NLP | 基于LLMs的文本分类任务

比赛链接：讯飞开放平台

来源：DataWhale AI夏令营3（NLP）

Roberta-base（BERT的改进）

①Roberta在预训练的阶段中没有对下一句话进行预测（NSP）

②采用了动态掩码 ③使用字符级和词级别表征的混合文本编码。

论文：https://arxiv.org/pdf/1907.11692.pdf

DataWhale Topline的改进：

  特征1：平均池化MeanPooling(768维) -> 全连接层fc(128维)

  特征2：末隐藏层Last_hidden (768维) -> 全连接层fc(128维)

运行方式：阿里云机器学习平台PAI-交互式建模DSW

镜像选择：pytorch:1.12-gpu-py39-cu113-ubuntu20.04

上传代码，解压指令：

unzip [filename]

运行py脚本指令（遇到网络错误重新运行即可）：
 

python [python_filename]

---

① 数据处理模块

导入需要的模块：

from transformers import AutoTokenizer  #文本分词
import pandas as pd
import numpy as np
from tqdm import tqdm  #显示进度条
import torch
from torch.nn.utils.rnn import pad_sequence
#填充序列，保证向量中各序列维度的大小一样MAX_LENGTH = 128  #定义最大序列长度为128

训练集制作：

def get_train(model_name, model_dict):model_index = model_dict[model_name]  # 获取模型索引train = pd.read_csv('./dataset/train.csv') #读取训练数据为DataFrametrain['content'] = train['title'] + train['author'] + train['abstract']  #将标题、作者和摘要拼接为训练内容tokenizer = AutoTokenizer.from_pretrained(model_name, max_length=MAX_LENGTH, cache_dir=f'./premodels/{model_name}_saved')  # 实例化分词器对象# 通过分词器对训练数据进行分词，并获取输入ID、注意力掩码和标记类型ID（这个可有可无）input_ids_list, attention_mask_list, token_type_ids_list = [], [], []y_train = []  # 存储训练数据的标签for i in tqdm(range(len(train['content']))):  # 遍历训练数据sample = train['content'][i]  # 获取样本内容tokenized = tokenizer(sample, truncation='longest_first') #分词处理，【最长优先方式】截断input_ids, attention_mask = tokenized['input_ids'], tokenized['attention_mask']  # 获取输入ID和注意力掩码input_ids, attention_mask = torch.tensor(input_ids), torch.tensor(attention_mask)  # 转换为PyTorch张量try:token_type_ids = tokenized['token_type_ids']  # 获取标记类型IDtoken_type_ids = torch.tensor(token_type_ids) # 转换为PyTorch张量except:token_type_ids = input_ids #异常处理input_ids_list.append(input_ids)  # 将输入ID添加到列表中attention_mask_list.append(attention_mask)  # 将注意力掩码添加到列表中token_type_ids_list.append(token_type_ids)  # 将标记类型ID添加到列表中y_train.append(train['label'][i])  # 将训练数据的标签添加到列表中# 保存 对下述对象进行填充，保证向量中各序列维度的大小一样，生成张量# 输入      ID input_ids_tensor、# 注意力掩码 attention_mask_tensor# 标记类型ID token_type_ids_tensorinput_ids_tensor = pad_sequence(input_ids_list, batch_first=True, padding_value=0)attention_mask_tensor = pad_sequence(attention_mask_list, batch_first=True, padding_value=0)token_type_ids_tensor = pad_sequence(token_type_ids_list, batch_first=True, padding_value=0) x_train = torch.stack([input_ids_tensor, attention_mask_tensor, token_type_ids_tensor], dim=1)  # 将输入张量堆叠为一个张量x_train = x_train.numpy()   # 转换为NumPy数组（ndarray）np.save(f'./models_input_files/x_train{model_index}.npy', x_train) #保存训练数据y_train = np.array(y_train) # 转换为NumPy数组（ndarray）np.save(f'./models_input_files/y_train{model_index}.npy', y_train) #保存标签数据

