mirror of https://github.com/alibaba/MNN.git
267 lines
11 KiB
Markdown
267 lines
11 KiB
Markdown
|
# 使用表达式接口训练
|
|||
|
|
[从源码编译](../compile/tools.html#id3)
|
|||
|
|
## 获取可训练模型
|
|||
|
|
首先我们需要获得一个可训练的模型结构，而在MNN中可以用两种方式达到这一目的。
|
|||
|
|
|
|||
|
|
一、将其他框架，如TensorFlow，Pytorch训练得到的模型转成MNN可训练模型，这一过程可使用 MNNConverter 工具实现。典型的应用场景为 1. 使用MNN Finetune，2. 使用MNN进行训练量化。在模型转换过程中建议使用 MNNConverter 的 --forTraining 选项，保留BatchNorm，Dropout等训练过程中会用到的算子。
|
|||
|
|
|
|||
|
|
二、使用MNN从零开始搭建一个模型，并使用MNN进行训练，这可以省去模型转换的步骤，并且也可以十分容易地转换为训练量化模型。在 MNN_ROOT/tools/train/source/models/ 目录中我们提供了Lenet，MobilenetV1，MobilenetV2等使用MNN训练框架搭建的模型。
|
|||
|
|
|
|||
|
|
### 将其他框架模型转换为MNN可训练模型
|
|||
|
|
以MobilenetV2的训练量化为例。首先我们需要到下载TensorFlow官方提供的MobilenetV2模型，然后编译 MNNConverter，并执行以下命令进行转换：
|
|||
|
|
```bash
|
|||
|
|
./MNNConvert --modelFile mobilenet_v2_1.0_224_frozen.pb --MNNModel mobilenet_v2_tfpb_train.mnn --framework TF --bizCode AliNNTest --forTraining
|
|||
|
|
```
|
|||
|
|
注意，上述命令中使用到的 mobilenet_v2_1.0_224_frozen.pb 模型中含有 BatchNorm 算子，没有进行融合，通过在转换时使用 --forTraining 选项，我们保留了BatchNorm算子到转换出来的 mobilenet_v2_tfpb_train.mnn 模型之中。
|
|||
|
|
|
|||
|
|
如果你的模型中没有BN，Dropout等在转MNN推理模型时会被融合掉的算子，那么直接使用MNN推理模型也可以进行训练，不必重新进行转换。
|
|||
|
|
|
|||
|
|
接下来我们仿照 MNN_ROOT/tools/train/source/demo/mobilenetV2Train.cpp 中的示例，读取转换得到的模型，将其转换为MNN可训练模型。关键代码示例如下
|
|||
|
|
```cpp
|
|||
|
|
// mobilenetV2Train.cpp
|
|||
|
|
|
|||
|
|
// 读取转换得到的MNN模型
|
|||
|
|
auto varMap = Variable::loadMap(argv[1]);
|
|||
|
|
// 指定量化bit数
|
|||
|
|
int bits = 8;
|
|||
|
|
// 获取模型的输入和输出
|
|||
|
|
auto inputOutputs = Variable::getInputAndOutput(varMap);
|
|||
|
|
auto inputs = Variable::mapToSequence(inputOutputs.first);
|
|||
|
|
auto outputs = Variable::mapToSequence(inputOutputs.second);
|
|||
|
|
|
|||
|
|
// 将转换得到的模型转换为可训练模型(将推理模型中的卷积，BatchNorm，Dropout抽取出来，转换成可训练模块)
|
|||
|
|
std::shared_ptr<Module> model(PipelineModule::extract(inputs, outputs, true));
|
|||
|
|
// 将可训练模型转换为训练量化模型，如果不需要进行训练量化，则可不做这一步
|
|||
|
|
((PipelineModule*)model.get())->toTrainQuant(bits);
|
|||
|
|
// 进入训练环节
|
|||
|
|
MobilenetV2Utils::train(model, 1001, 1, trainImagesFolder, trainImagesTxt, testImagesFolder, testImagesTxt);
