NLP--The Multilayer Perceptron（多层感知器）

目录

一、实验背景

二、The Multilayer Perceptron（多层感知器）

三、实验步骤

四、卷积神经网络

一、实验背景

1.1 实验背景

感知器是现存最简单的神经网络。感知器的一个历史性的缺点是它不能学习数据中存在的一些非常重要的模式。例如，查看图4-1中绘制的数据点。这相当于非此即彼(XOR)的情况，在这种情况下，决策边界不能是一条直线(也称为线性可分)。在这个例子中，感知器失败了。

在本实验中，多层感知器和卷积神经网络被分组在一起，因为它们都是前馈神经网络，并且与另一类神经网络——递归神经网络(RNNs)形成对比，递归神经网络(RNNs)允许反馈(或循环)，这样每次计算都可以从之前的计算中获得信息。在实验6和实验7中，我们将介绍RNNs以及为什么允许网络结构中的循环是有益的。

在我们介绍这些不同的模型时，需要理解事物如何工作的一个有用方法是在计算数据张量时注意它们的大小和形状。每种类型的神经网络层对它所计算的数据张量的大小和形状都有特定的影响，理解这种影响可以极大地有助于对这些模型的深入理解。

1.2 实验要点

• 通过“示例:带有多层感知器的姓氏分类”，掌握多层感知器在多层分类中的应用
• 掌握每种类型的神经网络层对它所计算的数据张量的大小和形状的影响

1.3 实验环境

• Python 3.6.7

1.4 附件目录

• 请将本实验所需数据文件**(surnames.csv)上传至目录：/data/surnames/**.
• 示例完整代码：
• exp4-In-Text-Examples.ipynb
• exp4-munging_surname_dataset.ipynb
• exp4-2D-Perceptron-MLP.ipynb
• exp4_4_Classify_Surnames_CNN.ipynb
• exp4_4_Classify_Surnames_MLP.ipynb

二、The Multilayer Perceptron（多层感知器）

多层感知器(MLP)被认为是最基本的神经网络构建模块之一。最简单的MLP是对第3章感知器的扩展。感知器将数据向量作为输入，计算出一个输出值。在MLP中，许多感知器被分组，以便单个层的输出是一个新的向量，而不是单个输出值。在PyTorch中，正如您稍后将看到的，这只需设置线性层中的输出特性的数量即可完成。MLP的另一个方面是，它将多个层与每个层之间的非线性结合在一起。

最简单的MLP，如图所示，由三个表示阶段和两个线性层组成。第一阶段是输入向量。第二个线性层计算一个输出向量。第三阶段是MPL的可视化展示

MLP除了简单的感知器之外，还有一个额外的计算层。我们用PyTorch的两个线性模块实例化了这个想法。线性对象被命名为fc1和fc2，它们遵循一个通用约定，即将线性模块称为“完全连接层”，简称为“fc层”。除了这两个线性层外，还有一个修正的线性单元(ReLU)非线性(在实验3“激活函数”一节中介绍)，它在被输入到第二个线性层之前应用于第一个线性层的输出。由于层的顺序性，必须确保层中的输出数量等于下一层的输入数量。使用两个线性层之间的非线性是必要的，因为没有它，两个线性层在数学上等价于一个线性层4，因此不能建模复杂的模式。MLP的实现只实现反向传播的前向传递。这是因为PyTorch根据模型的定义和向前传递的实现，自动计算出如何进行向后传递和梯度更新。

2.1 多层感知器

import torch.nn as nn
import torch.nn.functional as F

class MultilayerPerceptron(nn.Module):
    def __init__(self, input_dim, hidden_dim, output_dim):
        """
        Args:
            input_dim (int): the size of the input vectors
            hidden_dim (int): the output size of the first Linear layer
            output_dim (int): the output size of the second Linear layer
        """
        super(MultilayerPerceptron, self).__init__()
        self.fc1 = nn.Linear(input_dim, hidden_dim)
        self.fc2 = nn.Linear(hidden_dim, output_dim)

    def forward(self, x_in, apply_softmax=False):
        """The forward pass of the MLP

