[HFUT Student]当我使用CNCD_F模型时出现问题【我已解决问题】

[senker216]

报错代码
from edustudio.quickstart import run_edustudio

run_edustudio(
dataset='AAAI_2023',
cfg_file_name=None,
traintpl_cfg_dict={
'cls': 'GeneralTrainTPL',
},
datatpl_cfg_dict={
'cls': 'CNCDFDataTPL',
},
modeltpl_cfg_dict={
'cls': 'CNCD_F'
},
evaltpl_cfg_dict={
'clses': ['PredictionEvalTPL'],
}
)
报错为
2024-12-21 11:49:40[ERROR]: Traceback (most recent call last):
File "E:\EduStudio-main\edustudio\quickstart\quickstart.py", line 58, in run_edustudio
traintpl.start()
File "E:\EduStudio-main\edustudio\traintpl\gd_traintpl.py", line 79, in start
metrics = self.one_fold_start(fold_id)
File "E:\EduStudio-main\edustudio\traintpl\general_traintpl.py", line 53, in one_fold_start
self.fit(train_loader=self.train_loader, valid_loader=self.valid_loader)
File "E:\EduStudio-main\edustudio\traintpl\general_traintpl.py", line 96, in fit
loss_dict = self.model.get_loss_dict(**batch_dict)
File "E:\EduStudio-main\edustudio\model\CD\cncd_f.py", line 177, in get_loss_dict
return self.get_main_loss(**kwargs)
File "E:\EduStudio-main\edustudio\model\CD\cncd_f.py", line 170, in get_main_loss
pd = self(stu_id, exer_id).flatten()
File "D:\anac\envs\wl\lib\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "E:\EduStudio-main\edustudio\model\CD\cncd_f.py", line 144, in forward
items_Q_mat = self.Q_mat[exer_id].to(self.traintpl_cfg['device'])
RuntimeError: indices should be either on cpu or on the same device as the indexed tensor (cpu)
这个问题意思是在使用 PyTorch 进行索引操作时，张量（tensor）和索引（indices）不在同一个设备（device）上。加入调试信息查看后发现
stu_id device: cuda:0
exer_id device: cuda:0
self.word_ids device: cuda:0
self.Q_mat device: cpu
找到问题的根源是self.Q_mat没有转移到GPU，修改cncd_f.py代码

def build_model(self):
# prediction sub-net
self.textcnn = TextCNN(self.traintpl_cfg["batch_size"], self.n_cpt, self.text_dim)
self.word2id = self.textcnn.prepare_embedding(self.content_list)
self.word_ids = self.get_from_text().to(self.traintpl_cfg['device']) # 确保在指定设备上

# 确保 Q_mat 在正确的设备上
self.Q_mat = self.Q_mat.to(self.traintpl_cfg['device'])

# 打印设备信息
print(f"self.word_ids device: {self.word_ids.device}")
print(f"self.Q_mat device: {self.Q_mat.device}")

self.out_text_factor = nn.Linear(self.text_dim, 1)
self.student_emb = nn.Embedding(self.n_user, self.n_cpt+1)
self.k_difficulty = nn.Embedding(self.n_item, self.n_cpt)
self.e_difficulty = nn.Embedding(self.n_item, 1)
self.pd_net = PosMLP(
    input_dim=self.n_cpt+1, output_dim=1, activation=self.modeltpl_cfg['activation'],
    dnn_units=self.modeltpl_cfg['dnn_units'], dropout_rate=self.modeltpl_cfg['dropout_rate']
)

def forward(self, stu_id, exer_id):
# 确保索引和张量在同一设备上
stu_id = stu_id.to(self.traintpl_cfg['device'])
exer_id = exer_id.to(self.traintpl_cfg['device'])

# before prednet
items_Q_mat = self.Q_mat[exer_id].to(self.traintpl_cfg['device'])  # 移动到 GPU
items_content = self.word_ids[exer_id]
text_embedding = self.textcnn(items_content)
text_factor = torch.sigmoid(self.out_text_factor(text_embedding))
stu_emb = self.student_emb(stu_id)
stat_emb = torch.sigmoid(stu_emb)
k_difficulty = torch.sigmoid(self.k_difficulty(exer_id))
k_difficulty = torch.cat((k_difficulty, text_factor), dim=1)
e_difficulty = torch.sigmoid(self.e_difficulty(exer_id)) * self.modeltpl_cfg['disc_scale']
# prednet
text_factor_q = torch.ones((items_Q_mat.shape[0], 1), device=self.traintpl_cfg['device'])
input_knowledge_point = torch.cat((items_Q_mat, text_factor_q), dim=1)
input_x = e_difficulty * (stat_emb - k_difficulty) * input_knowledge_point
pd = self.pd_net(input_x).sigmoid()
return pd

更换这些代码即可

[jsxie9]

你好，运行run_cncd_f_demo.py并没有复现你提到的问题