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### 使用MNN从零开始搭建模型
|
|||
|
|
以Lenet为例，我们来看一下，如何使用MNN从零搭建一个模型。MNN提供了丰富的算子可供使用，下面的例子就不详细展开。值得注意的是Pooling输出为NC4HW4格式，需要转换到 NCHW 格式才能进入全连接层进行计算。
|
|||
|
|
```cpp
|
|||
|
|
class MNN_PUBLIC Lenet : public Module {
|
|||
|
|
public:
|
|||
|
|
Lenet();
|
|||
|
|
|
|||
|
|
virtual std::vector<Express::VARP> onForward(const std::vector<Express::VARP>& inputs) override;
|
|||
|
|
|
|||
|
|
std::shared_ptr<Module> conv1;
|
|||
|
|
std::shared_ptr<Module> conv2;
|
|||
|
|
std::shared_ptr<Module> ip1;
|
|||
|
|
std::shared_ptr<Module> ip2;
|
|||
|
|
std::shared_ptr<Module> dropout;
|
|||
|
|
};
|
|||
|
|
// 初始化
|
|||
|
|
Lenet::Lenet() {
|
|||
|
|
NN::ConvOption convOption;
|
|||
|
|
convOption.kernelSize = {5, 5};
|
|||
|
|
convOption.channel = {1, 20};
|
|||
|
|
conv1.reset(NN::Conv(convOption));
|
|||
|
|
convOption.reset();
|
|||
|
|
convOption.kernelSize = {5, 5};
|
|||
|
|
convOption.channel = {20, 50};
|
|||
|
|
conv2.reset(NN::Conv(convOption));
|
|||
|
|
ip1.reset(NN::Linear(800, 500));
|
|||
|
|
ip2.reset(NN::Linear(500, 10));
|
|||
|
|
dropout.reset(NN::Dropout(0.5));
|
|||
|
|
// 必须要进行register的参数才会进行更新
|
|||
|
|
registerModel({conv1, conv2, ip1, ip2, dropout});
|
|||
|
|
}
|
|||
|
|
// 前向计算
|
|||
|
|
std::vector<Express::VARP> Lenet::onForward(const std::vector<Express::VARP>& inputs) {
|
|||
|
|
using namespace Express;
|
|||
|
|
VARP x = inputs[0];
|
|||
|
|
x = conv1->forward(x);
|
|||
|
|
x = _MaxPool(x, {2, 2}, {2, 2});
|
|||
|
|
x = conv2->forward(x);
|
|||
|
|
x = _MaxPool(x, {2, 2}, {2, 2});
|
|||
|
|
// Pooling输出为NC4HW4格式，需要转换到NCHW才能进入全连接层进行计算
|
|||
|
|
x = _Convert(x, NCHW);
|
|||
|
|
x = _Reshape(x, {0, -1});
|
|||
|
|
x = ip1->forward(x);
|
|||
|
|
x = _Relu(x);
|
|||
|
|
x = dropout->forward(x);
|
|||
|
|
x = ip2->forward(x);
|
|||
|
|
x = _Softmax(x, 1);
|
|||
|
|
return {x};
|
|||
|
|
}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## 实现数据集接口
|
|||
|
|
这部分在MNN文档中[加载训练数据](train.html#id8) 部分有详细描述。
|
|||
|
|
|
|||
|
|
## 训练并保存模型
|
|||
|
|
以MNIST模型训练为例，代码在 MNN_ROOT/tools/train/source/demo/MnistUtils.cpp
|
|||
|
|
|
|||
|
|
```cpp
|
|||
|
|
// MnistUtils.cpp
|
|||
|
|
|
|||
|
|
......
