飞桨高层API使用指南¶

print("视觉相关数据集：", paddle.vision.datasets.__all__)
print("自然语言相关数据集：", paddle.text.__all__)

视觉相关数据集： ['DatasetFolder', 'ImageFolder', 'MNIST', 'FashionMNIST', 'Flowers', 'Cifar10', 'Cifar100', 'VOC2012']
自然语言相关数据集： ['Conll05st', 'Imdb', 'Imikolov', 'Movielens', 'UCIHousing', 'WMT14', 'WMT16', 'ViterbiDecoder', 'viterbi_decode']

from paddle.vision.transforms import ToTensor

# 训练数据集
train_dataset = vision.datasets.MNIST(mode="train", transform=ToTensor())

# 验证数据集
val_dataset = vision.datasets.MNIST(mode="test", transform=ToTensor())

from paddle.io import Dataset


class MyDataset(Dataset):
    """
    步骤一：继承paddle.io.Dataset类
    """

    def __init__(self, mode="train"):
        """
        步骤二：实现构造函数，定义数据读取方式，划分训练和测试数据集
        """
        super().__init__()

        if mode == "train":
            self.data = [
                ["traindata1", "label1"],
                ["traindata2", "label2"],
                ["traindata3", "label3"],
                ["traindata4", "label4"],
            ]
        else:
            self.data = [
                ["testdata1", "label1"],
                ["testdata2", "label2"],
                ["testdata3", "label3"],
                ["testdata4", "label4"],
            ]

    def __getitem__(self, index):
        """
        步骤三：实现__getitem__方法，定义指定index时如何获取数据，并返回单条数据（训练数据，对应的标签）
        """
        data = self.data[index][0]
        label = self.data[index][1]

        return data, label

    def __len__(self):
        """
        步骤四：实现__len__方法，返回数据集总数目
        """
        return len(self.data)


# 测试定义的数据集
train_dataset_2 = MyDataset(mode="train")
val_dataset_2 = MyDataset(mode="test")

print("=============train dataset=============")
for data, label in train_dataset_2:
    print(data, label)

print("=============evaluation dataset=============")
for data, label in val_dataset_2:
    print(data, label)

=============train dataset=============
traindata1 label1
traindata2 label2
traindata3 label3
traindata4 label4
=============evaluation dataset=============
testdata1 label1
testdata2 label2
testdata3 label3
testdata4 label4

# Sequential形式组网
mnist = paddle.nn.Sequential(
    paddle.nn.Flatten(),
    paddle.nn.Linear(784, 512),
    paddle.nn.ReLU(),
    paddle.nn.Dropout(0.2),
    paddle.nn.Linear(512, 10),
)

# Layer类继承方式组网
class Mnist(paddle.nn.Layer):
    def __init__(self):
        super().__init__()

        self.flatten = paddle.nn.Flatten()
        self.linear_1 = paddle.nn.Linear(784, 512)
        self.linear_2 = paddle.nn.Linear(512, 10)
        self.relu = paddle.nn.ReLU()
        self.dropout = paddle.nn.Dropout(0.2)

    def forward(self, inputs):
        y = self.flatten(inputs)
        y = self.linear_1(y)
        y = self.relu(y)
        y = self.dropout(y)
        y = self.linear_2(y)

        return y


mnist_2 = Mnist()

# 使用GPU训练
# paddle.set_device('gpu')

# 模型封装

## 场景1：动态图模式
## 1.1 为模型预测部署场景进行模型训练
## 需要添加input和label数据描述，否则会导致使用model.save(training=False)保存的预测模型在使用时出错
inputs = paddle.static.InputSpec([-1, 1, 28, 28], dtype="float32", name="input")
label = paddle.static.InputSpec([-1, 1], dtype="int8", name="label")
model = paddle.Model(mnist, inputs, label)

## 1.2 面向实验而进行的模型训练
## 可以不传递input和label信息
# model = paddle.Model(mnist)

## 场景2：静态图模式
# paddle.enable_static()
# paddle.set_device('gpu')
# input = paddle.static.InputSpec([None, 1, 28, 28], dtype='float32')
# label = paddle.static.InputSpec([None, 1], dtype='int8')
# model = paddle.Model(mnist, input, label)

model.summary((1, 28, 28))

---------------------------------------------------------------------------
 Layer (type)       Input Shape          Output Shape         Param #    
===========================================================================
   Flatten-1       [[1, 28, 28]]           [1, 784]              0       
   Linear-1          [[1, 784]]            [1, 512]           401,920    
    ReLU-1           [[1, 512]]            [1, 512]              0       
   Dropout-1         [[1, 512]]            [1, 512]              0       
   Linear-2          [[1, 512]]            [1, 10]             5,130     
===========================================================================
Total params: 407,050
Trainable params: 407,050
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.02
Params size (MB): 1.55
Estimated Total Size (MB): 1.57
---------------------------------------------------------------------------






{'total_params': 407050, 'trainable_params': 407050}

paddle.summary(mnist, (1, 28, 28))