测试集制作：

def get_test(model_name, model_dict):model_index = model_dict[model_name]  # 获取模型索引test = pd.read_csv('./dataset/testB.csv')  # 从CSV文件中读取测试数据为DataFrametest['content'] = test['title'] + ' ' + test['author'] + ' ' + test['abstract']  # 将标题、作者和摘要拼接为测试内容tokenizer = AutoTokenizer.from_pretrained(model_name, max_length=MAX_LENGTH,cache_dir=f'./premodels/{model_name}_saved')  # 实例化分词器对象# 通过分词器对测试数据进行分词，创建输入ID、注意力掩码和标记类型ID列表进行记录（可有可无）input_ids_list, attention_mask_list, token_type_ids_list = [], [], []for i in tqdm(range(len(test['content']))):  # 遍历测试数据sample = test['content'][i]  # 获取样本内容tokenized = tokenizer(sample, truncation='longest_first')  # 分词处理，使用最长优先方式截断input_ids, attention_mask = tokenized['input_ids'], tokenized['attention_mask']  # 获取输入ID和注意力掩码input_ids, attention_mask = torch.tensor(input_ids), torch.tensor(attention_mask)  # 转换为PyTorch张量try:token_type_ids = tokenized['token_type_ids']  # 获取标记类型IDtoken_type_ids = torch.tensor(token_type_ids)  # 转换为PyTorch张量except:token_type_ids = input_ids #异常处理input_ids_list.append(input_ids)  # 将输入ID添加到列表中attention_mask_list.append(attention_mask)  # 将注意力掩码添加到列表中token_type_ids_list.append(token_type_ids)  # 将标记类型ID添加到列表中# 保存，对输入ID、注意力掩码、标记类型ID进行填充，保证向量中各序列维度的大小一样，生成张量input_ids_tensor = pad_sequence(input_ids_list, batch_first=True, padding_value=0) attention_mask_tensor = pad_sequence(attention_mask_list, batch_first=True, padding_value=0)token_type_ids_tensor = pad_sequence(token_type_ids_list, batch_first=True, padding_value=0) x_test = torch.stack([input_ids_tensor, attention_mask_tensor, token_type_ids_tensor], dim=1)  # 将输入张量堆叠为一个张量x_test = x_test.numpy()  # 转换为NumPy数组np.save(f'./models_input_files/x_test{model_index}.npy', x_test)  # 保存测试数据

划分训练集和验证集：

def split_train(model_name, model_dict):# 训练集:验证集 = 9 : 1split_rate = 0.90# 处理样本内容model_index = model_dict[model_name]  # 获取模型索引train = np.load(f'./models_input_files/x_train{model_index}.npy')  # 加载训练数据state = np.random.get_state()  # 获取随机数状态，保证样本间的随机是可重复的# 或者也可以设置经典随机种子random_seed=42np.random.shuffle(train)  # 随机打乱训练数据，数据洗牌val = train[int(train.shape[0] * split_rate):]  # 划分验证集 validationtrain = train[:int(train.shape[0] * split_rate)]  # 划分训练集 train setnp.save(f'./models_input_files/x_train{model_index}.npy', train)  # 保存训练集np.save(f'./models_input_files/x_val{model_index}.npy', val)  # 保存验证集train = np.load(f'./models_input_files/y_train{model_index}.npy')  # 加载标签数据# 处理样本标签np.random.set_state(state)  # 恢复随机数状态，让样本标签的随机可重复np.random.shuffle(train)  # 随机打乱标签数据val = train[int(train.shape[0] * split_rate):]  # 划分验证集 validationtrain = train[:int(train.shape[0] * split_rate)]  # 划分训练集 train setnp.save(f'./models_input_files/y_train{model_index}.npy', train)  # 保存训练集标签np.save(f'./models_input_files/y_val{model_index}.npy', val)  # 保存验证集标签print('split done.')

