tflite_micro移植到MCU

总述

MCU在现在的产品中随出可见，功能也越发于强大，价格也越趋于便宜。一提到人工智能，彷佛只有强大的硬件才能实现，需要什么神经网络加速器之类的硬件支撑。其实NPU、KPU等只是加快相关的计算能力，以达到较好的实时性，对于简单的、功能不复杂的神经网络结构，一般稍强的MCU完全跑得动。Tensorflow在深度学习平台中比较有名，目前已经适配了嵌入式平台，tflite_micro就比较适用于一般的MCU平台。

tflite_micro

完全由c++语言实现，可以说凡是带有c++编译器的MCU平台都可以移植使用。其中已经实现许多的神经网络算子，比如:卷积、池化等等。对于熟悉c++的可以使用c++开发，对于只会c语言的就可以直接使用c语言调用c++了。

移植步骤

1. 下载源码包
   见下文获取
2. 改一下MCU平台的编译规则，加入tflite_micro源码文件，以下以XR872平台作为示例

gcc.mk

%.o: %.cc
	$(Q)$(CPP) $(CC_FLAGS) $(CC_SYMBOLS) -std=c++11 -Wno-type-limits  -Wno-maybe-uninitialized -Wno-missing-field-initializers -Wstrict-aliasing -DTF_LITE_STATIC_MEMORY -DNDEBUG -Wsign-compare -Wdouble-promotion -Wshadow -Wunused-variable -Wmissing-field-initializers -Wunused-function -Wswitch -Wvla -O3  -Wno-return-type -Wno-strict-aliasing $(INCLUDE_PATHS) -o $@ $<


工程下的makefile文件
SRCS += $(basename $(foreach dir,$(DIRS),$(wildcard $(dir)/*.cc)))  #查找所有当前工程下的.cc文件

#包含源码的头文件路径
INCLUDE_PATHS += -I$(PRJ_ROOT_PATH)/tfmicro
INCLUDE_PATHS += -I$(PRJ_ROOT_PATH)/tfmicro/third_party/gemmlowp  
INCLUDE_PATHS += -I$(PRJ_ROOT_PATH)/tfmicro/third_party/flatbuffers/include  
INCLUDE_PATHS += -I$(PRJ_ROOT_PATH)/tfmicro/third_party/ruy

模型生成以及转换

模型生成

使用keras随便搭建一个模型(二分类等)，待训练完毕，保存为.h5模型文件。

model = Sequential([ #模型结构])

model.summary()

epochs=120

model.compile(optimizer='adam',
              loss=tf.keras.losses.BinaryCrossentropy(),
              metrics=['accuracy'])

#log_dir = "./" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard( histogram_freq=1)

model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(
    filepath="./trans_model",
    monitor='val_accuracy',
    mode='max',
    save_best_only=True)
#train_dataset 数据集
history = model.fit(
    train_dataset,
    steps_per_epoch=len(X_train) // batch_size,  #每个batch包含的样本数,一次epochs分多少步
    epochs=epochs,   #训练轮数
    validation_data=train_dataset,
    validation_steps=1,
    callbacks=[tensorboard_callback, model_checkpoint_callback]
)

model.save("trained.model.h5")

模型转换（h5-tflite）

#mnist_ds为模型数据集 data labels
def representative_dataset_gen():
    for input_value in mnist_ds.take(100):
        yield [input_value]


saved_model_dir= "trained.model.h5"
converter2 = tf.compat.v1.lite.TFLiteConverter.from_keras_model_file(model_file=saved_model_dir)
converter2.optimizations = [tf.lite.Optimize.DEFAULT]
converter2.representative_dataset = tf.lite.RepresentativeDataset(representative_dataset_gen)
converter2.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
tflite_quant_model = converter2.convert()
open("converted_model.tflite", "wb").write(tflite_quant_model)

转化tflite文件为c语言数组

linux 下使用xxd -i命令可以将文件转化为数据格式，如下：

C语言调用模型

// 以下代码参考过github
const int kArenaSize = 55000; //模型需要的内存空间

NeuralNetwork::NeuralNetwork()
{
    #if  1
    m_error_reporter = new tflite::MicroErrorReporter();

    m_tensor_arena = (uint8_t *)malloc(kArenaSize);
    if((int)m_tensor_arena%16 != 0)
    {
         m_tensor_arena = m_tensor_arena + (16-(int)m_tensor_arena%16);
    }
    if (!m_tensor_arena)
    {
        TF_LITE_REPORT_ERROR(m_error_reporter, "Could not allocate arena");
        return;
    }
    TF_LITE_REPORT_ERROR(m_error_reporter, "Loading model");

    m_model = tflite::GetModel(converted_model_tflite);
    if (m_model->version() != TFLITE_SCHEMA_VERSION)
    {
        TF_LITE_REPORT_ERROR(m_error_reporter, "Model provided is schema version %d not equal to supported version %d.",
                             m_model->version(), TFLITE_SCHEMA_VERSION);
        return;
    }
    // This pulls in the operators implementations we need
    m_resolver = new tflite::MicroMutableOpResolver<10>();
    m_resolver->AddConv2D();
    m_resolver->AddDepthwiseConv2D();
    m_resolver->AddMaxPool2D();
    m_resolver->AddFullyConnected();
    m_resolver->AddMul();
    m_resolver->AddAdd();
    m_resolver->AddLogistic();
    m_resolver->AddReshape();
    m_resolver->AddQuantize();
    m_resolver->AddDequantize();

    // Build an interpreter to run the model with.
    m_interpreter = new tflite::MicroInterpreter(
        m_model, *m_resolver, m_tensor_arena, kArenaSize, m_error_reporter);

    // Allocate memory from the tensor_arena for the model's tensors.
    TfLiteStatus allocate_status = m_interpreter->AllocateTensors();
    if (allocate_status != kTfLiteOk)
    {
        TF_LITE_REPORT_ERROR(m_error_reporter, "AllocateTensors() failed");
        return;
    }
    
    size_t used_bytes = m_interpreter->arena_used_bytes();
    TF_LITE_REPORT_ERROR(m_error_reporter, "Used bytes %d\n", used_bytes);

    // Obtain pointers to the model's input and output tensors.
    input = m_interpreter->input(0);
    output = m_interpreter->output(0);
   // TfliteIntArray *dims = m_interpreter->tensor(m_interpreter->inputs()[0])->dims;
   // printf("height:%d width:%d channels:%d\r\n", dims->data[1], dims->data[2],dims->data[3]);
   //printf("input len:%d\r\n", m_interpreter->inputs_size());
    int uinput = m_interpreter->inputs()[0];
// get the dims of input tensor
    TfLiteIntArray* dims = m_interpreter->tensor(uinput)->dims;
    int wanted_height = dims->data[1];
    int wanted_width = dims->data[2];
    int wanted_channels = dims->data[3];
    printf("wanted_height:%d, wanted_width:%d, wanted_channels:%d\r\n", wanted_height, wanted_width, wanted_channels);
    #endif
}

NeuralNetwork::~NeuralNetwork()
{
    free(m_tensor_arena);
}

float *NeuralNetwork::getInputBuffer()
{
    return input->data.f;
}

float NeuralNetwork::predict()
{
#if  1
    m_interpreter->Invoke();
    return output->data.f[0];
 #endif
}

资源获取

关注公众号发送 872_tflite
代码仓库.