Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

模型效果复现相关 #6

Open
KpiHang opened this issue Dec 5, 2023 · 8 comments
Open

模型效果复现相关 #6

KpiHang opened this issue Dec 5, 2023 · 8 comments

Comments

@KpiHang
Copy link

KpiHang commented Dec 5, 2023

对代码中提供的zhihu数据集,进行相关实验:

使用README.md中给出的命令:
python -u DreamRec.py --data zhihu --timesteps 500 --lr 0.01 --beta_sche linear --w 4 --optimizer adamw --diffuser_type mlp1 --random_seed 100

达不到论文中的效果,HR@20 < 2.00%,NDCG@20<0.70%,

使用optuna调参后,得到最大的ndcg@20效果;
image

ndcg@20次优结果:
image

不过所有的HR@20 均达不到论文中给出的效果,(调参实验中:HR@20 <= 0.020380)

调参范围严格按照论文中的:
“We leverage AdamW as the optimizer. The embedding dimension of items is fixed as 64 across all models. The learning rate is tuned in the range of [0.01, 0.005, 0.001, 0.0005, 0.0001, 0.00005]. Despite that DreamRec does not require L2 regularization, we tune the weight of L2 regularization for all baselines in the range of [1e-3, 1e-4, 1e-5, 1e-6, 1e-7]. For all baselines, we conduct negative sampling from the uniform distribution at the ratio of 1: 1, which is not conducted in DreamRec. For our DreamRec, we fix the unconditional training probability pu as 0.1 suggested by [36]. We search the total diffusion step T in the range of [50, 100, 200, 500, 1000, 2000], and the personalized guidance strength w in the range of [0, 2, 4, 6, 8, 10].”

论文中的模型效果参数是如何设置?感谢。

@Yewx0310
Copy link

Yewx0310 commented Dec 5, 2023

+1,我跑zhihu也跑不出来论文中的结果

@YangZhengyi98
Copy link
Owner

你好,

可能是由于实验环境不同,在论文中给定的实验环境下(GPU type: 3090, python==3.7.13, pytorch==1.12.1, numpy==1.21.5),按提供参数的实验结果为:

8660cd9c52f8d2dd1aaf4950c6131ae

738b9706bb2f394526be4f7e76e2bfc

并且我在试验中也并未观测到ndcg@20高达0.008787的情况。

由于diffusion对参数极其敏感,因此在其他环境下可能调参范围和尺度会有所变化。

如果方便的话,可以提供一下你的GPU型号,我如果有相同型号的话可以调调看。

@Yewx0310
Copy link

Yewx0310 commented Dec 8, 2023

你好,

可能是由于实验环境不同,在论文中给定的实验环境下(GPU type: 3090, python==3.7.13, pytorch==1.12.1, numpy==1.21.5),按提供参数的实验结果为:

8660cd9c52f8d2dd1aaf4950c6131ae

738b9706bb2f394526be4f7e76e2bfc

并且我在试验中也并未观测到ndcg@20高达0.008787的情况。

由于diffusion对参数极其敏感,因此在其他环境下可能调参范围和尺度会有所变化。

如果方便的话,可以提供一下你的GPU型号,我如果有相同型号的话可以调调看。

我的GPU是4090,但是我也有3090的GPU,我可以试试

@KpiHang
Copy link
Author

KpiHang commented Dec 12, 2023

我的GPU:Tesla V100
python 3.9.18
pytorch 2.1.1
numpy 1.24.4

@Smgdx2002
Copy link

作者你好,我想了解下对于yc和ks数据集,大概要训练多少个epoch才能差不多达到论文中的效果呢?
我大致训练了200个epoch,但是效果距离论文中提及的还是相差甚远。

@YangZhengyi98
Copy link
Owner

你好,我的实验中yc在100个epoch左右,ks应该不到50个epoch。

@GithubX-F
Copy link

I trained for 600 epochs, but never achieved a case where HR@20 was greater than 2.00% or NDCG@20 was greater than 0.70%, haha.🤡

@Dreamhepeng
Copy link

作者你好,我想了解下对于yc和ks数据集,大概要训练多少个epoch才能差不多达到论文中的效果呢? 我大致训练了200个epoch,但是效果距离论文中提及的还是相差甚远。

你好,后面你是怎么查看ks的结果的呢?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

6 participants