数据处理主函数：

if __name__ == '__main__':model_dict = {'xlm-roberta-base':1,'roberta-base':2, 'bert-base-uncased':3, 'microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext':4, 'dmis-lab/biobert-base-cased-v1.2':5, 'marieke93/MiniLM-evidence-types':6,'microsoft/MiniLM-L12-H384-uncased':7, 'cambridgeltl/SapBERT-from-PubMedBERT-fulltext':8,'microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract':9,'microsoft/BiomedNLP-PubMedBERT-large-uncased-abstract':10}model_name = 'roberta-base'get_train(model_name, model_dict) #读取训练集get_test(model_name, model_dict)  #读取测试集split_train(model_name, model_dict) #划分训练集和测试集

---

② 模型训练

导入需要的模块：

import numpy as np
import torch
import torch.nn as nn
from sklearn import metrics
import os
import time
from transformers import AutoModel, AutoConfig
# 导入AutoModel和AutoConfig类，用于加载预训练模型
from tqdm import tqdm  #显示进度条

超参数类（可修改的所有超参数）：

class opt:seed               = 42 # 随机种子batch_size         = 16 # 批处理大小set_epoch          = 5  # 训练轮数 early_stop         = 5  # 提前停止epoch数learning_rate      = 1e-5 # 学习率weight_decay       = 2e-6 # 权重衰减,L2正则化device             = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 选择设备,GPU或CPUgpu_num            = 1 # GPU个数use_BCE            = False # 是否使用BCE损失函数models             = ['xlm-roberta-base', 'roberta-base', 'bert-base-uncased',  'microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext', 'dmis-lab/biobert-base-cased-v1.2', 'marieke93/MiniLM-evidence-types',  'microsoft/MiniLM-L12-H384-uncased','cambridgeltl/SapBERT-from-PubMedBERT-fulltext', 'microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract','microsoft/BiomedNLP-PubMedBERT-large-uncased-abstract'] # 模型名称列表model_index        = 2 # 根据上面选择使用的模型，这里填对应的模型索引model_name         = models[model_index-1] # 使用的模型名称continue_train     = False # 是否继续训练show_val           = False # 是否显示验证过程

定义模型类：

# 定义模型
class MODEL(nn.Module):def __init__(self, model_index):super(MODEL, self).__init__()# 若是第一次下载权重，则下载至同级目录的./premodels/内，以防占主目录的存储空间self.model = AutoModel.from_pretrained(opt.models[model_index-1], cache_dir='./premodels/'+opt.models[model_index-1]+'_saved', from_tf=False) # 加载预训练语言模型# 加载模型配置，可以直接获得模型最后一层的维度，而不需要手动修改config = AutoConfig.from_pretrained(opt.models[model_index-1], cache_dir='./premodels/'+opt.models[model_index-1]+'_saved') # 获取配置last_dim = config.hidden_size # 最后一层的维度if opt.use_BCE:out_size = 1 # 损失函数如果使用BCE,则输出大小为1else          :out_size = 2 # 否则则使用CE,输出大小为2feature_size = 128 # 设置特征的维度大小self.fc1 = nn.Linear(last_dim, feature_size) # 全连接层1self.fc2 = nn.Linear(last_dim, feature_size) # 全连接层2self.classifier = nn.Linear(feature_size, out_size) # 分类器self.dropout = nn.Dropout(0.3) # Dropout层def forward(self, x): #BPinput_ids, attention_mask, token_type_ids = x[:,0],x[:,1],x[:,2] # 获取输入x = self.model(input_ids, attention_mask) # 通过模型all_token     = x[0] # 全部序列分词的表征向量pooled_output = x[1] # [CLS]的表征向量+一个全连接层+Tanh激活函数feature1 = all_token.mean(dim=1) # 对全部序列分词的表征向量取均值feature1 = self.fc1(feature1)    # 再输入进全连接层，得到feature1feature2 = pooled_output      # [CLS]的表征向量+一个全连接层+Tanh激活函数feature2 = self.fc2(feature2) # 再输入进全连接层，得到feature2feature  = 0.5*feature1 + 0.5*feature2 # 加权融合特征feature  = self.dropout(feature) # Dropoutx  = self.classifier(feature) # 分类return x

