深度学总结：skip-gram pytorch实现

skip-gram pytorch 朴素实现

网络结构

class SkipGram(nn.Module):
    def __init__(self, n_vocab, n_embed):
        super().__init__()
        
        self.embed = nn.Embedding(n_vocab, n_embed)
        self.output = nn.Linear(n_embed, n_vocab)
        self.log_softmax = nn.LogSoftmax(dim=1)
    
    def forward(self, x):
        x = self.embed(x)
        scores = self.output(x)
        log_ps = self.log_softmax(scores)
        
        return log_ps

训练过程:使用nn.NLLLoss()

# check if GPU is available
device = 'cuda' if torch.cuda.is_available() else 'cpu'

embedding_dim=300 # you can change, if you want

model = SkipGram(len(vocab_to_int), embedding_dim).to(device)
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.parameters(), lr=0.003)

print_every = 500
steps = 0
epochs = 5

# train for some number of epochs
for e in range(epochs):
    
    # get input and target batches
    for inputs, targets in get_batches(train_words, 512):
        steps += 1
        inputs, targets = torch.LongTensor(inputs), torch.LongTensor(targets)
        inputs, targets = inputs.to(device), targets.to(device)
        
        log_ps = model(inputs)
        loss = criterion(log_ps, targets)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

batch的准备，为unsupervised，准备数据获取（center,contex)的pair：

def get_target(words, idx, window_size=5):
    ''' Get a list of words in a window around an index. '''
    
    R = np.random.randint(1, window_size+1)
    start = idx - R if (idx - R) > 0 else 0
    stop = idx + R
    target_words = words[start:idx] + words[idx+1:stop+1]
    
    return list(target_words)
def get_batches(words, batch_size, window_size=5):
    ''' Create a generator of word batches as a tuple (inputs, targets) '''
    
    n_batches = len(words)//batch_size
    
    # only full batches
    words = words[:n_batches*batch_size]
    
    for idx in range(0, len(words), batch_size):
        x, y = [], []
        batch = words[idx:idx+batch_size]
        for ii in range(len(batch)):
            batch_x = batch[ii]
            batch_y = get_target(batch, ii, window_size)
            y.extend(batch_y)
            x.extend([batch_x]*len(batch_y))
        yield x, y

采样时的优化：Subsampling降低高频词的概率

Words that show up often such as “the”, “of”, and “for” don’t provide much context to the nearby words. If we discard some of them, we can remove some of the noise from our data and in return get faster training and better representations. This process is called subsampling by Mikolov. For each word $w_i$ in the training set, we’ll discard it with probability given by

$$P(w_i)=1-\sqrt{\frac{t}{f(w_i)}}$$
where $t$ is a threshold parameter and $f(w_i)$ is the frequency of word $w_i$ in the total dataset.

from collections import Counter
import random
import numpy as np

threshold = 1e-5
word_counts = Counter(int_words)
#print(list(word_counts.items())[0])  # dictionary of int_words, how many times they appear

total_count = len(int_words)
freqs = {word: count/total_count for word, count in word_counts.items()}
p_drop = {word: 1 - np.sqrt(threshold/freqs[word]) for word in word_counts}
# discard some frequent words, according to the subsampling equation
# create a new list of words for training
train_words = [word for word in int_words if random.random() < (1 - p_drop[word])]

skip-gram 进阶：negative sampling

一般都是针对计算效率优化的方法：negative sampling和hierachical softmax

negative sampling实现：

negative sampling原理：

class NegativeSamplingLoss(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, input_vectors, output_vectors, noise_vectors):
        
        batch_size, embed_size = input_vectors.shape
        
        # Input vectors should be a batch of column vectors
        input_vectors = input_vectors.view(batch_size, embed_size, 1)
        
        # Output vectors should be a batch of row vectors
        output_vectors = output_vectors.view(batch_size, 1, embed_size)
        
