NLP 学习笔记 1:pytorch基础操作以及Perceptron 和 FF networks实现

2023-10-25 12:20

本文主要是介绍NLP 学习笔记 1:pytorch基础操作以及Perceptron 和 FF networks实现,希望对大家解决编程问题提供一定的参考价值,需要的开发者们随着小编来一起学习吧!

一些自己的nlp学习笔记

一:一些基础的pytorch操作

1 tensor的建立

import torch
import numpy as np
x = torch.Tensor(2,3) # 建立两行三列的torch tensorprint(x.type())       # type是Tensor类的一个mothod,返回Python string# torch.FloatTensor是real number的默认类型,一般来说GPU都能很好的处理x = torch.rand(2,3)   # uniform distribution
x = torch.randn(2,3)  # normal distributionx = torch.zeros(2, 3) # 全0 tensor
x = torch.ones(2, 3)  # 全1 tensor
x.fill_(5)            # 将tensor中全填入某相同的值#Tensor from list
x = torch.Tensor([[1, 2, 3], [4, 5, 6]])    #从list中获取tensor#From numpy to torch
a = np.random.rand(2, 3)                          
x = torch.from_numpy(a)                          # 用from_numpy将numpy类型转为tensor
x = torch.from_numpy(a).type(torch.FloatTensor)  # 可以用type来指定数据类型
y = torch.from_numpy(a).type_as(x)               # 可以用type_as来指定与其他tensor相同的数 # 据类型#数据类型以及数据类型的转换,一般默认为FloatTensor
z = x.long()                                     # 转为long

 2 tensor基础操作

# 求和
print(torch.add(x,x))
print(torch.sum(x, dim=0))          #按列求和# 对应元素求积
print(torch.mul(x,x))
print(x*x)# range tensor
print(torch.arange(6))# 返回不同shape的tensor
print(x.view(3, 2))x1 = torch.arange(6).view(2,3)# indexing + sum
x2 = torch.ones(3, 2).long()
x2[:, 1] += 1print('x1 =', x1)
print('x2 =', x2)# 矩阵乘
print(torch.mm(x1, x2))

3 检查pytorch所需硬件

import torchprint(torch.cuda.is_available())
print(torch.cuda.current_device())
print(torch.cuda.device(0))
print(torch.cuda.device_count())
print(torch.cuda.get_device_name(0))

4 pytorch 中的 Automatic differentiation

x = torch.ones(1, requires_grad=True)
print(x)y = x+42
print(y)z = 3*y*y
print(z)z.backward()     # 计算梯度
print(x.grad)    # ∂z/∂x = 6(x+42) = 6*1+252 = 258
print(y.grad)    # y的gradient没有保存因为没有requires_grad=True

二:The Perceptron

import torch
import torch.nn as nn# nn.Module 是所有神经网络的基类
class Perceptron(nn.Module):"""Our perceptron class"""def __init__(self, input_dim):"""Constructor"""super().__init__()self.fc = nn.Linear(input_dim, 1)self.sigmoid = torch.nn.Sigmoid()def forward(self, x_in):# squeeze unwraps the result from the singleton listreturn self.sigmoid(self.fc(x_in)) #.squeeze()print(Perceptron(10).forward(torch.ones(10)))
Activation functions

Sigmoid :f\left ( x \right )=\frac{1}{1+e^{-x}}

Tanh : f(x)=\frac{e^{x}-e^{-x}}{e^{x}+e^{-x}}

Relu :f(x)=max(0,x)

Loss function

MSE Loss:L(y,\hat{y})=\frac{1}{n}\sum_{i=1}^{n}(y_{i}-\hat{y_{i}})^{2}

import torch
import torch.nn as nnmse_loss = nn.MSELoss()
produced = torch.randn(2, 4, requires_grad=True)
print(produced)
expected = torch.randn(2, 4)
print(expected)
loss = mse_loss(produced, expected)
print(loss)

categorical cross-entropy loss

L(y,\hat{y})=-\sum_{i=1}^{n}y_{i}log(\hat{y})

import torch
import torch.nn as nnce_loss = nn.CrossEntropyLoss() # for binary classification, we can use nn.BCELoss()
produced = torch.randn(2, 4, requires_grad=True) # 2*4, normal distribution
print(produced)
# input is an index for each vector indicating the correct category/class
expected = torch.tensor([1, 0], dtype=torch.int64)
loss = ce_loss(produced, expected)
print(loss)

三:Language classification with the Perceptron

1 setup

from random import randintimport torch
from torch.utils.data import Dataset, DataLoaderimport torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