数据加载：

def load_data():#数据集路径train_data_path     = f'models_input_files/x_train{model_index}.npy'train_label_path    = f'models_input_files/y_train{model_index}.npy'val_data_path       = f'models_input_files/x_val{model_index}.npy'# 验证集val_label_path      = f'models_input_files/y_val{model_index}.npy'# 验证集标签test_data_path      = f'models_input_files/x_test{model_index}.npy'# 测试集输入#数据集读取#data=torch.tensor([path],allow_pickle=True).tolist())train_data          = torch.tensor(np.load(train_data_path  , allow_pickle=True).tolist())train_label         = torch.tensor(np.load(train_label_path  , allow_pickle=True).tolist()).long() val_data            = torch.tensor(np.load(val_data_path  , allow_pickle=True).tolist()) val_label           = torch.tensor(np.load(val_label_path  , allow_pickle=True).tolist()).long()test_data           = torch.tensor(np.load(test_data_path  , allow_pickle=True).tolist()) #构造训练集、验证集、测试集train_dataset       = torch.utils.data.TensorDataset(train_data  , train_label) val_dataset         = torch.utils.data.TensorDataset(val_data  , val_label) test_dataset        = torch.utils.data.TensorDataset(test_data) return train_dataset, val_dataset, test_dataset # 返回数据集

