深度学总结:skip-gram pytorch实现
文章目录
skip-gram pytorch 朴素实现
网络结构
训练过程:使用nn.NLLLoss()
batch的准备,为unsupervised,准备数据获取(center,contex)的pair:
采样时的优化:Subsampling降低高频词的概率
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:
negative sampling抽样方法:
negative sampling前向传递过程:
negative sampling训练过程:
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
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
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()
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
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
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
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 wi w_iw
i
in the training set, we’ll discard it with probability given by
P(wi)=1−tf(wi)−−−−√ P(w_i) = 1 - \sqrt{\frac{t}{f(w_i)}}
P(w
i
)=1−
f(w
i
)
t
where t tt is a threshold parameter and f(wi) f(w_i)f(w
i
) is the frequency of word wi w_iw
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])]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
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()
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
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)))
1
2
3
4
5
6
7
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
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
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()
深度学总结:skip-gram pytorch实现的更多相关文章
- pytorch深度学习书、论坛和比赛地址
pytorch深度学习书.论坛和比赛地址 待办 https://zhuanlan.zhihu.com/p/85353963 http://zh.d2l.ai/ https://discuss.gluo ...
- 深度学习识别CIFAR10:pytorch训练LeNet、AlexNet、VGG19实现及比较(三)
版权声明:本文为博主原创文章,欢迎转载,并请注明出处.联系方式:460356155@qq.com VGGNet在2014年ImageNet图像分类任务竞赛中有出色的表现.网络结构如下图所示: 同样的, ...
- 深度学习识别CIFAR10:pytorch训练LeNet、AlexNet、VGG19实现及比较(二)
版权声明:本文为博主原创文章,欢迎转载,并请注明出处.联系方式:460356155@qq.com AlexNet在2012年ImageNet图像分类任务竞赛中获得冠军.网络结构如下图所示: 对CIFA ...
- 深度学习框架Keras与Pytorch对比
对于许多科学家.工程师和开发人员来说,TensorFlow是他们的第一个深度学习框架.TensorFlow 1.0于2017年2月发布,可以说,它对用户不太友好. 在过去的几年里,两个主要的深度学习库 ...
- 深度学*点云语义分割:CVPR2019论文阅读
深度学*点云语义分割:CVPR2019论文阅读 Point Cloud Oversegmentation with Graph-Structured Deep Metric Learning 摘要 本 ...
- 3D点云深度学*
3D点云深度学* 在自动驾驶中关于三维点云的深度学*方法应用.三维场景语义理解的方法以及对应的关键技术介绍. 1. 数据 但是对于3D点云,数据正在迅速增长.大有从2D向3D发展的趋势,比如在open ...
- 深度学习调用TensorFlow、PyTorch等框架
深度学习调用TensorFlow.PyTorch等框架 一.开发目标目标 提供统一接口的库,它可以从C++和Python中的多个框架中运行深度学习模型.欧米诺使研究人员能够在自己选择的框架内轻松建立模 ...
- 动手学深度学习9-多层感知机pytorch
多层感知机 隐藏层 激活函数 小结 多层感知机 之前已经介绍过了线性回归和softmax回归在内的单层神经网络,然后深度学习主要学习多层模型,后续将以多层感知机(multilayer percetro ...
- 【深度学习 01】线性回归+PyTorch实现
1. 线性回归 1.1 线性模型 当输入包含d个特征,预测结果表示为: 记x为样本的特征向量,w为权重向量,上式可表示为: 对于含有n个样本的数据集,可用X来表示n个样本的特征集合,其中行代表样本,列 ...
随机推荐
- 数据库MySQL技术-基础知识
数据库技术: SQL,关系数据库标准 注意: 环境编码: cmd客户端是固定的gbk编码 而php网页中,是该网页文件的编码(现在主流都是utf8). mysql> set names gb ...
- CodeForces 719A Vitya in the Countryside (水题)
题意:根据题目,给定一些数字,让你判断是上升还是下降. 析:注意只有0,15时特别注意一下,然后就是14 15 1 0注意一下就可以了. 代码如下: #pragma comment(linker, & ...
- Codeforces Round #422 (Div. 2) C. Hacker, pack your bags!(更新数组)
传送门 题意 给出n个区间[l,r]及花费\(cost_i\),找两个区间满足 1.区间和为指定值x 2.花费最小 分析 先用vector记录(l,r,cost)和(r,l,cost),按l排序,再设 ...
- Codeforces703B Mishka and trip
题意: 就是有n个点,本来相邻点之间就有一条边,1和n之间也有一条,然后给你几个特殊点,说这些特殊点和其他所有点都连起来了,然后算一个所有边的权值和,每条边的权值等于两个点的c相乘. 思路: 水题啊- ...
- js的NaN变量
js中,我们经常在parseInt函数的时候遇到NaN变量,这个变量到底是什么呢? w3c上这样解释: NaN 属性是代表非数字值的特殊值.该属性用于指示某个值不是数字.可以把 Number 对象设置 ...
- C++ C# 默认对齐是8字节
C++ C# 默认对齐是8字节 以上,一直以为是4字节,尼玛
- js对象注册清除
<!DOCTYPE html><html> <head> <meta charset="UTF-8"> <title>& ...
- Unity插值函数Lerp()与增量时间Time.deltatime
一.Unity插值函数Lerp() 通过官方文档简单了解插值函数(https://docs.unity3d.com/ScriptReference/index.html),可以看到插值函数有很多 Ma ...
- easyUI Uncaught TypeError: Cannot read property 'length' of undefined
dataGrid json 封装数据格式为 List<Object> 格式
- Caffe实战二(手写体识别例程:CPU、GPU、cuDNN速度对比)
上一篇文章成功在CPU模式下编译了Caffe,接下来需要运行一个例程来直观的了解Caffe的作用.(参考:<深度学习 21天实战Caffe>第6天 运行手写体数字识别例程) 编译步骤: C ...