        Args:
            x_in (torch.Tensor): an input data tensor.
                x_in.shape should be (batch, input_dim)
            apply_softmax (bool): a flag for the softmax activation
                should be false if used with the Cross Entropy losses
        Returns:
            the resulting tensor. tensor.shape should be (batch, output_dim)
        """
        intermediate = F.relu(self.fc1(x_in))
        output = self.fc2(intermediate)

        if apply_softmax:
            output = F.softmax(output, dim=1)
        return output

2.2 MLP 实例化示例

batch_size = 2 # number of samples input at once
input_dim = 3
hidden_dim = 100
output_dim = 4

# Initialize model
mlp = MultilayerPerceptron(input_dim, hidden_dim, output_dim)
print(mlp)

MultilayerPerceptron(
(fc1): Linear(in_features=3, out_features=100, bias=True)
(fc2): Linear(in_features=100, out_features=4, bias=True)
)
 

2.3 具有随机输入的 MLP

import torch
def describe(x):
    print("Type: {}".format(x.type()))
    print("Shape/size: {}".format(x.shape))
    print("Values: \n{}".format(x))


x_input = torch.rand(batch_size, input_dim)
describe(x_input)

Type: torch.FloatTensor
Shape/size: torch.Size([2, 3])
Values:
tensor([[0.7883, 0.0888, 0.1100],
[0.0639, 0.9812, 0.7767]])
 

上述代码运行结果：

Type: torch.FloatTensor
Shape/size: torch.Size([2, 3])
Values: 
tensor([[0.6193, 0.7045, 0.7812],
        [0.6345, 0.4476, 0.9909]])

三、实验步骤

3.1 自定义的PyTorch数据集类 SurnameDataset，用于处理姓氏数据集

from torch.utils.data import Dataset
from argparse import Namespace

class SurnameDataset(Dataset):
    def __init__(self, surname_df, vectorizer):
        """
        Args:
            surname_df (pandas.DataFrame): the dataset
            vectorizer (SurnameVectorizer): vectorizer instatiated from dataset
        """
        self.surname_df = surname_df
        self._vectorizer = vectorizer

        self.train_df = self.surname_df[self.surname_df.split=='train']
        self.train_size = len(self.train_df)

        self.val_df = self.surname_df[self.surname_df.split=='val']
        self.validation_size = len(self.val_df)

        self.test_df = self.surname_df[self.surname_df.split=='test']
        self.test_size = len(self.test_df)

        self._lookup_dict = {'train': (self.train_df, self.train_size),
                             'val': (self.val_df, self.validation_size),
                             'test': (self.test_df, self.test_size)}

        self.set_split('train')
        
        # Class weights
        class_counts = surname_df.nationality.value_counts().to_dict()
        def sort_key(item):
            return self._vectorizer.nationality_vocab.lookup_token(item[0])
        sorted_counts = sorted(class_counts.items(), key=sort_key)
        frequencies = [count for _, count in sorted_counts]
        self.class_weights = 1.0 / torch.tensor(frequencies, dtype=torch.float32)

    @classmethod
    def load_dataset_and_make_vectorizer(cls, surname_csv):
        """Load dataset and make a new vectorizer from scratch
        
        Args:
            surname_csv (str): location of the dataset
        Returns:
            an instance of SurnameDataset
        """
        surname_df = pd.read_csv(surname_csv)
        train_surname_df = surname_df[surname_df.split=='train']
        return cls(surname_df, SurnameVectorizer.from_dataframe(train_surname_df))

    @classmethod
    def load_dataset_and_load_vectorizer(cls, surname_csv, vectorizer_filepath):
        """Load dataset and the corresponding vectorizer. 
        Used in the case in the vectorizer has been cached for re-use
        
        Args:
            surname_csv (str): location of the dataset
            vectorizer_filepath (str): location of the saved vectorizer
        Returns:
            an instance of SurnameDataset
        """
        surname_df = pd.read_csv(surname_csv)
        vectorizer = cls.load_vectorizer_only(vectorizer_filepath)
        return cls(surname_df, vectorizer)

