Fine-Tuning Model with Covariates and Target Time Series

[thisthq]

Thank you for your contributions and those of your colleagues.

I am working on a multivariate time series forecasting task. My data consists of 31 time series, including 30 time series as covariates and 1 time series as a target that I need to forecast. Can I fine-tune MOIRAI with these 30 covariates and 1 target as described above?

Thank you very much.

[gorold]

Sure, you could do that. You would need to process your data into the appropriate format.

[Riofd]

Did you successfully generate dataset with correct feat_dynamic_real? If so, I believe you can fine-tune the model with it. Moreover, my fine-tuned improvement is compared with not fine-tuning that performs very bad on my dataset.

[Houyikai]

我测是不行。我读他代码，没发现处理feat_dynamic_real这个字段，只用了target。我感觉这个代码有问题，指定是哪不对。

[gorold]

The data generation script looks correct. For the current pre-training/fine-tuning implementation, feat_dynamic_real are not included, but with some minor edits, they should be supported.

[lilyniu]

I would like to proceed with including feat_dynamic_real in the training process. Could you provide some guidance on how we can make the necessary modifications to support this feature? Specifically, what kind of minor edits do you think would be required to accommodate feat_dynamic_real?

[gorold]

you'll need to process the data appropriately, and update the transforms to consider feat_dynamic_real too.