGPU */ false);
+ srand(time(0));
+ paddle_real* array;
+ // Get First row.
+ error |= paddle_matrix_get_row(mat, 0, &array);
+
+ for (int i = 0; i < input_size; ++i) {
+ array[i] = rand() / ((float)RAND_MAX);
+ }
+
+ error |= paddle_arguments_set_value(in_args, 0, mat);
+
+ paddle_arguments out_args = paddle_arguments_create_none();
+
+ if (error != kPD_NO_ERROR) {
+ std::cout << "paddle init input data!" << std::endl;
+ }
+
+ error |= paddle_gradient_machine_forward(machine,
+ in_args,
+ out_args,
+ /* isTrain */ false);
+
+ {
+ Timer time("forward time", 20);
+ for (int i = 0; i < 20; i++) {
+ error |= paddle_gradient_machine_forward(machine,
+ in_args,
+ out_args,
+ /* isTrain */ false);
+ }
+ }
+
+ if (error != kPD_NO_ERROR) {
+ std::cout << "paddle forward error!" << std::endl;
+ }
+
+ paddle_arguments_destroy(out_args);
+ paddle_matrix_destroy(mat);
+ paddle_arguments_destroy(in_args);
+ paddle_gradient_machine_destroy(machine);
+
+ return 0;
+}