    @staticmethod
    def load_vectorizer_only(vectorizer_filepath):
        """a static method for loading the vectorizer from file
        
        Args:
            vectorizer_filepath (str): the location of the serialized vectorizer
        Returns:
            an instance of SurnameVectorizer
        """
        with open(vectorizer_filepath) as fp:
            return SurnameVectorizer.from_serializable(json.load(fp))

    def save_vectorizer(self, vectorizer_filepath):
        """saves the vectorizer to disk using json
        
        Args:
            vectorizer_filepath (str): the location to save the vectorizer
        """
        with open(vectorizer_filepath, "w") as fp:
            json.dump(self._vectorizer.to_serializable(), fp)

    def get_vectorizer(self):
        """ returns the vectorizer """
        return self._vectorizer

    def set_split(self, split="train"):
        """ selects the splits in the dataset using a column in the dataframe """
        self._target_split = split
        self._target_df, self._target_size = self._lookup_dict[split]

    def __len__(self):
        return self._target_size

    def __getitem__(self, index):
        """the primary entry point method for PyTorch datasets
        
        Args:
            index (int): the index to the data point 
        Returns:
            a dictionary holding the data point's:
                features (x_surname)
                label (y_nationality)
        """
        row = self._target_df.iloc[index]

        surname_vector = \
            self._vectorizer.vectorize(row.surname)

        nationality_index = \
            self._vectorizer.nationality_vocab.lookup_token(row.nationality)

        return {'x_surname': surname_vector,
                'y_nationality': nationality_index}

    def get_num_batches(self, batch_size):
        """Given a batch size, return the number of batches in the dataset
        
        Args:
            batch_size (int)
        Returns:
            number of batches in the dataset
        """
        return len(self) // batch_size

    
def generate_batches(dataset, batch_size, shuffle=True,
                     drop_last=True, device="cpu"): 
    """
    A generator function which wraps the PyTorch DataLoader. It will 
      ensure each tensor is on the write device location.
    """
    dataloader = DataLoader(dataset=dataset, batch_size=batch_size,
                            shuffle=shuffle, drop_last=drop_last)

    for data_dict in dataloader:
        out_data_dict = {}
        for name, tensor in data_dict.items():
            out_data_dict[name] = data_dict[name].to(device)
        yield out_data_dict

3.2 构建多层感知机模型

class SurnameClassifier(nn.Module):
    """ A 2-layer Multilayer Perceptron for classifying surnames """
    def __init__(self, input_dim, hidden_dim, output_dim):
        """
        Args:
            input_dim (int): the size of the input vectors
            hidden_dim (int): the output size of the first Linear layer
            output_dim (int): the output size of the second Linear layer
        """
        super(SurnameClassifier, self).__init__()
        self.fc1 = nn.Linear(input_dim, hidden_dim)
        self.fc2 = nn.Linear(hidden_dim, output_dim)
 
    def forward(self, x_in, apply_softmax=False):
        """The forward pass of the classifier
        
        Args:
            x_in (torch.Tensor): an input data tensor. 
                x_in.shape should be (batch, input_dim)
            apply_softmax (bool): a flag for the softmax activation
                should be false if used with the Cross Entropy losses
        Returns:
            the resulting tensor. tensor.shape should be (batch, output_dim)
        """
        intermediate_vector = F.relu(self.fc1(x_in))
        prediction_vector = self.fc2(intermediate_vector)
 
        if apply_softmax:
            prediction_vector = F.softmax(prediction_vector, dim=1)
 
        return prediction_vector
 
def make_train_state(args):
    return {'stop_early': False,
            'early_stopping_step': 0,
            'early_stopping_best_val': 1e8,
            'learning_rate': args.learning_rate,
            'epoch_index': 0,
            'train_loss': [],
            'train_acc': [],
            'val_loss': [],
            'val_acc': [],
            'test_loss': -1,
            'test_acc': -1,
            'model_filename': args.model_state_file}
 
def update_train_state(args, model, train_state):
    """Handle the training state updates.
    Components:
     - Early Stopping: Prevent overfitting.
     - Model Checkpoint: Model is saved if the model is better
    :param args: main arguments
    :param model: model to train
    :param train_state: a dictionary representing the training state values
    :returns:
        a new train_state
    """
 