2 Data Preparation

建立LanguageRecognitionDataset类,用于处理原始data,生成我们language classification训练所需要的dataset

class LanguageRecognitionDataset(Dataset):"""An automatically generated dataset for our language classification task."""def _get_bigrams(self, sentence_list):big  rams = {}# for each sentencefor s in sentence_list:# for each bigramfor k in range(len(s)-1):bigrams[s[k:k+2]] = 1.0return bigrams.keys()def _get_bigram_vector(self, sentence):sent_bigrams = self._get_bigrams([sentence])vector = []for bigram in self.bigrams:vector.append(1.0 if bigram in sent_bigrams else 0.0)return vectordef __init__(self, sample, training_bigrams = None):"""Args:sample: List of sentences with their classification (True/False)"""self.num_samples = len(sample)if not training_bigrams:self.bigrams = self._get_bigrams([x for x, _ in sample])else:self.bigrams = training_bigramsself.data = []for sentence, gold_label in sample:sentence = sentence.lower()item = {'inputs': torch.tensor(self._get_bigram_vector(sentence)), 'outputs': torch.tensor([gold_label])}self.data.append(item)def __len__(self):return self.num_samplesdef __getitem__(self, idx):return self.data[idx]LanguageRecognitionDataset([("ciao ciao pippo", 1), ("la casa si trova in collina", 1)])[1]

3 建立一个简单的dataset

training_sentences = [("Scienziata italiana scopre la più grande esplosione nell’Universo.", 1.0),("Nell’ammasso di galassie di Ofiuco, distante 390 milioni di anni luce.", 1.0),("Ha rilasciato una quantità di energia 5 volte più grande della precedente che deteneva il primato.", 1.0),("Syria war: Turkey says thousands of migrants have crossed to EU.", 0.0),("Turkey could no longer deal with the amount of people fleeing Syria's civil war, he added.", 0.0),("Greece says it has blocked thousands of migrants from entering illegally from Turkey.", 0.0),("Tutto perfetto? Non proprio. Ci sono elementi problematici che vanno considerati.", 1.0),("Il primo è l’autonomia degli studenti, che devono essere in grado di gestire la tecnologia.", 1.0),("Il secondo, è la durata e la cadenza delle lezioni.", 1.0),("Per motivi di connessione, di competenze, di strumenti.", 1.0),("Serve un’assistenza dedicata.", 1.0),("Potremmo completare l’anno scolastico in versione virtuale?", 1.0),("Siamo preparati per affiancare la didattica tradizionale a quella virtuale, ma non siamo pronti per sostituirla", 1.0),("Various architectures of recurrent neural networks have been successful.", 0.0),("They perform tasks relating to sequence measuring", 0.0),("The networks operate by processing input components sequentially", 0.0),("They retain a hidden vector between iterations", 0.0),("It is constantly used and modified throughout the sequence.", 0.0),("They are able to model arbitrarily complicated programs.", 0.0),("L’Istituto, che raccoglie studenti di liceo scientifico, linguistico e tecnico economico, è l’esempio ideale.", 1.0),]validation_sentences = [("L’Istituto superiore di sanità ha confermato tutti i casi esaminati.", 1.0),("Measures announced after an emergency cabinet meeting also include the cancellation of the Paris half-marathon which was to be held on Sunday.", 0.0),("Lavagne in condivisione, documenti scaricabili sulla piattaforma gratuita, esercizi collaborativi.", 1.0),("Each encoder consists of two major components", 0.0),]test_sentences = [("Il ministro della Salute francese ha raccomandato di salutarsi mantenendo le distanze, mentre l’Organizzazione mondiale della sanità alza l’allerta a molto alta.", 1.0),("Possiamo riammalarci ma in questo caso si parla di ricaduta.", 1.0),("The vast majority of infections and deaths are in China, where the virus originated late last year.", 0.0),("France has banned all indoor gatherings of more than 5,000 people, as part of efforts to contain the country's coronavirus outbreak", 0.0)]def test_dataset_class():simple_dataset = LanguageRecognitionDataset(training_sentences)print('Dataset test:')for i in range(len(training_sentences)):print(f'  sample {i}: {simple_dataset[i]}')test_dataset_class()

4 Model training 

我们建立一个trainer类,其中包含了以下几个部分

  • training loop:使用模型,在数据集上迭代,来解决我们的问题
  • evaluation function:来评估我们模型的学习状态
  • prediction function:获取我们模型的output