|
|||
|
|
|
|||
|
|
void MnistUtils::train(std::shared_ptr<Module> model, std::string root) {
|
|||
|
|
{
|
|||
|
|
// Load snapshot
|
|||
|
|
// 模型结构 + 模型参数
|
|||
|
|
auto para = Variable::load("mnist.snapshot.mnn");
|
|||
|
|
model->loadParameters(para);
|
|||
|
|
}
|
|||
|
|
// 配置训练框架参数
|
|||
|
|
auto exe = Executor::getGlobalExecutor();
|
|||
|
|
BackendConfig config;
|
|||
|
|
// 使用CPU，4线程
|
|||
|
|
exe->setGlobalExecutorConfig(MNN_FORWARD_CPU, config, 4);
|
|||
|
|
// SGD求解器
|
|||
|
|
std::shared_ptr<SGD> sgd(new SGD(model));
|
|||
|
|
// SGD求解器参数设置
|
|||
|
|
sgd->setMomentum(0.9f);
|
|||
|
|
sgd->setWeightDecay(0.0005f);
|
|||
|
|
|
|||
|
|
// 创建数据集和DataLoader
|
|||
|
|
auto dataset = MnistDataset::create(root, MnistDataset::Mode::TRAIN);
|
|||
|
|
// the stack transform, stack [1, 28, 28] to [n, 1, 28, 28]
|
|||
|
|
const size_t batchSize = 64;
|
|||
|
|
const size_t numWorkers = 0;
|
|||
|
|
bool shuffle = true;
|
|||
|
|
|
|||
|
|
auto dataLoader = std::shared_ptr<DataLoader>(dataset.createLoader(batchSize, true, shuffle, numWorkers));
|
|||
|
|
|
|||
|
|
size_t iterations = dataLoader->iterNumber();
|
|||
|
|
|
|||
|
|
auto testDataset = MnistDataset::create(root, MnistDataset::Mode::TEST);
|
|||
|
|
const size_t testBatchSize = 20;
|
|||
|
|
const size_t testNumWorkers = 0;
|
|||
|
|
shuffle = false;
|
|||
|
|
|
|||
|
|
auto testDataLoader = std::shared_ptr<DataLoader>(testDataset.createLoader(testBatchSize, true, shuffle, testNumWorkers));
|
|||
|
|
|
|||
|
|
size_t testIterations = testDataLoader->iterNumber();
|
|||
|
|
|
|||
|
|
// 开始训练
|
|||
|
|
for (int epoch = 0; epoch < 50; ++epoch) {
|
|||
|
|
model->clearCache();
|
|||
|
|
exe->gc(Executor::FULL);
|
|||
|
|
exe->resetProfile();
|
|||
|
|
{
|
|||
|
|
AUTOTIME;
|
|||
|
|
dataLoader->reset();
|
|||
|
|
// 训练阶段需设置isTraining Flag为true
|
|||
|
|
model->setIsTraining(true);
|
|||
|
|
Timer _100Time;
|
|||
|
|
int lastIndex = 0;
|
|||
|
|
int moveBatchSize = 0;
|
|||
|
|
for (int i = 0; i < iterations; i++) {
|
|||
|
|
// AUTOTIME;
|
|||
|
|
// 获得一个batch的数据，包括数据及其label
|
|||
|
|
auto trainData = dataLoader->next();
|
|||
|
|
auto example = trainData[0];
|
|||
|
|
auto cast = _Cast<float>(example.first[0]);
|
|||
|
|
example.first[0] = cast * _Const(1.0f / 255.0f);
|
|||
|
|
moveBatchSize += example.first[0]->getInfo()->dim[0];
|
|||
|
|
|
|||
|
|
// Compute One-Hot
|
|||
|
|
auto newTarget = _OneHot(_Cast<int32_t>(example.second[0]), _Scalar<int>(10), _Scalar<float>(1.0f),
|
|||
|
|
_Scalar<float>(0.0f));
|
|||
|
|
// 前向计算
|
|||
|
|
auto predict = model->forward(example.first[0]);
|
|||
|
|
// 计算loss
|
|||
|
|
auto loss = _CrossEntropy(predict, newTarget);
|
|||
|
|
// 调整学习率
|
|||
|
|
float rate = LrScheduler::inv(0.01, epoch * iterations + i, 0.0001, 0.75);
|
|||
|
|
sgd->setLearningRate(rate);
|
|||
|
|
|
|||
|
|
if (moveBatchSize % (10 * batchSize) == 0 || i == iterations - 1) {
|
|||
|
|
std::cout << "epoch: " << (epoch);
|
|||
|
|
std::cout << " " << moveBatchSize << " / " << dataLoader->size();
|
|||
|
|
std::cout << " loss: " << loss->readMap<float>()[0];
|
|||
|
|
std::cout << " lr: " << rate;
|
|||
|
|
