使用的是这个工程 /
1.数据集标注
2.模型训练
for obj in objects:
print(filename)
class_name = obj.find('name').text
class_name = class_name.lower()
if class_name == 'thanks-1':
class_name = 'thanks'
class_id = self.classes.index(class_name)
3.利用tensorflow lite布置到移动端
a.
上面的demo如果要使用非量化模型需要做如下改动
// If you have your own model, modify this to the file name, and make sure
// you've added the file to your app resources too.
//static NSString* model_file_name = @"mobilenet_quant_v1_224";
static NSString* model_file_name = @"mobilenet_v1_1.0_224";
static NSString* model_file_type = @"tflite";
// Returns the top N confidence values over threshold in the provided vector,
// sorted by confidence in descending order.
static void GetTopN_float(const float* prediction, const int prediction_size, const int num_results,
const float threshold, std::vector<std::pair<float, int>>* top_results) {
// Will contain top N results in ascending order.
std::priority_queue<std::pair<float, int>, std::vector<std::pair<float, int>>,
std::greater<std::pair<float, int>>>
top_result_pq;
const long count = prediction_size;
for (int i = 0; i < count; ++i) {
const float value = prediction[i];//#######不用除以255.0
// Only add it if it beats the threshold and has a chance at being in
// the top N.
if (value < threshold) {
continue;
}
top_result_pq.push(std::pair<float, int>(value, i));
// If at capacity, kick the smallest value out.
if (top_result_pq.size() > num_results) {
top_result_pq.pop();
}
}
// Copy to output vector and reverse into descending order.
while (!top_result_pq.empty()) {
top_results->push_back(top_result_pq.top());
top_result_pq.pop();
}
std::reverse(top_results->begin(), top_results->end());
}
//注意如果相机是默认设置的话，pixelBuffer得到的图片是横着的
- (void)runModelOnFrame_float:(CVPixelBufferRef)pixelBuffer {
assert(pixelBuffer != NULL);
OSType sourcePixelFormat = CVPixelBufferGetPixelFormatType(pixelBuffer);
assert(sourcePixelFormat == kCVPixelFormatType_32ARGB ||
sourcePixelFormat == kCVPixelFormatType_32BGRA);
const int sourceRowBytes = (int)CVPixelBufferGetBytesPerRow(pixelBuffer);
const int image_width = (int)CVPixelBufferGetWidth(pixelBuffer);
const int fullHeight = (int)CVPixelBufferGetHeight(pixelBuffer);
CVPixelBufferLockFlags unlockFlags = kNilOptions;
CVPixelBufferLockBaseAddress(pixelBuffer, unlockFlags);
unsigned char* sourceBaseAddr = (unsigned char*)(CVPixelBufferGetBaseAddress(pixelBuffer));
int image_height;
unsigned char* sourceStartAddr;
if (fullHeight <= image_width) {
image_height = fullHeight;
sourceStartAddr = sourceBaseAddr;
} else {
image_height = image_width;
const int marginY = ((fullHeight - image_width) / 2);
sourceStartAddr = (sourceBaseAddr + (marginY * sourceRowBytes));
}
const int image_channels = 4;
assert(image_channels >= wanted_input_channels);
uint8_t* in = sourceStartAddr;
//根据当前问题填充输入张量
//得到输入张量数组中的第一个张量，也就是classifier中唯一的那个输入张量,input是个整型值，语义是张量列表中的引索。第二条语句有两个作用，
//1）以input为索引，在TfLiteTensor* content_.tensors这个张量表得到具体的张量。
int input = interpreter->inputs()[0];
//2）返回该张量data.raw，它指示张量正关联着的内存块。有了out，app就把可以把要预测的图像数据填向它了。
float* out = interpreter->typed_tensor<float>(input);
const float input_mean = 127.5f;
const float input_std = 127.5f;
for (int y = 0; y < wanted_input_height; ++y) {
float* out_row = out + (y * wanted_input_width * wanted_input_channels);
for (int x = 0; x < wanted_input_width; ++x) {
const int in_x = (y * image_width) / wanted_input_width;
const int in_y = (x * image_height) / wanted_input_height;
uint8_t* in_pixel = in + (in_y * image_width * image_channels) + (in_x * image_channels);
float* out_pixel = out_row + (x * wanted_input_channels);
for (int c = 0; c < wanted_input_channels; ++c) {
//#############归一化
out_pixel[c] = (in_pixel[c] - input_mean) / input_std;
}
}
}
//调用Invoke进行预测
double startTimestamp = [[NSDate new] timeIntervalSince1970];
if (interpreter->Invoke() != kTfLiteOk) {
LOG(FATAL) << "Failed to invoke!";
}
double endTimestamp = [[NSDate new] timeIntervalSince1970];
total_latency += (endTimestamp - startTimestamp);
total_count += 1;
NSLog(@"Time: %.4lf, avg: %.4lf, count: %d", endTimestamp - startTimestamp,
total_latency / total_count, total_count);
const int output_size = 1000;
const int kNumResults = 5;
const float kThreshold = 0.1f;
std::vector<std::pair<float, int>> top_results;
//解析输出张量得到识别结果
//作用是得到输出张量关联的内存块。output存放的数据已是一维数组，之后就可用它得到识别结果了
float* output = interpreter->typed_output_tensor<float>(0);
//GetTopN用于计算output数组中最大的N个值(first)，以及它们的位置(second)
GetTopN_float(output, output_size, kNumResults, kThreshold, &top_results);
NSMutableDictionary* newValues = [NSMutableDictionary dictionary];
for (const auto& result : top_results) {
const float confidence = result.first;
const int index = result.second;
NSString* labelObject = [NSString stringWithUTF8String:labels[index].c_str()];
NSNumber* valueObject = [NSNumber numberWithFloat:confidence];
[newValues setObject:valueObject forKey:labelObject];
}
dispatch_async(dispatch_get_main_queue(), ^(void) {
[self setPredictionValues:newValues];
});
CVPixelBufferUnlockBaseAddress(pixelBuffer, unlockFlags);
CVPixelBufferUnlockBaseAddress(pixelBuffer, 0);
}
运行非量化模型预测图片
cd tensorflow/contrib/lite/examples/ios/simple
pod install
pod update
open tflite_simple_example.xcworkspace
b.生成TFLite文件,要先安装tensorflow源码和bazel方法,编译tensorflow源码生成tensorflow
bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
如果tensorflow环境已经搭建好的话，执行到上一步就可以了
如果build过程失败了，我们也可以执行如下指令，只生成我们在模型转化时需要的执行文件
.bazel build tensorflow/python/tools:freeze_graph
bazel build --config=opt tensorflow/contrib/lite/toco:toco
bazel build tensorflow/tools/graph_transforms:summarize_graph
bazel build tensorflow/tools/quantization:quantize_graph
bazel-bin/tensorflow/contrib/lite/toco/toco --input_file=quantized_graph.pb --input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE --output_file=quantized_graphh.tflite --inference_type=QUANTIZED_UINT8 --inference_input_type=QUANTIZED_UINT8 --input_shapes=1,224,224,3 --input_arrays=input --output_arrays=MobilenetV1/Predictions/Reshape_1 --mean_values=128 --std_values=128 --default_ranges_min=0 --default_ranges_max=6