        # bmm = batch matrix multiplication
        # correct log-sigmoid loss
        out_loss = torch.bmm(output_vectors, input_vectors).sigmoid().log()
        out_loss = out_loss.squeeze()
        
        # incorrect log-sigmoid loss
        noise_loss = torch.bmm(noise_vectors.neg(), input_vectors).sigmoid().log()
        noise_loss = noise_loss.squeeze().sum(1)  # sum the losses over the sample of noise vectors

        # negate and sum correct and noisy log-sigmoid losses
        # return average batch loss
        return -(out_loss + noise_loss).mean()

negative sampling抽样方法：

# Get our noise distribution
# Using word frequencies calculated earlier in the notebook
word_freqs = np.array(sorted(freqs.values(), reverse=True))
unigram_dist = word_freqs/word_freqs.sum()
noise_dist = torch.from_numpy(unigram_dist**(0.75)/np.sum(unigram_dist**(0.75)))

negative sampling前向传递过程：

class SkipGramNeg(nn.Module):
    def __init__(self, n_vocab, n_embed, noise_dist=None):
        super().__init__()
        
        self.n_vocab = n_vocab
        self.n_embed = n_embed
        self.noise_dist = noise_dist
        
        # define embedding layers for input and output words
        self.in_embed = nn.Embedding(n_vocab, n_embed)
        self.out_embed = nn.Embedding(n_vocab, n_embed)
        
        # Initialize embedding tables with uniform distribution
        # I believe this helps with convergence
        self.in_embed.weight.data.uniform_(-1, 1)
        self.out_embed.weight.data.uniform_(-1, 1)
        
    def forward_input(self, input_words):
        input_vectors = self.in_embed(input_words)
        return input_vectors
    
    def forward_output(self, output_words):
        output_vectors = self.out_embed(output_words)
        return output_vectors
    
    def forward_noise(self, batch_size, n_samples):
        """ Generate noise vectors with shape (batch_size, n_samples, n_embed)"""
        if self.noise_dist is None:
            # Sample words uniformly
            noise_dist = torch.ones(self.n_vocab)
        else:
            noise_dist = self.noise_dist
            
        # Sample words from our noise distribution
        noise_words = torch.multinomial(noise_dist,
                                        batch_size * n_samples,
                                        replacement=True)
        
        device = "cuda" if model.out_embed.weight.is_cuda else "cpu"
        noise_words = noise_words.to(device)
        
        noise_vectors = self.out_embed(noise_words).view(batch_size, n_samples, self.n_embed)
        
        return noise_vectors

negative sampling训练过程：

device = 'cuda' if torch.cuda.is_available() else 'cpu'

# Get our noise distribution
# Using word frequencies calculated earlier in the notebook
word_freqs = np.array(sorted(freqs.values(), reverse=True))
unigram_dist = word_freqs/word_freqs.sum()
noise_dist = torch.from_numpy(unigram_dist**(0.75)/np.sum(unigram_dist**(0.75)))

# instantiating the model
embedding_dim = 300
model = SkipGramNeg(len(vocab_to_int), embedding_dim, noise_dist=noise_dist).to(device)

# using the loss that we defined
criterion = NegativeSamplingLoss() 
optimizer = optim.Adam(model.parameters(), lr=0.003)

print_every = 1500
steps = 0
epochs = 5

# train for some number of epochs
for e in range(epochs):
    
    # get our input, target batches
    for input_words, target_words in get_batches(train_words, 512):
        steps += 1
        inputs, targets = torch.LongTensor(input_words), torch.LongTensor(target_words)
        inputs, targets = inputs.to(device), targets.to(device)
        
        # input, output, and noise vectors
        input_vectors = model.forward_input(inputs)
        output_vectors = model.forward_output(targets)
        noise_vectors = model.forward_noise(inputs.shape[0], 5)

        # negative sampling loss
        loss = criterion(input_vectors, output_vectors, noise_vectors)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()