    # Save one model at least
    if train_state['epoch_index'] == 0:
        torch.save(model.state_dict(), train_state['model_filename'])
        train_state['stop_early'] = False
 
    # Save model if performance improved
    elif train_state['epoch_index'] >= 1:
        loss_tm1, loss_t = train_state['val_loss'][-2:]
 
        # If loss worsened
        if loss_t >= train_state['early_stopping_best_val']:
            # Update step
            train_state['early_stopping_step'] += 1
        # Loss decreased
        else:
            # Save the best model
            if loss_t < train_state['early_stopping_best_val']:
                torch.save(model.state_dict(), train_state['model_filename'])
 
            # Reset early stopping step
            train_state['early_stopping_step'] = 0
 
        # Stop early ?
        train_state['stop_early'] = \
            train_state['early_stopping_step'] >= args.early_stopping_criteria
 
    return train_state
 
def compute_accuracy(y_pred, y_target):
    _, y_pred_indices = y_pred.max(dim=1)
    n_correct = torch.eq(y_pred_indices, y_target).sum().item()
    return n_correct / len(y_pred_indices) * 100
 
def set_seed_everywhere(seed, cuda):
    np.random.seed(seed)
    torch.manual_seed(seed)
    if cuda:
        torch.cuda.manual_seed_all(seed)
 
def handle_dirs(dirpath):
    if not os.path.exists(dirpath):
        os.makedirs(dirpath)

3.3 训练模型

import numpy as np
from tqdm import tqdm_notebook

def make_train_state(args):
    return {'stop_early': False,
            'early_stopping_step': 0,
            'early_stopping_best_val': 1e8,
            'learning_rate': args.learning_rate,
            'epoch_index': 0,
            'train_loss': [],
            'train_acc': [],
            'val_loss': [],
            'val_acc': [],
            'test_loss': -1,
            'test_acc': -1,
            'model_filename': args.model_state_file}

def update_train_state(args, model, train_state):
    """Handle the training state updates.

    Components:
     - Early Stopping: Prevent overfitting.
     - Model Checkpoint: Model is saved if the model is better

    :param args: main arguments
    :param model: model to train
    :param train_state: a dictionary representing the training state values
    :returns:
        a new train_state
    """

    # Save one model at least
    if train_state['epoch_index'] == 0:
        torch.save(model.state_dict(), train_state['model_filename'])
        train_state['stop_early'] = False

    # Save model if performance improved
    elif train_state['epoch_index'] >= 1:
        loss_tm1, loss_t = train_state['val_loss'][-2:]

        # If loss worsened
        if loss_t >= train_state['early_stopping_best_val']:
            # Update step
            train_state['early_stopping_step'] += 1
        # Loss decreased
        else:
            # Save the best model
            if loss_t < train_state['early_stopping_best_val']:
                torch.save(model.state_dict(), train_state['model_filename'])

            # Reset early stopping step
            train_state['early_stopping_step'] = 0

        # Stop early ?
        train_state['stop_early'] = \
            train_state['early_stopping_step'] >= args.early_stopping_criteria

    return train_state

def compute_accuracy(y_pred, y_target):
    _, y_pred_indices = y_pred.max(dim=1)
    n_correct = torch.eq(y_pred_indices, y_target).sum().item()
    return n_correct / len(y_pred_indices) * 100

def set_seed_everywhere(seed, cuda):
    np.random.seed(seed)
    torch.manual_seed(seed)
    if cuda:
        torch.cuda.manual_seed_all(seed)

def handle_dirs(dirpath):
    if not os.path.exists(dirpath):
        os.makedirs(dirpath)

classifier = classifier.to(args.device)
dataset.class_weights = dataset.class_weights.to(args.device)

