word2vec的高速化：CBOW模型

word2vec的高速化

1.CBOW模型的学习代码

import sys
sys.path.append('..')
from common import config
# 在用GPU运行时，请打开下面的注释（需要cupy）
# ===============================================
#config.GPU = True
# ===============================================
from common.np import *
import pickle
from common.trainer import Trainer
from common.optimizer import Adam
from cbow import CBOW
from skip_gram import SkipGram
from common.util import create_contexts_target, to_cpu, to_gpu
from dataset import ptb


# 设定超参数
window_size = 5
hidden_size = 100
batch_size = 100
max_epoch = 10

# 读入数据
corpus, word_to_id, id_to_word = ptb.load_data('train')		#加载熟练数据
vocab_size = len(word_to_id)

contexts, target = create_contexts_target(corpus, window_size)		#
if config.GPU:
    contexts, target = to_gpu(contexts), to_gpu(target)

# 生成模型等
model = CBOW(vocab_size, hidden_size, window_size, corpus)
# model = SkipGram(vocab_size, hidden_size, window_size, corpus)
optimizer = Adam()
trainer = Trainer(model, optimizer)

# 开始学习
trainer.fit(contexts, target, max_epoch, batch_size)
trainer.plot()

# 保存必要数据，以便后续使用
word_vecs = model.word_vecs
if config.GPU:
    word_vecs = to_cpu(word_vecs)
params = {}
params['word_vecs'] = word_vecs.astype(np.float16)
params['word_to_id'] = word_to_id
params['id_to_word'] = id_to_word
pkl_file = 'cbow_params.pkl'  # or 'skipgram_params.pkl'
with open(pkl_file, 'wb') as f:
    pickle.dump(params, f, -1)

2. create_contexts_target

def create_contexts_target(corpus, window_size=1):
    '''生成上下文和目标词

    :param corpus: 语料库（单词ID列表）
    :param window_size: 窗口大小（当窗口大小为1时，左右各1个单词为上下文）
    :return:
    '''
    target = corpus[window_size:-window_size]
    contexts = []

    for idx in range(window_size, len(corpus)-window_size):
        cs = []
        for t in range(-window_size, window_size + 1):
            if t == 0:
                continue
            cs.append(corpus[idx + t])
        contexts.append(cs)

    return np.array(contexts), np.array(target)

3.CBOW模型的实现

import sys
sys.path.append('..')
from common.np import *  # import numpy as np
from common.layers import Embedding
from ch04.negative_sampling_layer import NegativeSamplingLoss


class CBOW:
    def __init__(self, vocab_size, hidden_size, window_size, corpus):
        V, H = vocab_size, hidden_size

        # 初始化权重
        W_in = 0.01 * np.random.randn(V, H).astype('f')
        W_out = 0.01 * np.random.randn(V, H).astype('f')

        # 生成层
        self.in_layers = []
        for i in range(2 * window_size):		#创建 2 * window_size 个Embedding 层
            layer = Embedding(W_in)  # 使用Embedding层，输入层-隐藏层
            self.in_layers.append(layer)
        self.ns_loss = NegativeSamplingLoss(W_out, corpus, power=0.75, sample_size=5)

        # 将所有的权重和梯度整理到列表中
        layers = self.in_layers + [self.ns_loss]
        self.params, self.grads = [], []
        for layer in layers:
            self.params += layer.params
            self.grads += layer.grads

        # 将单词的分布式表示设置为成员变量
        self.word_vecs = W_in
	#contexts, target如图所示
    def forward(self, contexts, target):
        h = 0
        for i, layer in enumerate(self.in_layers):
            h += layer.forward(contexts[:, i])
        h *= 1 / len(self.in_layers)
        loss = self.ns_loss.forward(h, target)		#实现负采样，NegativeSamplingLoss，并返回损失函数
        return loss

    def backward(self, dout=1):
        dout = self.ns_loss.backward(dout)
        dout *= 1 / len(self.in_layers)
        for layer in self.in_layers:
            layer.backward(dout)
        return None

forward：contexts是目标词的上下的关联词，target使我们要预测的值，这里我们给出的正确解

​ 首先：先调用输入层Embedding, 相关词先经过Embedding，然后相加，再除以相关词的总数这就求出H层隐藏层

​ 然后：H层和target目标词 调用NegativeSamplingLoss，求出损失值

​ 最后：给出损失值。

2.1使用Embedding层（输入层到隐藏层）

class Embedding:
    def __init__(self, W):
        self.params = [W]
        self.grads = [np.zeros_like(W)]
        self.idx = None
        
    def forward(self, idx):
        W, = self.params   #加逗号把array外面的方括号去点了
        self.idx = idx
        out = W[idx]
        return out
    
    def backward(self, dout):
        dW, = self.grads
        dW[...] = 0
        for i, word_id in enumerate(self.idx):
            dW[word_id] += dout[i]
         # 或者
         # np.add.at(dW, self.idx, dout)
        return None

Embedding层是获取的：输入层到隐藏层的，输入的值第几个直接取到Win的第几个。

2.2 实现负采样，NegativeSamplingLoss

import sys
sys.path.append('..')
from common.np import *  # import numpy as np
from common.layers import Embedding, SigmoidWithLoss
import collections

class NegativeSamplingLoss:
    def __init__(self, W, corpus, power=0.75, sample_size=5):
        #输出侧权重的 W、语料库（单词 ID 列表）corpus、概率分布的次方值 power 和负例的采样数 sample_size。
        self.sample_size = sample_size
        self.sampler = UnigramSampler(corpus, power, sample_size)			#负样例类
         #成员变量 loss_layers 和 embed_dot_layers 中以列表格式保存了必要的层。
        #在这两个列表中生成 sample_size + 1 个层，这是因为需要生成一个正例用的层和 sample_size 个负例用的层。
        self.loss_layers = [SigmoidWithLoss() for _ in range(sample_size + 1)]	#损失函数
        self.embed_dot_layers = [EmbeddingDot(W) for _ in range(sample_size + 1)]	#这是w是Wout

        self.params, self.grads = [], []
        for layer in self.embed_dot_layers:
            self.params += layer.params
            self.grads += layer.grads
	
