深度知识追踪（DKT）实现pytorch（及常见问题）

Posted 2021-12-18 sereasuesue

发现代码跑了几遍还是没有自己按照思路写一遍清楚

参考代码

GitHub - dxywill/deepknowledgetracing: Pytorch implementation for Deep Knowledge tracing

论文 https://web.stanford.edu/~cpiech/bio/papers/deepKnowledgeTracing.pdf

数据集介绍

数据形式三行（第一行：答题数
第二行：题目编号（从0开始）
第三行：答题结果，0表示错，1表示对)

data.py

import csv
import random


def load_data(fileName):
    rows = []
    max_skill_num = 0  # max_skill_num是知识点(题目)个数
    max_num_problems = 0  # max_num_problems最长题目序列

    with open(fileName, "r") as csvfile:#打开文件
        reader = csv.reader(csvfile, delimiter=',')
        for row in reader:
            rows.append(row)
    index = 0
    print("the number of rows is " + str(len(rows)))
    tuple_rows = []
    #turn list to tuple
    while(index < len(rows)-1):
        problems_num = int(rows[index][0])
        tmp_max_skill = max(map(int, rows[index+1]))
        if(tmp_max_skill > max_skill_num):
            max_skill_num = tmp_max_skill
        if(problems_num <= 2):#去除了题目小于两个的数据
            index += 3
        else:
            if problems_num > max_num_problems:
                max_num_problems = problems_num
            tup = (rows[index], rows[index+1], rows[index+2])## tup:[题目个数, 题目序列, 答对情况]
            tuple_rows.append(tup)
            index += 3
    #shuffle the tuple
    random.shuffle(tuple_rows)
    print("The number of students is ", len(tuple_rows))
    print("max_skill_num  ", max_skill_num)
    print("max_num_problems  ", max_num_problems)
    print("Finish reading data")
    return tuple_rows, max_num_problems, max_skill_num+1
    #  skill序号从0开始所以最长题目序列为max_skill_num+1
# if __name__ == '__main__':
#     train_data_path='data/test1.csv'
#     load_data(train_data_path)

模型 

class DeepKnowledgeTracing(nn.Module):
    def __init__(self, input_size, hidden_size, num_skills, nlayers, dropout=0.6, tie_weights=False):
        super(DeepKnowledgeTracing, self).__init__()

        
        self.rnn = nn.LSTM(input_size, hidden_size, nlayers, batch_first=True, dropout=dropout)
        # self.rnn = nn.RNN(input_size, hidden_size, nlayers, nonlinearity='tanh', dropout=dropout)

        # nn.Linear是一个全连接层，hidden_size是输入层维数，num_skills是输出层维数
        # decoder是隐层(self.rnn)到输出层的网络
        self.decoder = nn.Linear(hidden_size, num_skills)
        self.nhid = hidden_size
        self.nlayers = nlayers

    # 前向计算, 网络结构是：input --> hidden(self.rnn) --> decoder(输出层)
    # 这里需要注意：在pytorch中，rnn的输入格式已经和tensorflow的rnn不太一样，具体见官网：
    # https://pytorch.org/docs/stable/generated/torch.nn.RNN.html?highlight=rnn#torch.nn.RNN
    # 根据官网,torch.nn.RNN接收的参数input形状是[时间步数, 批量大小, 特征维数], hidden: 旧的隐藏层的状态
    def forward(self, input, hidden):
        # output: 隐藏层在各个时间步上计算并输出的隐藏状态, 形状是[时间步数, 批量大小, 隐层维数]
        output, hidden = self.rnn(input, hidden)
        # decoded: 形状是[时间步数, 批量大小, num_skills]
        decoded = self.decoder(output.contiguous().view(output.size(0) * output.size(1), output.size(2)))
        return decoded, hidden

可改进的地方

其他代码与其的不同的地方

可参考文章：DKT学习_qq_40282662的博客-CSDN博客

Deep Knowledge Tracing [Pytorch] | 一切皆可解读 (chsong.live)

torch.optim.Adam(model.parameters(), lr=lr ,eps=args.epsilon)

影响模型效果的因素

学习率，batchsize 一般设置为8的倍数 还有epsilon不要过大 0.01或0.001都可