    
loss_func = nn.CrossEntropyLoss(dataset.class_weights)
optimizer = optim.Adam(classifier.parameters(), lr=args.learning_rate)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer=optimizer,
                                                 mode='min', factor=0.5,
                                                 patience=1)

train_state = make_train_state(args)

epoch_bar = tqdm_notebook(desc='training routine', 
                          total=args.num_epochs,
                          position=0)

dataset.set_split('train')
train_bar = tqdm_notebook(desc='split=train',
                          total=dataset.get_num_batches(args.batch_size), 
                          position=1, 
                          leave=True)
dataset.set_split('val')
val_bar = tqdm_notebook(desc='split=val',
                        total=dataset.get_num_batches(args.batch_size), 
                        position=1, 
                        leave=True)

try:
    for epoch_index in range(args.num_epochs):
        train_state['epoch_index'] = epoch_index

        # Iterate over training dataset

        # setup: batch generator, set loss and acc to 0, set train mode on

        dataset.set_split('train')
        batch_generator = generate_batches(dataset, 
                                           batch_size=args.batch_size, 
                                           device=args.device)
        running_loss = 0.0
        running_acc = 0.0
        classifier.train()

        for batch_index, batch_dict in enumerate(batch_generator):
            # the training routine is these 5 steps:

            # --------------------------------------
            # step 1. zero the gradients
            optimizer.zero_grad()

            # step 2. compute the output
            y_pred = classifier(batch_dict['x_surname'])

            # step 3. compute the loss
            loss = loss_func(y_pred, batch_dict['y_nationality'])
            loss_t = loss.item()
            running_loss += (loss_t - running_loss) / (batch_index + 1)

            # step 4. use loss to produce gradients
            loss.backward()

            # step 5. use optimizer to take gradient step
            optimizer.step()
            # -----------------------------------------
            # compute the accuracy
            acc_t = compute_accuracy(y_pred, batch_dict['y_nationality'])
            running_acc += (acc_t - running_acc) / (batch_index + 1)

            # update bar
            train_bar.set_postfix(loss=running_loss, acc=running_acc, 
                            epoch=epoch_index)
            train_bar.update()

        train_state['train_loss'].append(running_loss)
        train_state['train_acc'].append(running_acc)

        # Iterate over val dataset

        # setup: batch generator, set loss and acc to 0; set eval mode on
        dataset.set_split('val')
        batch_generator = generate_batches(dataset, 
                                           batch_size=args.batch_size, 
                                           device=args.device)
        running_loss = 0.
        running_acc = 0.
        classifier.eval()

        for batch_index, batch_dict in enumerate(batch_generator):

            # compute the output
            y_pred =  classifier(batch_dict['x_surname'])

            # step 3. compute the loss
            loss = loss_func(y_pred, batch_dict['y_nationality'])
            loss_t = loss.to("cpu").item()
            running_loss += (loss_t - running_loss) / (batch_index + 1)

            # compute the accuracy
            acc_t = compute_accuracy(y_pred, batch_dict['y_nationality'])
            running_acc += (acc_t - running_acc) / (batch_index + 1)
            val_bar.set_postfix(loss=running_loss, acc=running_acc, 
                            epoch=epoch_index)
            val_bar.update()

        train_state['val_loss'].append(running_loss)
        train_state['val_acc'].append(running_acc)

        train_state = update_train_state(args=args, model=classifier,
                                         train_state=train_state)

        scheduler.step(train_state['val_loss'][-1])

        if train_state['stop_early']:
            break

        train_bar.n = 0
        val_bar.n = 0
        epoch_bar.update()
except KeyboardInterrupt:
    print("Exiting loop")

3.4 用于执行国籍预测的函数

def predict_nationality(name, classifier, vectorizer):
    vectorized_name = vectorizer.vectorize(name)
    vectorized_name = torch.tensor(vectorized_name).view(1, -1)
    result = classifier(vectorized_name, apply_softmax=True)

    probability_values, indices = result.max(dim=1)
    index = indices.item()

    predicted_nationality = vectorizer.nationality_vocab.lookup_index(index)
    probability_value = probability_values.item()

    return {'nationality': predicted_nationality,
            'probability': probability_value}