---------------------------------------------------------------------------
 Layer (type)       Input Shape          Output Shape         Param #    
===========================================================================
   Flatten-1       [[1, 28, 28]]           [1, 784]              0       
   Linear-1          [[1, 784]]            [1, 512]           401,920    
    ReLU-1           [[1, 512]]            [1, 512]              0       
   Dropout-1         [[1, 512]]            [1, 512]              0       
   Linear-2          [[1, 512]]            [1, 10]             5,130     
===========================================================================
Total params: 407,050
Trainable params: 407,050
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.02
Params size (MB): 1.55
Estimated Total Size (MB): 1.57
---------------------------------------------------------------------------






{'total_params': 407050, 'trainable_params': 407050}

# 使用GPU训练
# paddle.set_device('gpu')

# 构建模型训练用的Model，告知需要训练哪个模型
model = paddle.Model(mnist)

# 为模型训练做准备，设置优化器，损失函数和精度计算方式
model.prepare(
    paddle.optimizer.Adam(parameters=model.parameters()),
    paddle.nn.CrossEntropyLoss(),
    paddle.metric.Accuracy(),
)

# 启动模型训练，指定训练数据集，设置训练轮次，设置每次数据集计算的批次大小，设置日志格式
model.fit(train_dataset, epochs=5, batch_size=64, verbose=1)

The loss value printed in the log is the current step, and the metric is the average value of previous steps.
Epoch 1/5
step 938/938 [==============================] - loss: 0.0060 - acc: 0.9871 - 27ms/step        
Epoch 2/5
step 938/938 [==============================] - loss: 0.0042 - acc: 0.9905 - 27ms/step         
Epoch 3/5
step 938/938 [==============================] - loss: 0.0289 - acc: 0.9911 - 27ms/step        
Epoch 4/5
step 938/938 [==============================] - loss: 0.0018 - acc: 0.9929 - 27ms/step        
Epoch 5/5
step 938/938 [==============================] - loss: 0.0394 - acc: 0.9935 - 27ms/step           

class SelfDefineLoss(paddle.nn.Layer):
    """
    1. 继承paddle.nn.Layer
    """
    def __init__(self):
        """
        2. 构造函数根据自己的实际算法需求和使用需求进行参数定义即可
        """
        super().__init__()

    def forward(self, input, label):
        """
        3. 实现forward函数，forward在调用时会传递两个参数：input和label
            - input：单个或批次训练数据经过模型前向计算输出结果
            - label：单个或批次训练数据对应的标签数据

            接口返回值是一个Tensor，根据自定义的逻辑加和或计算均值后的损失
        """
        # 使用Paddle中相关API自定义的计算逻辑
        # output = xxxxx
        # return output

class SoftmaxWithCrossEntropy(paddle.nn.Layer):
    def __init__(self):
        super().__init__()

    def forward(self, input, label):
        loss = F.softmax_with_cross_entropy(input, 
                                            label, 
                                            return_softmax=False,
                                            axis=1)
        return paddle.mean(loss)

class SelfDefineMetric(paddle.metric.Metric):
    """
    1. 继承paddle.metric.Metric
    """
    def __init__(self):
        """
        2. 构造函数实现，自定义参数即可
        """
        super().__init__()

    def name(self):
        """
        3. 实现name方法，返回定义的评估指标名字
        """
        return '自定义评价指标的名字'

    def compute(self, ...)
        """
        4. 本步骤可以省略，实现compute方法，这个方法主要用于`update`的加速，可以在这个方法中调用一些paddle实现好的Tensor计算API，编译到模型网络中一起使用低层C++ OP计算。
        """

        return 自己想要返回的数据，会做为update的参数传入。

    def update(self, ...):
        """
        5. 实现update方法，用于单个batch训练时进行评估指标计算。
        - 当`compute`类函数未实现时，会将模型的计算输出和标签数据的展平作为`update`的参数传入。
        - 当`compute`类函数做了实现时，会将compute的返回结果作为`update`的参数传入。
        """
        return acc value
    
    def accumulate(self):
        """
        6. 实现accumulate方法，返回历史batch训练积累后计算得到的评价指标值。
        每次`update`调用时进行数据积累，`accumulate`计算时对积累的所有数据进行计算并返回。
        结算结果会在`fit`接口的训练日志中呈现。
        """
        # 利用update中积累的成员变量数据进行计算后返回
        return accumulated acc value

    def reset(self):
        """
        7. 实现reset方法，每个Epoch结束后进行评估指标的重置，这样下个Epoch可以重新进行计算。
        """
        # do reset action

from paddle.metric import Metric


class Precision(Metric):
    """
    Precision (also called positive predictive value) is the fraction of
    relevant instances among the retrieved instances. Refer to
    https://en.wikipedia.org/wiki/Evaluation_of_binary_classifiers

    Noted that this class manages the precision score only for binary
    classification task.
    
    ......