为了让模型正确的学习,我们需要loss function来评估模型输出与真实值的差距,需要optimizer来基于loss更正模型参数

class Trainer():"""Utility class to train and evaluate a model."""def __init__(self,model,loss_function,optimizer):"""Args:model: the model we want to train.loss_function: the loss_function to minimize.optimizer: the optimizer used to minimize the loss_function."""self.model = modelself.loss_function = loss_functionself.optimizer = optimizerdef train(self, train_dataset, valid_dataset, epochs=1):"""Args:train_dataset: a Dataset or DatasetLoader instance containingthe training instances.valid_dataset: a Dataset or DatasetLoader instance used to evaluatelearning progress.epochs: the number of times to iterate over train_dataset.Returns:avg_train_loss: the average training loss on train_dataset overepochs."""assert epochs > 1 and isinstance(epochs, int)print('Training...')train_loss = 0.0for epoch in range(epochs):print(' Epoch {:03d}'.format(epoch + 1))epoch_loss = 0.0for step, sample in enumerate(train_dataset):inputs = sample['inputs']labels = sample['outputs']# we need to set the gradients to zero before starting to do backpropragation# because PyTorch accumulates the gradients on subsequent backward passesself.optimizer.zero_grad()predictions = self.model(inputs)sample_loss = self.loss_function(predictions, labels)#print("Before BP:", list(model.parameters()))sample_loss.backward()self.optimizer.step()#print("After BP:", list(model.parameters()))# sample_loss is a Tensor, tolist returns a float (alternative: use float() instead of .tolist())epoch_loss += sample_loss.tolist()print('    [E: {:2d} @ step {}] current avg loss = {:0.4f}'.format(epoch, step, epoch_loss / (step + 1)))avg_epoch_loss = epoch_loss / len(train_dataset)train_loss += avg_epoch_lossprint('  [E: {:2d}] train loss = {:0.4f}'.format(epoch, avg_epoch_loss))valid_loss = self.evaluate(valid_dataset)print('  [E: {:2d}] valid loss = {:0.4f}'.format(epoch, valid_loss))print('... Done!')avg_epoch_loss = train_loss / epochsreturn avg_epoch_lossdef evaluate(self, valid_dataset):"""Args:valid_dataset: the dataset to use to evaluate the model.Returns:avg_valid_loss: the average validation loss over valid_dataset."""valid_loss = 0.0# no gradient updates herewith torch.no_grad():for sample in valid_dataset:inputs = sample['inputs']labels = sample['outputs']predictions = self.model(inputs)sample_loss = self.loss_function(predictions, labels)valid_loss += sample_loss.tolist()return valid_loss / len(valid_dataset)def predict(self, x):"""Returns: hopefully the right prediction."""return self.model(x).tolist()

5 最后,定义dataset,setup trainer,训练我们的模型

training_dataset = DataLoader(LanguageRecognitionDataset(training_sentences), batch_size=6)
validation_dataset = DataLoader(LanguageRecognitionDataset(validation_sentences, training_dataset.dataset.bigrams), batch_size=2)
test_dataset = DataLoader(LanguageRecognitionDataset(test_sentences, training_dataset.dataset.bigrams), batch_size=2)print("Number of input dimensions", len(training_dataset.dataset.bigrams))
model = Perceptron(len(training_dataset.dataset.bigrams))
trainer = Trainer(model,loss_function = nn.MSELoss(),optimizer = optim.SGD(model.parameters(), lr=0.01)
)avg_epoch_loss = trainer.train(training_dataset, validation_dataset,epochs=50)

5 evaluation

检查我们的模型是否真的学习了一些东西

trainer.evaluate(test_dataset)for step, batch in enumerate(test_dataset):print(step, trainer.predict(batch['inputs']), batch['outputs'])

四:Language classification with a Feedforward Neural Network

1 model definition

class LanguageRecognitionFF(nn.Module):"""A simple model that classifies language"""def __init__(self, input_dim, hparams):super().__init__()# Hidden layer: transforms the input value/scalar into# a hidden vector representation.self.fc1 = nn.Linear(input_dim, hparams.hidden_size)self.relu = nn.ReLU()# Output layer: transforms the hidden vector representation# into a value/scalar (hopefully the input value + 1).self.fc2 = nn.Linear(hparams.hidden_size, 1)self.sigmoid = nn.Sigmoid()def forward(self, x):hidden = self.fc1(x)relu = self.relu(hidden)result = self.fc2(relu)return self.sigmoid(result)

2 Model Building

尽量把超参数与model definition分开,因为这样可以我们可以在不碰模型的情况下改变超参数

class HParams():hidden_size = 16

instance

model_ff = LanguageRecognitionFF(len(training_dataset.dataset.bigrams), HParams)

3 Model Training

trainer = Trainer(model = model_ff,loss_function = nn.MSELoss(),optimizer = optim.SGD(model_ff.parameters(), lr=1e-5)
)
trainer.train(training_dataset, validation_dataset, 50)