四、卷积神经网络

卷积神经网络(CNN)，这是一种非常适合检测空间子结构(并因此创建有意义的空间子结构)的神经网络。CNNs通过使用少量的权重来扫描输入数据张量来实现这一点。通过这种扫描，它们产生表示子结构检测(或不检测)的输出张量。

CNNs的名称和基本功能源于经典的数学运算卷积。卷积已经应用于各种工程学科，包括数字信号处理和计算机图形学。一般来说，卷积使用程序员指定的参数。这些参数被指定来匹配一些功能设计，如突出边缘或抑制高频声音。事实上，许多Photoshop滤镜都是应用于图像的固定卷积运算。然而，在深度学习和本实验中，我们从数据中学习卷积滤波器的参数，因此它对于解决当前的任务是最优的。

4.1 姓氏矢量化

class SurnameVectorizer(object):
    """ The Vectorizer which coordinates the Vocabularies and puts them to use"""
    def vectorize(self, surname):
        """
        Args:
            surname (str): the surname
        Returns:
            one_hot_matrix (np.ndarray): a matrix of one-hot vectors
        """

        one_hot_matrix_size = (len(self.character_vocab), self.max_surname_length)
        one_hot_matrix = np.zeros(one_hot_matrix_size, dtype=np.float32)

        for position_index, character in enumerate(surname):
            character_index = self.character_vocab.lookup_token(character)
            one_hot_matrix[character_index][position_index] = 1

        return one_hot_matrix

    @classmethod
    def from_dataframe(cls, surname_df):
        """Instantiate the vectorizer from the dataset dataframe

        Args:
            surname_df (pandas.DataFrame): the surnames dataset
        Returns:
            an instance of the SurnameVectorizer
        """
        character_vocab = Vocabulary(unk_token="@")
        nationality_vocab = Vocabulary(add_unk=False)
        max_surname_length = 0

        for index, row in surname_df.iterrows():
            max_surname_length = max(max_surname_length, len(row.surname))
            for letter in row.surname:
                character_vocab.add_token(letter)
            nationality_vocab.add_token(row.nationality)

        return cls(character_vocab, nationality_vocab, max_surname_length)

4.2 基于 CNN 的 SurnameClassifier

import torch.nn as nn
import torch.nn.functional as F

class SurnameClassifier(nn.Module):
    def __init__(self, initial_num_channels, num_classes, num_channels):
        """
        Args:
            initial_num_channels (int): size of the incoming feature vector
            num_classes (int): size of the output prediction vector
            num_channels (int): constant channel size to use throughout network
        """
        super(SurnameClassifier, self).__init__()

        self.convnet = nn.Sequential(
            nn.Conv1d(in_channels=initial_num_channels,
                      out_channels=num_channels, kernel_size=3),
            nn.ELU(),
            nn.Conv1d(in_channels=num_channels, out_channels=num_channels,
                      kernel_size=3, stride=2),
            nn.ELU(),
            nn.Conv1d(in_channels=num_channels, out_channels=num_channels,
                      kernel_size=3, stride=2),
            nn.ELU(),
            nn.Conv1d(in_channels=num_channels, out_channels=num_channels,
                      kernel_size=3),
            nn.ELU()
        )
        self.fc = nn.Linear(num_channels, num_classes)

    def forward(self, x_surname, apply_softmax=False):
        """The forward pass of the classifier

        Args:
            x_surname (torch.Tensor): an input data tensor.
                x_surname.shape should be (batch, initial_num_channels,
                                           max_surname_length)
            apply_softmax (bool): a flag for the softmax activation
                should be false if used with the Cross Entropy losses
        Returns:
            the resulting tensor. tensor.shape should be (batch, num_classes)
        """
        features = self.convnet(x_surname).squeeze(dim=2)
        prediction_vector = self.fc(features)

        if apply_softmax:
            prediction_vector = F.softmax(prediction_vector, dim=1)

        return prediction_vector