    #H是隐藏层，target是目标词的正确解
    def forward(self, h, target):
        batch_size = target.shape[0]		#代表一共有多少个样例
        negative_sample = self.sampler.get_negative_sample(target)  #负样例采集，返回收集的负样例

        # 正例的正向传播，
        score = self.embed_dot_layers[0].forward(h, target)
        correct_label = np.ones(batch_size, dtype=np.int32)		#正确解标签都是1
        loss = self.loss_layers[0].forward(score, correct_label)	#求出损失值

        # 负例的正向传播
        negative_label = np.zeros(batch_size, dtype=np.int32)
        for i in range(self.sample_size):
            negative_target = negative_sample[:, i]
            score = self.embed_dot_layers[1 + i].forward(h, negative_target)
            loss += self.loss_layers[1 + i].forward(score, negative_label)

        return loss

    def backward(self, dout=1):
        dh = 0
        for l0, l1 in zip(self.loss_layers, self.embed_dot_layers):
            dscore = l0.backward(dout)
            dh += l1.backward(dscore)

        return dh

负采样的采样方法:UnigramSampler

import sys
sys.path.append('..')
from common.np import *  # import numpy as np
from common.layers import Embedding, SigmoidWithLoss
import collections


class UnigramSampler:
    def __init__(self, corpus, power, sample_size):
        self.sample_size = sample_size
        self.vocab_size = None				￥#一共有多少个单词，(重复的算成一个)
        self.word_p = None					#负样例

        counts = collections.Counter()
        for word_id in corpus:		#统计每个单词出现的次数
            counts[word_id] += 1

        vocab_size = len(counts)		#代表一共有多少个单词，(重复的算成一个)
        self.vocab_size = vocab_size

        self.word_p = np.zeros(vocab_size)
        for i in range(vocab_size):
            self.word_p[i] = counts[i]

        self.word_p = np.power(self.word_p, power)
        self.word_p /= np.sum(self.word_p)			#这里是单词出现的频率大小

    def get_negative_sample(self, target):
        batch_size = target.shape[0]		#共有多少个样例

        if not GPU:
            negative_sample = np.zeros((batch_size, self.sample_size), dtype=np.int32)

            for i in range(batch_size):
                p = self.word_p.copy()		#这里是单词出现的频率
                target_idx = target[i]		#这里是取出正例的索引
                p[target_idx] = 0	#把正例概率设为0
                p /= p.sum()
                negative_sample[i, :] = np.random.choice(self.vocab_size, size=self.sample_size, replace=False, p=p)
        else:
            # 在用GPU(cupy）计算时，优先速度
            # 有时目标词存在于负例中
            negative_sample = np.random.choice(self.vocab_size, size=(batch_size, self.sample_size),
                                               replace=True, p=self.word_p)

        return negative_sample

（详细过程见https://blog.csdn.net/weixin_44953928/article/details/121864848?spm=1001.2014.3001.5501 4.2.6  负采样的采样方法中）

Sigmoid求损失函数：SigmoidWithLoss

class SigmoidWithLoss:
    def __init__(self):
        self.params, self.grads = [], []
        self.loss = None
        self.y = None  # sigmoid的输出
        self.t = None  # 监督标签

    def forward(self, x, t):
        self.t = t
        self.y = 1 / (1 + np.exp(-x))

        self.loss = cross_entropy_error(np.c_[1 - self.y, self.y], self.t)

        return self.loss

    def backward(self, dout=1):
        batch_size = self.t.shape[0]

        dx = (self.y - self.t) * dout / batch_size
        return dx

EmbeddingDot（隐藏层到输出层）

先调用Embedding 返回的是 Wout 对应的某一行，然后h再于Embedding相乘

class EmbeddingDot:
    def __init__(self, W):
        self.embed = Embedding(W)
        self.params = self.embed.params
        self.grads = self.embed.grads
        self.cache = None

    def forward(self, h, idx):
        target_W = self.embed.forward(idx)
        out = np.sum(target_W * h, axis=1)  #返回的矩阵左右，一行相加

        self.cache = (h, target_W)
        return out

    def backward(self, dout):
        h, target_W = self.cache
        dout = dout.reshape(dout.shape[0], 1)

        dtarget_W = dout * h
        self.embed.backward(dtarget_W)
        dh = dout * target_W
        return dh

Embedding返回的知识 Wout 对应的某一行

#这里EmbeddingDot调用Embedding，这里的w是Wout,输出侧的w
class Embedding:
    def __init__(self, W):
        self.params = [W]
        self.grads = [np.zeros_like(W)]
        self.idx = None
        
    def forward(self, idx):
        W, = self.params   #加逗号把array外面的方括号去点了
        self.idx = idx
        out = W[idx]
        return out
    
    def backward(self, dout):
        dW, = self.grads
        dW[...] = 0
        for i, word_id in enumerate(self.idx):
            dW[word_id] += dout[i]
         # 或者
         # np.add.at(dW, self.idx, dout)
        return None