模型预训练：

def model_pretrain(model_index, train_loader, val_loader):# 超参数设置set_epoch          = opt.set_epoch  # 训练轮数early_stop         = opt.early_stop # 提前停止epoch数learning_rate      = opt.learning_rate # 学习率weight_decay       = opt.weight_decay  # 权重衰减device             = opt.device  # 设备 gpu_num            = opt.gpu_num # GPU个数continue_train     = opt.continue_train # 是否继续训练model_save_dir     = 'checkpoints' # 模型保存路径# 是否要继续训练，若是，则加载模型进行训练；若否，则跳过训练，直接对测试集进行推理if not continue_train:# 判断最佳模型是否已经存在,若存在则直接读取,若不存在则进行训练if os.path.exists(f'checkpoints/best_model{model_index}.pth'): best_model = MODEL(model_index)best_model.load_state_dict(torch.load(f'checkpoints/best_model{model_index}.pth')) # 加载模型return best_modelelse:pass# 模型初始化model = MODEL(model_index).to(device) if continue_train:model.load_state_dict(torch.load(f'checkpoints/best_model{model_index}.pth')) # 继续训练加载模型# 优化器初始化if device    != 'cpu' and gpu_num > 1:  # 多张显卡optimizer = torch.optim.AdamW(model.module.parameters(), lr=learning_rate, weight_decay=weight_decay)optimizer = torch.nn.DataParallel(optimizer, device_ids=list(range(gpu_num))) # 多GPUelse: # 单张显卡optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay) # 单GPU# 损失函数初始化if opt.use_BCE:loss_func = nn.BCEWithLogitsLoss() # BCE损失else:loss_func = nn.CrossEntropyLoss() # 交叉熵损失（CE）# 模型训练best_epoch         = 0 # 最佳epochbest_train_loss    = 100000 # 最佳训练损失train_acc_list     = [] # 训练准确率列表train_loss_list    = [] # 训练损失列表val_acc_list       = [] # 验证准确率列表 val_loss_list      = [] # 验证损失列表start_time         = time.time() # 训练开始时间for epoch in range(set_epoch): # 轮数model.train() # 模型切换到训练模式train_loss = 0 # 训练损失train_acc = 0 # 训练准确率for x, y in tqdm(train_loader): # 遍历训练集# 训练前先将数据放到GPU上x        = x.to(device)y        = y.to(device)outputs  = model(x) # 前向传播if opt.use_BCE: # BCE损失loss = loss_func(outputs, y.float().unsqueeze(1)) else: # 交叉熵损失loss = loss_func(outputs, y)train_loss += loss.item() # 累加训练损失optimizer.zero_grad() # 清空梯度loss.backward() # 反向传播if device != 'cpu' and gpu_num > 1: # 多GPU更新optimizer.module.step()  else:optimizer.step() # 单GPU更新if not opt.use_BCE: # 非BCE损失_, predicted = torch.max(outputs.data, 1) # 预测结果else:predicted = (outputs > 0.5).int() # 预测结果predicted = predicted.squeeze(1) train_acc   += (predicted == y).sum().item() # 计算训练准确率average_mode = 'binary'# 计算F1、Precision、Recalltrain_f1     = metrics.f1_score(y.cpu(), predicted.cpu(), average=average_mode)train_pre    = metrics.precision_score(y.cpu(), predicted.cpu(), average=average_mode)train_recall = metrics.recall_score(y.cpu(), predicted.cpu(), average=average_mode)train_loss /= len(train_loader) # 平均所有步数的训练损失作为一个epoch的训练损失train_acc  /= len(train_loader.dataset) # 平均所有步数训练准确率作为一个epoch的准确率train_acc_list.append(train_acc)   # 添加训练准确率train_loss_list.append(train_loss) # 添加训练损失print('-'*50)print('Epoch [{}/{}]\n Train Loss: {:.4f}, Train Acc: {:.4f}'.format(epoch + 1, set_epoch, train_loss, train_acc))print('Train-f1: {:.4f}, Train-precision: {:.4f} Train-recall: {:.4f}'.format(train_f1, train_pre, train_recall))if opt.show_val: # 显示验证过程# 验证model.eval() # 模型切换到评估模式val_loss = 0 # 验证损失val_acc = 0 # 验证准确率for x, y in tqdm(val_loader): # 遍历验证集# 训练前先将数据放到GPU上x = x.to(device) y = y.to(device)outputs = model(x) # 前向传播if opt.use_BCE: # BCE损失loss = loss_func(outputs, y.float().unsqueeze(1))else: # 交叉熵损失  loss = loss_func(outputs, y)val_loss += loss.item() # 累加验证损失if not opt.use_BCE: # 非BCE损失_, predicted = torch.max(outputs.data, 1) else:predicted = (outputs > 0.5).int() # 预测结果predicted = predicted.squeeze(1)val_acc += (predicted == y).sum().item() # 计算验证准确率#计算F1、Precision、Recallval_f1     = metrics.f1_score(y.cpu(), predicted.cpu(), average=average_mode)val_pre    = metrics.precision_score(y.cpu(), predicted.cpu(), average=average_mode)val_recall = metrics.recall_score(y.cpu(), predicted.cpu(), average=average_mode)val_loss /= len(val_loader) # 平均验证损失val_acc /= len(val_loader.dataset) # 平均验证准确率val_acc_list.append(val_acc)   # 添加验证准确率val_loss_list.append(val_loss) # 添加验证损失print('\nVal Loss: {:.4f}, Val Acc: {:.4f}'.format(val_loss, val_acc))print('Val-f1: {:.4f}, Val-precision: {:.4f} Val-recall: {:.4f}'.format(val_f1, val_pre, val_recall))if train_loss < best_train_loss: # 更新最佳训练损失best_train_loss = train_lossbest_epoch = epoch + 1if device == 'cuda' and gpu_num > 1: # 多GPU保存模型torch.save(model.module.state_dict(), f'{model_save_dir}/best_model{model_index}.pth')else:torch.save(model.state_dict(), f'{model_save_dir}/best_model{model_index}.pth') # 单GPU保存模型# 提前停止判断if epoch+1 - best_epoch == early_stop:  print(f'{early_stop} epochs later, the loss of the validation set no longer continues to decrease, so the training is stopped early.')end_time = time.time()print(f'Total time is {end_time - start_time}s.')breakbest_model = MODEL(model_index) # 初始化最佳模型best_model.load_state_dict(torch.load(f'checkpoints/best_model{model_index}.pth')) # 加载模型参数return best_model # 返回最佳模型