    """

    def __init__(self, name='precision', *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.tp = 0  # true positive
        self.fp = 0  # false positive
        self._name = name

    def update(self, preds, labels):
        """
        Update the states based on the current mini-batch prediction results.

        Args:
            preds (numpy.ndarray): The prediction result, usually the output
                of two-class sigmoid function. It should be a vector (column
                vector or row vector) with data type: 'float64' or 'float32'.
            labels (numpy.ndarray): The ground truth (labels),
                the shape should keep the same as preds.
                The data type is 'int32' or 'int64'.
        """
        if isinstance(preds, paddle.Tensor):
            preds = preds.numpy()
        elif not _is_numpy_(preds):
            raise ValueError("The 'preds' must be a numpy ndarray or Tensor.")

        if isinstance(labels, paddle.Tensor):
            labels = labels.numpy()
        elif not _is_numpy_(labels):
            raise ValueError("The 'labels' must be a numpy ndarray or Tensor.")

        sample_num = labels.shape[0]
        preds = np.floor(preds + 0.5).astype("int32")

        for i in range(sample_num):
            pred = preds[i]
            label = labels[i]
            if pred == 1:
                if pred == label:
                    self.tp += 1
                else:
                    self.fp += 1

    def reset(self):
        """
        Resets all of the metric state.
        """
        self.tp = 0
        self.fp = 0

    def accumulate(self):
        """
        Calculate the final precision.

        Returns:
            A scaler float: results of the calculated precision.
        """
        ap = self.tp + self.fp
        return float(self.tp) / ap if ap != 0 else .0

    def name(self):
        """
        Returns metric name
        """
        return self._name

class SelfDefineCallback(paddle.callbacks.Callback):
    """
    1. 继承paddle.callbacks.Callback
    2. 按照自己的需求实现以下类成员方法：
        def on_train_begin(self, logs=None)                 训练开始前，`Model.fit`接口中调用
        def on_train_end(self, logs=None)                   训练结束后，`Model.fit`接口中调用
        def on_eval_begin(self, logs=None)                  评估开始前，`Model.evaluate`接口调用
        def on_eval_end(self, logs=None)                    评估结束后，`Model.evaluate`接口调用
        def on_test_begin(self, logs=None)                  预测测试开始前，`Model.predict`接口中调用
        def on_test_end(self, logs=None)                    预测测试结束后，`Model.predict`接口中调用 
        def on_epoch_begin(self, epoch, logs=None)          每轮训练开始前，`Model.fit`接口中调用 
        def on_epoch_end(self, epoch, logs=None)            每轮训练结束后，`Model.fit`接口中调用 
        def on_train_batch_begin(self, step, logs=None)     单个Batch训练开始前，`Model.fit`和`Model.train_batch`接口中调用
        def on_train_batch_end(self, step, logs=None)       单个Batch训练结束后，`Model.fit`和`Model.train_batch`接口中调用
        def on_eval_batch_begin(self, step, logs=None)      单个Batch评估开始前，`Model.evalute`和`Model.eval_batch`接口中调用
        def on_eval_batch_end(self, step, logs=None)        单个Batch评估结束后，`Model.evalute`和`Model.eval_batch`接口中调用
        def on_test_batch_begin(self, step, logs=None)      单个Batch预测测试开始前，`Model.predict`和`Model.test_batch`接口中调用
        def on_test_batch_end(self, step, logs=None)        单个Batch预测测试结束后，`Model.predict`和`Model.test_batch`接口中调用
    """
    def __init__(self):
        super().__init__()

    按照需求定义自己的类成员方法

class ModelCheckpoint(Callback):
    def __init__(self, save_freq=1, save_dir=None):
        self.save_freq = save_freq
        self.save_dir = save_dir

    def on_epoch_begin(self, epoch=None, logs=None):
        self.epoch = epoch

    def _is_save(self):
        return self.model and self.save_dir and ParallelEnv().local_rank == 0

    def on_epoch_end(self, epoch, logs=None):
        if self._is_save() and self.epoch % self.save_freq == 0:
            path = '{}/{}'.format(self.save_dir, epoch)
            print('save checkpoint at {}'.format(os.path.abspath(path)))
            self.model.save(path)

    def on_train_end(self, logs=None):
        if self._is_save():
            path = '{}/final'.format(self.save_dir)
            print('save checkpoint at {}'.format(os.path.abspath(path)))
            self.model.save(path)

class Mnist(paddle.nn.Layer):
    def __init__(self):
        super().__init__()

        self.flatten = paddle.nn.Flatten()
        self.linear_1 = paddle.nn.Linear(784, 512)
        self.linear_2 = paddle.nn.Linear(512, 10)
        self.relu = paddle.nn.ReLU()
        self.dropout = paddle.nn.Dropout(0.2)

    @paddle.jit.to_static
    def forward(self, inputs):
        y = self.flatten(inputs)
        y = self.linear_1(y)
        y = self.relu(y)
        y = self.dropout(y)
        y = self.linear_2(y)

        return y

model.save("~/model/mnist")