4 Model Evaluation

trainer.evaluate(test_dataset)for step, batch in enumerate(test_dataset):print(trainer.predict(batch['inputs']), batch['outputs'])

这篇关于NLP 学习笔记 1:pytorch基础操作以及Perceptron 和 FF networks实现的文章就介绍到这儿,希望我们推荐的文章对编程师们有所帮助!



http://www.chinasem.cn/article/282525

相关文章

SpringBoot全局域名替换的实现

《SpringBoot全局域名替换的实现》本文主要介绍了SpringBoot全局域名替换的实现,文中通过示例代码介绍的非常详细,对大家的学习或者工作具有一定的参考学习价值,需要的朋友们下面随着小编来一... 目录 项目结构⚙️ 配置文件application.yml️ 配置类AppProperties.Ja

Python实现批量CSV转Excel的高性能处理方案

《Python实现批量CSV转Excel的高性能处理方案》在日常办公中,我们经常需要将CSV格式的数据转换为Excel文件,本文将介绍一个基于Python的高性能解决方案,感兴趣的小伙伴可以跟随小编一... 目录一、场景需求二、技术方案三、核心代码四、批量处理方案五、性能优化六、使用示例完整代码七、小结一、

Java实现将HTML文件与字符串转换为图片

《Java实现将HTML文件与字符串转换为图片》在Java开发中,我们经常会遇到将HTML内容转换为图片的需求,本文小编就来和大家详细讲讲如何使用FreeSpire.DocforJava库来实现这一功... 目录前言核心实现:html 转图片完整代码场景 1:转换本地 HTML 文件为图片场景 2:转换 H

C#使用Spire.Doc for .NET实现HTML转Word的高效方案

《C#使用Spire.Docfor.NET实现HTML转Word的高效方案》在Web开发中,HTML内容的生成与处理是高频需求,然而,当用户需要将HTML页面或动态生成的HTML字符串转换为Wor... 目录引言一、html转Word的典型场景与挑战二、用 Spire.Doc 实现 HTML 转 Word1

C#实现一键批量合并PDF文档

《C#实现一键批量合并PDF文档》这篇文章主要为大家详细介绍了如何使用C#实现一键批量合并PDF文档功能,文中的示例代码简洁易懂,感兴趣的小伙伴可以跟随小编一起学习一下... 目录前言效果展示功能实现1、添加文件2、文件分组(书签)3、定义页码范围4、自定义显示5、定义页面尺寸6、PDF批量合并7、其他方法

SpringBoot实现不同接口指定上传文件大小的具体步骤

《SpringBoot实现不同接口指定上传文件大小的具体步骤》:本文主要介绍在SpringBoot中通过自定义注解、AOP拦截和配置文件实现不同接口上传文件大小限制的方法,强调需设置全局阈值远大于... 目录一  springboot实现不同接口指定文件大小1.1 思路说明1.2 工程启动说明二 具体实施2

Python实现精确小数计算的完全指南

《Python实现精确小数计算的完全指南》在金融计算、科学实验和工程领域,浮点数精度问题一直是开发者面临的重大挑战,本文将深入解析Python精确小数计算技术体系,感兴趣的小伙伴可以了解一下... 目录引言:小数精度问题的核心挑战一、浮点数精度问题分析1.1 浮点数精度陷阱1.2 浮点数误差来源二、基础解决

Java实现在Word文档中添加文本水印和图片水印的操作指南

《Java实现在Word文档中添加文本水印和图片水印的操作指南》在当今数字时代,文档的自动化处理与安全防护变得尤为重要,无论是为了保护版权、推广品牌,还是为了在文档中加入特定的标识,为Word文档添加... 目录引言Spire.Doc for Java:高效Word文档处理的利器代码实战:使用Java为Wo

Java实现远程执行Shell指令

《Java实现远程执行Shell指令》文章介绍使用JSch在SpringBoot项目中实现远程Shell操作,涵盖环境配置、依赖引入及工具类编写,详解分号和双与号执行多指令的区别... 目录软硬件环境说明编写执行Shell指令的工具类总结jsch(Java Secure Channel)是SSH2的一个纯J

使用Python实现Word文档的自动化对比方案

《使用Python实现Word文档的自动化对比方案》我们经常需要比较两个Word文档的版本差异,无论是合同修订、论文修改还是代码文档更新,人工比对不仅效率低下,还容易遗漏关键改动,下面通过一个实际案例... 目录引言一、使用python-docx库解析文档结构二、使用difflib进行差异比对三、高级对比方