模型推理：

def model_predict(model, model_index, test_loader):device = 'cuda'model.to(device) # 模型到GPUmodel.eval()  # 切换到评估模式test_outputs = Nonewith torch.no_grad():  # 禁用梯度计算for i, data in enumerate(tqdm(test_loader)):data = data[0].to(device) # 测试数据到GPUoutputs = model(data) # 前向传播if i == 0: test_outputs = outputs # 第一个batch直接赋值else:test_outputs = torch.cat([test_outputs, outputs], dim=0) # 其余batch拼接del data, outputs  # 释放不再需要的Tensor# 保存预测结果    if not opt.use_BCE: test_outputs = torch.softmax(test_outputs, dim=1) # 转换为概率torch.save(test_outputs, f'./models_prediction/{model_index}_prob.pth') # 保存概率

模型训练主函数：

def run(model_index):# 固定随机种子seed = opt.seed  torch.seed = seednp.random.seed(seed)torch.manual_seed(seed)torch.cuda.manual_seed(seed)torch.cuda.manual_seed_all(seed)torch.backends.cudnn.deterministic = Truetrain_dataset, val_dataset, test_dataset = load_data() # 加载数据集# 打印数据集信息print('-数据集信息:')print(f'-训练集样本数:{len(train_dataset)},测试集样本数:{len(test_dataset)}')train_labels = len(set(train_dataset.tensors[1].numpy()))# 查看训练样本类别均衡状况print(f'-训练集的标签种类个数为:{train_labels}') numbers = [0] * train_labelsfor i in train_dataset.tensors[1].numpy():numbers[i] += 1print(f'-训练集各种类样本的个数:')for i in range(train_labels):print(f'-{i}的样本个数为:{numbers[i]}')batch_size   = opt.batch_size # 批处理大小# 构建DataLoadertrain_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True) val_loader   = torch.utils.data.DataLoader(dataset=val_dataset,   batch_size=batch_size, shuffle=True)test_loader  = torch.utils.data.DataLoader(dataset=test_dataset,  batch_size=batch_size, shuffle=False)best_model   = model_pretrain(model_index, train_loader, val_loader)# 使用验证集评估模型model_predict(best_model, model_index, test_loader) # 模型推理if __name__ == '__main__':model_index = opt.model_index # 获取模型索引run(model_index) # 运行程序

