search

KimMeen / Time-LLM

[ICLR 2024] Official implementation of " 🦙 Time-LLM: Time Series Forecasting by Reprogramming Large Language Models"
https://arxiv.org/abs/2310.01728
Apache License 2.0
1.46k stars 252 forks source link

checkpoint解析 #122

Closed leecmy08 closed 3 months ago

leecmy08 commented 4 months ago

您好，打扰您了，我想尝试复现您的代码，目前已经生成了checkpoint压缩包，如何解析该压缩包，请问是否有对应代码？

kwuking commented 4 months ago

你好，感谢对于我们工作的关注，checkpoint压缩包是指训练完TimeLLM之后的权重文件嘛，如果您处理后是tar包可以直接解压，参考exp文件中的text方法的代码加载模型权重即可。

leecmy08 commented 4 months ago

checkpoint压缩包是指训练完TimeLLM之后的权重文件，我是希望看到代码跑的最终的结果，包括MAE,MSE这些指标等，不知道您是否可以提供指导，目前我仅有一个checkpoint压缩包，我希望看到实验的结果，感谢您了

leecmy08 commented 4 months ago

请问模型的预测结果 也是在checkpoint文件里面吗

leecmy08 commented 4 months ago

作者 还有一个问题问您，我现在加载模型后，做推理后的ETTH1数据集预测和真实值是其中的某一列，并不是最后的OT油温，不知道这个问题应该如何解决，模型是一个多输入多输出的模型吗？

lphhh123 commented 3 months ago

Hello, sorry to bother you. I have been reproducing this paper recently. How did I obtain the compressed checkpoint file? After configuring the environment, running the '. sh' file will result in an error， and running run_ma4.py will also result in an error. Do I need to debug and modify in the reverse direction? Thank you very much image

kwuking commented 3 months ago

作者 还有一个问题问您，我现在加载模型后，做推理后的ETTH1数据集预测和真实值是其中的某一列，并不是最后的OT油温，不知道这个问题应该如何解决，模型是一个多输入多输出的模型吗？

感谢对于我们工作的关注，目前我们使用了channel independent策略（这个是从PatchTST开始被广泛采用的），目前是将多元时序转换成了一元时序来进行处理的，因此并不会只输出最后的OT列的预测结果。

kwuking commented 3 months ago

Hello, sorry to bother you. I have been reproducing this paper recently. How did I obtain the compressed checkpoint file? After configuring the environment, running the '. sh' file will result in an error， and running run_ma4.py will also result in an error. Do I need to debug and modify in the reverse direction? Thank you very much

Thanks for your attention. Currently, I am running the script without any issues. Could you please provide more logs to help with the analysis? Also, since we are using LLM as the base model and employing complex frameworks like DeepSpeed, you might need to check the related configurations and environment to ensure there are no problems.