std::cout << " time: " << (float)_100Time.durationInUs() / 1000.0f << " ms / " << (i - lastIndex) << " iter" << std::endl;
|
|||
|
|
std::cout.flush();
|
|||
|
|
_100Time.reset();
|
|||
|
|
lastIndex = i;
|
|||
|
|
}
|
|||
|
|
// 根据loss反向计算，并更新网络参数
|
|||
|
|
sgd->step(loss);
|
|||
|
|
}
|
|||
|
|
}
|
|||
|
|
// 保存模型参数，便于重新载入训练
|
|||
|
|
Variable::save(model->parameters(), "mnist.snapshot.mnn");
|
|||
|
|
{
|
|||
|
|
model->setIsTraining(false);
|
|||
|
|
auto forwardInput = _Input({1, 1, 28, 28}, NC4HW4);
|
|||
|
|
forwardInput->setName("data");
|
|||
|
|
auto predict = model->forward(forwardInput);
|
|||
|
|
predict->setName("prob");
|
|||
|
|
// 优化网络结构【可选】
|
|||
|
|
Transformer::turnModelToInfer()->onExecute({predict});
|
|||
|
|
// 保存模型和结构，可脱离Module定义使用
|
|||
|
|
Variable::save({predict}, "temp.mnist.mnn");
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
// 测试模型
|
|||
|
|
int correct = 0;
|
|||
|
|
testDataLoader->reset();
|
|||
|
|
// 测试时，需设置标志位
|
|||
|
|
model->setIsTraining(false);
|
|||
|
|
int moveBatchSize = 0;
|
|||
|
|
for (int i = 0; i < testIterations; i++) {
|
|||
|
|
auto data = testDataLoader->next();
|
|||
|
|
auto example = data[0];
|
|||
|
|
moveBatchSize += example.first[0]->getInfo()->dim[0];
|
|||
|
|
if ((i + 1) % 100 == 0) {
|
|||
|
|
std::cout << "test: " << moveBatchSize << " / " << testDataLoader->size() << std::endl;
|
|||
|
|
}
|
|||
|
|
auto cast = _Cast<float>(example.first[0]);
|
|||
|
|
example.first[0] = cast * _Const(1.0f / 255.0f);
|
|||
|
|
auto predict = model->forward(example.first[0]);
|
|||
|
|
predict = _ArgMax(predict, 1);
|
|||
|
|
auto accu = _Cast<int32_t>(_Equal(predict, _Cast<int32_t>(example.second[0]))).sum({});
|
|||
|
|
correct += accu->readMap<int32_t>()[0];
|
|||
|
|
}
|
|||
|
|
// 计算准确率
|
|||
|
|
auto accu = (float)correct / (float)testDataLoader->size();
|
|||
|
|
std::cout << "epoch: " << epoch << " accuracy: " << accu << std::endl;
|
|||
|
|
exe->dumpProfile();
|
|||
|
|
}
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## 保存和恢复模型
|
|||
|
|
一、只保存模型参数，不保存模型结构，需要对应的模型结构去加载这些参数
|
|||
|
|
保存：
|
|||
|
|
```cpp
|
|||
|
|
Variable::save(model->parameters(), "mnist.snapshot.mnn");
|
|||
|
|
```
|
|||
|
|
恢复：
|
|||
|
|
```cpp
|
|||
|
|
// 模型结构 + 模型参数
|
|||
|
|
auto para = Variable::load("mnist.snapshot.mnn");
|
|||
|
|
model->loadParameters(para);
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
二、同时保存模型结构和参数，便于推理
|
|||
|
|
保存：
|
|||
|
|
```cpp
|
|||
|
|
model->setIsTraining(false);
|
|||
|
|
auto forwardInput = _Input({1, 1, 28, 28}, NC4HW4);
|
|||
|
|
forwardInput->setName("data");
|
|||
|
|
auto predict = model->forward(forwardInput);
|
|||
|
|
predict->setName("prob");
|
|||
|
|
|
|||
|
|
// 保存输出节点，会连同结构参数一并存储下来
|
|||
|
|
Variable::save({predict}, "temp.mnist.mnn");
|
|||
|
|
```
|
|||
|
|
恢复（进行推理）：
|
|||
|
|
```cpp
|
|||
|
|
auto varMap = Variable::loadMap("temp.mnist.mnn");
|
|||
|
|
|
|||
|
|
//输入节点名与保存时设定的名字一致，为 data，维度大小与保存时设定的大小一致，为 [1, 1, 28, 28]
|
|||
|
|
float* inputPtr = varMap["data"]->writeMap<float>();
|
|||
|
|
//填充 inputPtr
|
|||
|
|
|
|||
|
|
//输出节点名与保存时设定的名字一致，为 prob
|
|||
|
|
float* outputPtr = varMap["prob"]->readMap<float>();
|
|||
|
|
// 使用 outputPtr 的数据
|
|||
|
|
|
|||
|
|
```
|