---

③ 模型评估

import torch
import pandas as pd
from models_training import MODEL  # 从本地文件models_training.py中导入MODEL类
from tqdm import tqdm
from sklearn.metrics import classification_report
import numpy as np# 推理
def inference(model_indexs, use_BCE):device = 'cuda'  # 设备选择为cudafor model_index in model_indexs:# 加载模型model = MODEL(model_index).to(device)  # 创建MODEL类的实例，并将模型移至设备(device)model.load_state_dict(torch.load(f'checkpoints/best_model{model_index}.pth'))  # 加载模型的权重参数model.eval()  # 切换到评估模式# 加载val数据val_data_path = f'models_input_files/x_val{model_index}.npy'  # val数据的路径val_data = torch.tensor(np.load(val_data_path, allow_pickle=True).tolist())  # 加载val数据，并转换为Tensor格式val_dataset = torch.utils.data.TensorDataset(val_data)  # 创建val数据集val_loader  = torch.utils.data.DataLoader(dataset=val_dataset, batch_size=32, shuffle=False)  # 创建val数据的数据加载器val_outputs = None  # 初始化val_outputs变量with torch.no_grad():  # 禁用梯度计算for i, data in enumerate(tqdm(val_loader)):  # 遍历val_loader，显示进度条data = data[0].to(device)  # 将数据移至GPUoutputs = model(data)  # 模型推理，获取输出if i == 0:val_outputs = outputs  # 若为第一次迭代，直接赋值给val_outputselse:val_outputs = torch.cat([val_outputs, outputs], dim=0)  # 否则在dim=0上拼接val_outputs和outputsdel data, outputs  # 释放不再需要的Tensor对象# 输出预测概率if not use_BCE:val_outputs = torch.softmax(val_outputs, dim=1)  # 对val_outputs进行softmax操作torch.save(val_outputs, f'evaluate_prediction/{model_index}_prob.pth')  # 保存预测概率结果def run(model_indexs, use_BCE):# 读取所有的model_prob.pth，并全加在一起avg_pred = None  # 初始化avg_pred变量for i in model_indexs:pred = torch.load(f'evaluate_prediction/{i}_prob.pth').data  # 加载预测概率结果if use_BCE:# 选取大于0.5的作为预测结果pred = (pred > 0.5).int()  # 将大于0.5的值转换为整数（0或1）pred = pred.reshape(-1)  # 将预测结果进行形状重塑else:# 选取最大的概率作为预测结果pred = torch.argmax(pred, dim=1)  # 获取最大概率的索引作为预测结果pred = pred.cpu().numpy()  # 将预测结果转移到CPU上，并转换为NumPy数组# to_evaluate# 读取真实标签val_label_path = f'models_input_files/y_val{i}.npy'  # 真实标签的路径y_true = np.load(val_label_path)  # 加载真实标签# 分类报告print(f'model_index = {i}:')print(classification_report(y_true, pred, digits=4))  # 打印分类报告，包括精确度、召回率等指标zero_acc = 0; one_acc = 0 # 初始化0类和1类的准确率zero_num = 0; one_num= 0  # 初始化0类和1类的样本数量for i in range(pred.shape[0]):if y_true[i] == 0:zero_num += 1  # 统计0类的样本数量elif y_true[i] == 1:one_num += 1  # 统计1类的样本数量if pred[i] == y_true[i]:if pred[i] == 0:zero_acc += 1  # 统计0类的正确预测数量elif pred[i] == 1:one_acc += 1  # 统计1类的正确预测数量zero = np.sum(pred == 0) / pred.shape[0]  # 计算预测为0类的样本占比zero_acc /= zero_num  # 计算0类的正确率print(f'预测0类占比：{zero}  0类正确率：{zero_acc}')one = np.sum(pred == 1) / pred.shape[0]  # 计算预测为1类的样本占比one_acc /= one_num  # 计算1类的正确率print(f'预测1类占比：{one}  1类正确率：{one_acc}')print('-' * 80)if __name__ == '__main__':use_BCE = False  # 是否使用BCE损失函数的标志，这里我只用交叉熵CE，所以是Falseinference([2], use_BCE=use_BCE)  # 进行推理，传入模型索引和use_BCE标志model_indexs = [2]  # 模型索引列表run(model_indexs, use_BCE=use_BCE)  # 进行运行，传入模型索引和use_BCE标志

---

④ 测试集推理

import torch
import pandas as pd
import warnings # 过滤警告
warnings.filterwarnings('ignore')def run(model_indexs, use_BCE):# 记录模型数量model_num = len(model_indexs)# 读取所有的model_prob.pth，并全加在一起for i in model_indexs:# 加载模型在训练完成后对测试集推理所得的预测文件pred = torch.load(f'./models_prediction/{i}_prob.pth', map_location='cpu').data# 这里的操作是将每个模型对测试集推理的概率全加在一起if i == model_indexs[0]:avg_pred = predelse:avg_pred += pred# 取平均avg_pred /= model_num # 使用全加在一起的预测概率除以模型数量if use_BCE:# 选取概率大于0.5的作为预测结果pred = (avg_pred > 0.5).int()pred = pred.reshape(-1)else:# 后处理 - 根据标签数目的反馈，对预测阈值进行调整pred[:, 0][pred[:, 0]>0.001] = 1pred[:, 1][pred[:, 1]>0.999] = 1.2# 选取最大的概率作为预测结果pred = torch.argmax(avg_pred, dim=1)pred = pred.cpu().numpy()# to_submit# 读取test.csv文件test = pd.read_csv('./dataset/testB_submit_exsample.csv')# 开始写入预测结果for i in range(len(pred)):test['label'][i] = pred[i]print(test['label'].value_counts())# 保存为提交文件test.to_csv(f'submit.csv',index=False)if __name__ == '__main__':run([2], use_BCE=False)# run([1,2,3,4,5,6,7,8,9,10], use_BCE=False)

模型优化的思路：

超参数调整、最大序列长度调整、损失函数更改、模型参数冻结

特征工程、模型集成、对比学习、提示学习サ

ChatGML2-6B

LLMs：自回归模型

Pretrained => prompt、finetune => RLHF 强化对齐学习

LoRA低秩适应：冻结预训练好的模型权重参数，在冻结原模型参数的情况下，通过往模型中加入额外的网络层，并只训练这些新增的网络层参数。

「instruction --> 」「input: X」「output: Y」

P-tuning v2：在原有的大型语言模型上添加一些新的参数，这些新的参数可以帮助模型更好地理解和处理特定的任务。

微调应用：垂直领域、个性化

在阿里云Pytorch环境中，克隆代码、下载chatglm2-6b模型，

安装依赖，并且运行训练脚本。

xfg_train.sh

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \--model_name_or_path chatglm2-6b \ 本地模型的目录--stage sft \ 微调方法--use_v2 \ 使用glm2模型微调，默认值true--do_train \ 是否训练，默认值true--dataset paper_label \ 数据集名字--finetuning_type lora \ --lora_rank 8 \  LoRA 微调中的秩大小--output_dir ./output/label_xfg \ 输出lora权重存放目录--per_device_train_batch_size 4 \ 用于训练的批处理大小--gradient_accumulation_steps 4 \ 梯度累加次数--lr_scheduler_type cosine \--logging_steps 10 \ 日志输出间隔--save_steps 1000 \ 断点保存间隔--learning_rate 5e-5 \ 学习率--num_train_epochs 4.0 \ 训练轮数--fp16 是否使用 fp16 半精度 默认值：False

导入数据

import pandas as pd
train_df = pd.read_csv('./csv_data/train.csv')
testB_df = pd.read_csv('./csv_data/testB.csv')

制作数据集

res = [] #存储数据样本for i in range(len(train_df)):# 遍历训练数据的每一行paper_item = train_df.loc[i] # 获取当前行的数据# 创建一个字典，包含LoRA的指令、输入和输出信息tmp = {"instruction": "Please judge whether it is a medical field paper according to the given paper title and abstract, output 1 or 0, the following is the paper title and abstract -->","input": f"title:{paper_item[1]},abstract:{paper_item[3]}","output": str(paper_item[5])}res.append(tmp) # 将字典添加到结果列表中import json #用于保存数据集# 将制作好的数据集保存到data目录下
with open('./data/paper_label.json', mode='w', encoding='utf-8') as f:json.dump(res, f, ensure_ascii=False, indent=4)

修改data/data_info.json

{"paper_label": {"file_name": "paper_label.json"}
}

加载训练好的LoRA权重，进行预测

from peft import PeftModel
from transformers import AutoTokenizer, AutoModel, GenerationConfig, AutoModelForCausalLM# 定义预训练模型的路径
model_path = "../chatglm2-6b"
model = AutoModel.from_pretrained(model_path, trust_remote_code=True).half().cuda()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)# 加载 label lora权重
model = PeftModel.from_pretrained(model, './output/label_xfg').half()
model = model.eval()# 使用加载的模型和分词器进行聊天，生成回复
response, history = model.chat(tokenizer, "你好", history=[])
response

预测函数：

def predict(text):# 使用加载的模型和分词器进行聊天，生成回复response, history = model.chat(tokenizer, f"Please judge whether it is a medical field paper according to the given paper title and abstract, output 1 or 0, the following is the paper title and abstract -->{text}", history=[],temperature=0.01)return response

预测，导出csv

from tqdm import tqdm #预测过程的进度条label = [] #存储预测结果for i in tqdm(range(len(testB_df))): # 遍历测试集中的每一条样本test_item = testB_df.loc[i]      # 测试集中的每一条样本# 构建预测函数的输入：prompttest_input = f"title:{test_item[1]},author:{test_item[2]},abstract:{test_item[3]}"label.append(int(predict(test_input)))# 预测结果存入lable列表testB_df['label'] = label # 把label列表存入testB_df# task1虽然只需要label，但需要有一个keywords列，用个随意的字符串代替
testB_df['Keywords'] = ['tmp' for _ in range(2000)]# 制作submit，提交submit
submit = testB_df[['uuid', 'Keywords', 'label']]
submit.to_csv('submit.csv', index=False)

提交结果：

 ライト